1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Imports the given method and converts it to semantic trees XX
12 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
13 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
23 #define Verify(cond, msg) \
28 verRaiseVerifyExceptionIfNeeded(INDEBUG(msg) DEBUGARG(__FILE__) DEBUGARG(__LINE__)); \
32 #define VerifyOrReturn(cond, msg) \
37 verRaiseVerifyExceptionIfNeeded(INDEBUG(msg) DEBUGARG(__FILE__) DEBUGARG(__LINE__)); \
42 #define VerifyOrReturnSpeculative(cond, msg, speculative) \
56 verRaiseVerifyExceptionIfNeeded(INDEBUG(msg) DEBUGARG(__FILE__) DEBUGARG(__LINE__)); \
62 /*****************************************************************************/
64 void Compiler::impInit()
68 impTreeList = nullptr;
69 impTreeLast = nullptr;
70 impInlinedCodeSize = 0;
74 /*****************************************************************************
76 * Pushes the given tree on the stack.
79 void Compiler::impPushOnStack(GenTree* tree, typeInfo ti)
81 /* Check for overflow. If inlining, we may be using a bigger stack */
83 if ((verCurrentState.esStackDepth >= info.compMaxStack) &&
84 (verCurrentState.esStackDepth >= impStkSize || ((compCurBB->bbFlags & BBF_IMPORTED) == 0)))
86 BADCODE("stack overflow");
90 // If we are pushing a struct, make certain we know the precise type!
91 if (tree->TypeGet() == TYP_STRUCT)
93 assert(ti.IsType(TI_STRUCT));
94 CORINFO_CLASS_HANDLE clsHnd = ti.GetClassHandle();
95 assert(clsHnd != NO_CLASS_HANDLE);
98 if (tiVerificationNeeded && !ti.IsDead())
100 assert(typeInfo::AreEquivalent(NormaliseForStack(ti), ti)); // types are normalized
102 // The ti type is consistent with the tree type.
105 // On 64-bit systems, nodes whose "proper" type is "native int" get labeled TYP_LONG.
106 // In the verification type system, we always transform "native int" to "TI_INT".
107 // Ideally, we would keep track of which nodes labeled "TYP_LONG" are really "native int", but
108 // attempts to do that have proved too difficult. Instead, we'll assume that in checks like this,
109 // when there's a mismatch, it's because of this reason -- the typeInfo::AreEquivalentModuloNativeInt
110 // method used in the last disjunct allows exactly this mismatch.
111 assert(ti.IsDead() || ti.IsByRef() && (tree->TypeGet() == TYP_I_IMPL || tree->TypeGet() == TYP_BYREF) ||
112 ti.IsUnboxedGenericTypeVar() && tree->TypeGet() == TYP_REF ||
113 ti.IsObjRef() && tree->TypeGet() == TYP_REF || ti.IsMethod() && tree->TypeGet() == TYP_I_IMPL ||
114 ti.IsType(TI_STRUCT) && tree->TypeGet() != TYP_REF ||
115 typeInfo::AreEquivalentModuloNativeInt(NormaliseForStack(ti),
116 NormaliseForStack(typeInfo(tree->TypeGet()))));
118 // If it is a struct type, make certain we normalized the primitive types
119 assert(!ti.IsType(TI_STRUCT) ||
120 info.compCompHnd->getTypeForPrimitiveValueClass(ti.GetClassHandle()) == CORINFO_TYPE_UNDEF);
124 if (VERBOSE && tiVerificationNeeded)
127 printf(TI_DUMP_PADDING);
128 printf("About to push to stack: ");
131 #endif // VERBOSE_VERIFY
135 verCurrentState.esStack[verCurrentState.esStackDepth].seTypeInfo = ti;
136 verCurrentState.esStack[verCurrentState.esStackDepth++].val = tree;
138 if ((tree->gtType == TYP_LONG) && (compLongUsed == false))
142 else if (((tree->gtType == TYP_FLOAT) || (tree->gtType == TYP_DOUBLE)) && (compFloatingPointUsed == false))
144 compFloatingPointUsed = true;
148 inline void Compiler::impPushNullObjRefOnStack()
150 impPushOnStack(gtNewIconNode(0, TYP_REF), typeInfo(TI_NULL));
153 // This method gets called when we run into unverifiable code
154 // (and we are verifying the method)
156 inline void Compiler::verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* msg) DEBUGARG(const char* file)
157 DEBUGARG(unsigned line))
159 // Remember that the code is not verifiable
160 // Note that the method may yet pass canSkipMethodVerification(),
161 // and so the presence of unverifiable code may not be an issue.
162 tiIsVerifiableCode = FALSE;
165 const char* tail = strrchr(file, '\\');
171 if (JitConfig.JitBreakOnUnsafeCode())
173 assert(!"Unsafe code detected");
177 JITLOG((LL_INFO10000, "Detected unsafe code: %s:%d : %s, while compiling %s opcode %s, IL offset %x\n", file, line,
178 msg, info.compFullName, impCurOpcName, impCurOpcOffs));
180 if (verNeedsVerification() || compIsForImportOnly())
182 JITLOG((LL_ERROR, "Verification failure: %s:%d : %s, while compiling %s opcode %s, IL offset %x\n", file, line,
183 msg, info.compFullName, impCurOpcName, impCurOpcOffs));
184 verRaiseVerifyException(INDEBUG(msg) DEBUGARG(file) DEBUGARG(line));
188 inline void DECLSPEC_NORETURN Compiler::verRaiseVerifyException(INDEBUG(const char* msg) DEBUGARG(const char* file)
189 DEBUGARG(unsigned line))
191 JITLOG((LL_ERROR, "Verification failure: %s:%d : %s, while compiling %s opcode %s, IL offset %x\n", file, line,
192 msg, info.compFullName, impCurOpcName, impCurOpcOffs));
195 // BreakIfDebuggerPresent();
196 if (getBreakOnBadCode())
198 assert(!"Typechecking error");
202 RaiseException(SEH_VERIFICATION_EXCEPTION, EXCEPTION_NONCONTINUABLE, 0, nullptr);
206 // helper function that will tell us if the IL instruction at the addr passed
207 // by param consumes an address at the top of the stack. We use it to save
209 bool Compiler::impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle)
211 assert(!compIsForInlining());
215 opcode = (OPCODE)getU1LittleEndian(codeAddr);
219 // case CEE_LDFLDA: We're taking this one out as if you have a sequence
225 // of a primitivelike struct, you end up after morphing with addr of a local
226 // that's not marked as addrtaken, which is wrong. Also ldflda is usually used
227 // for structs that contain other structs, which isnt a case we handle very
228 // well now for other reasons.
232 // We won't collapse small fields. This is probably not the right place to have this
233 // check, but we're only using the function for this purpose, and is easy to factor
234 // out if we need to do so.
236 CORINFO_RESOLVED_TOKEN resolvedToken;
237 impResolveToken(codeAddr + sizeof(__int8), &resolvedToken, CORINFO_TOKENKIND_Field);
239 CORINFO_CLASS_HANDLE clsHnd;
240 var_types lclTyp = JITtype2varType(info.compCompHnd->getFieldType(resolvedToken.hField, &clsHnd));
242 // Preserve 'small' int types
243 if (!varTypeIsSmall(lclTyp))
245 lclTyp = genActualType(lclTyp);
248 if (varTypeIsSmall(lclTyp))
262 void Compiler::impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind)
264 pResolvedToken->tokenContext = impTokenLookupContextHandle;
265 pResolvedToken->tokenScope = info.compScopeHnd;
266 pResolvedToken->token = getU4LittleEndian(addr);
267 pResolvedToken->tokenType = kind;
269 if (!tiVerificationNeeded)
271 info.compCompHnd->resolveToken(pResolvedToken);
275 Verify(eeTryResolveToken(pResolvedToken), "Token resolution failed");
279 /*****************************************************************************
281 * Pop one tree from the stack.
284 StackEntry Compiler::impPopStack()
286 if (verCurrentState.esStackDepth == 0)
288 BADCODE("stack underflow");
293 if (VERBOSE && tiVerificationNeeded)
296 printf(TI_DUMP_PADDING);
297 printf("About to pop from the stack: ");
298 const typeInfo& ti = verCurrentState.esStack[verCurrentState.esStackDepth - 1].seTypeInfo;
301 #endif // VERBOSE_VERIFY
304 return verCurrentState.esStack[--verCurrentState.esStackDepth];
307 /*****************************************************************************
309 * Peep at n'th (0-based) tree on the top of the stack.
312 StackEntry& Compiler::impStackTop(unsigned n)
314 if (verCurrentState.esStackDepth <= n)
316 BADCODE("stack underflow");
319 return verCurrentState.esStack[verCurrentState.esStackDepth - n - 1];
322 unsigned Compiler::impStackHeight()
324 return verCurrentState.esStackDepth;
327 /*****************************************************************************
328 * Some of the trees are spilled specially. While unspilling them, or
329 * making a copy, these need to be handled specially. The function
330 * enumerates the operators possible after spilling.
333 #ifdef DEBUG // only used in asserts
334 static bool impValidSpilledStackEntry(GenTree* tree)
336 if (tree->gtOper == GT_LCL_VAR)
341 if (tree->OperIsConst())
350 /*****************************************************************************
352 * The following logic is used to save/restore stack contents.
353 * If 'copy' is true, then we make a copy of the trees on the stack. These
354 * have to all be cloneable/spilled values.
357 void Compiler::impSaveStackState(SavedStack* savePtr, bool copy)
359 savePtr->ssDepth = verCurrentState.esStackDepth;
361 if (verCurrentState.esStackDepth)
363 savePtr->ssTrees = new (this, CMK_ImpStack) StackEntry[verCurrentState.esStackDepth];
364 size_t saveSize = verCurrentState.esStackDepth * sizeof(*savePtr->ssTrees);
368 StackEntry* table = savePtr->ssTrees;
370 /* Make a fresh copy of all the stack entries */
372 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++, table++)
374 table->seTypeInfo = verCurrentState.esStack[level].seTypeInfo;
375 GenTree* tree = verCurrentState.esStack[level].val;
377 assert(impValidSpilledStackEntry(tree));
379 switch (tree->gtOper)
386 table->val = gtCloneExpr(tree);
390 assert(!"Bad oper - Not covered by impValidSpilledStackEntry()");
397 memcpy(savePtr->ssTrees, verCurrentState.esStack, saveSize);
402 void Compiler::impRestoreStackState(SavedStack* savePtr)
404 verCurrentState.esStackDepth = savePtr->ssDepth;
406 if (verCurrentState.esStackDepth)
408 memcpy(verCurrentState.esStack, savePtr->ssTrees,
409 verCurrentState.esStackDepth * sizeof(*verCurrentState.esStack));
413 /*****************************************************************************
415 * Get the tree list started for a new basic block.
417 inline void Compiler::impBeginTreeList()
419 assert(impTreeList == nullptr && impTreeLast == nullptr);
421 impTreeList = impTreeLast = new (this, GT_BEG_STMTS) GenTree(GT_BEG_STMTS, TYP_VOID);
424 /*****************************************************************************
426 * Store the given start and end stmt in the given basic block. This is
427 * mostly called by impEndTreeList(BasicBlock *block). It is called
428 * directly only for handling CEE_LEAVEs out of finally-protected try's.
431 inline void Compiler::impEndTreeList(BasicBlock* block, GenTree* firstStmt, GenTree* lastStmt)
433 assert(firstStmt->gtOper == GT_STMT);
434 assert(lastStmt->gtOper == GT_STMT);
436 /* Make the list circular, so that we can easily walk it backwards */
438 firstStmt->gtPrev = lastStmt;
440 /* Store the tree list in the basic block */
442 block->bbTreeList = firstStmt;
444 /* The block should not already be marked as imported */
445 assert((block->bbFlags & BBF_IMPORTED) == 0);
447 block->bbFlags |= BBF_IMPORTED;
450 /*****************************************************************************
452 * Store the current tree list in the given basic block.
455 inline void Compiler::impEndTreeList(BasicBlock* block)
457 assert(impTreeList->gtOper == GT_BEG_STMTS);
459 GenTree* firstTree = impTreeList->gtNext;
463 /* The block should not already be marked as imported */
464 assert((block->bbFlags & BBF_IMPORTED) == 0);
466 // Empty block. Just mark it as imported
467 block->bbFlags |= BBF_IMPORTED;
471 // Ignore the GT_BEG_STMTS
472 assert(firstTree->gtPrev == impTreeList);
474 impEndTreeList(block, firstTree, impTreeLast);
478 if (impLastILoffsStmt != nullptr)
480 impLastILoffsStmt->gtStmt.gtStmtLastILoffs = compIsForInlining() ? BAD_IL_OFFSET : impCurOpcOffs;
481 impLastILoffsStmt = nullptr;
484 impTreeList = impTreeLast = nullptr;
488 /*****************************************************************************
490 * Check that storing the given tree doesnt mess up the semantic order. Note
491 * that this has only limited value as we can only check [0..chkLevel).
494 inline void Compiler::impAppendStmtCheck(GenTree* stmt, unsigned chkLevel)
499 assert(stmt->gtOper == GT_STMT);
501 if (chkLevel == (unsigned)CHECK_SPILL_ALL)
503 chkLevel = verCurrentState.esStackDepth;
506 if (verCurrentState.esStackDepth == 0 || chkLevel == 0 || chkLevel == (unsigned)CHECK_SPILL_NONE)
511 GenTree* tree = stmt->gtStmt.gtStmtExpr;
513 // Calls can only be appended if there are no GTF_GLOB_EFFECT on the stack
515 if (tree->gtFlags & GTF_CALL)
517 for (unsigned level = 0; level < chkLevel; level++)
519 assert((verCurrentState.esStack[level].val->gtFlags & GTF_GLOB_EFFECT) == 0);
523 if (tree->gtOper == GT_ASG)
525 // For an assignment to a local variable, all references of that
526 // variable have to be spilled. If it is aliased, all calls and
527 // indirect accesses have to be spilled
529 if (tree->gtOp.gtOp1->gtOper == GT_LCL_VAR)
531 unsigned lclNum = tree->gtOp.gtOp1->gtLclVarCommon.gtLclNum;
532 for (unsigned level = 0; level < chkLevel; level++)
534 assert(!gtHasRef(verCurrentState.esStack[level].val, lclNum, false));
535 assert(!lvaTable[lclNum].lvAddrExposed ||
536 (verCurrentState.esStack[level].val->gtFlags & GTF_SIDE_EFFECT) == 0);
540 // If the access may be to global memory, all side effects have to be spilled.
542 else if (tree->gtOp.gtOp1->gtFlags & GTF_GLOB_REF)
544 for (unsigned level = 0; level < chkLevel; level++)
546 assert((verCurrentState.esStack[level].val->gtFlags & GTF_GLOB_REF) == 0);
553 /*****************************************************************************
555 * Append the given GT_STMT node to the current block's tree list.
556 * [0..chkLevel) is the portion of the stack which we will check for
557 * interference with stmt and spill if needed.
560 inline void Compiler::impAppendStmt(GenTree* stmt, unsigned chkLevel)
562 assert(stmt->gtOper == GT_STMT);
563 noway_assert(impTreeLast != nullptr);
565 /* If the statement being appended has any side-effects, check the stack
566 to see if anything needs to be spilled to preserve correct ordering. */
568 GenTree* expr = stmt->gtStmt.gtStmtExpr;
569 unsigned flags = expr->gtFlags & GTF_GLOB_EFFECT;
571 // Assignment to (unaliased) locals don't count as a side-effect as
572 // we handle them specially using impSpillLclRefs(). Temp locals should
575 if ((expr->gtOper == GT_ASG) && (expr->gtOp.gtOp1->gtOper == GT_LCL_VAR) &&
576 !(expr->gtOp.gtOp1->gtFlags & GTF_GLOB_REF) && !gtHasLocalsWithAddrOp(expr->gtOp.gtOp2))
578 unsigned op2Flags = expr->gtOp.gtOp2->gtFlags & GTF_GLOB_EFFECT;
579 assert(flags == (op2Flags | GTF_ASG));
583 if (chkLevel == (unsigned)CHECK_SPILL_ALL)
585 chkLevel = verCurrentState.esStackDepth;
588 if (chkLevel && chkLevel != (unsigned)CHECK_SPILL_NONE)
590 assert(chkLevel <= verCurrentState.esStackDepth);
594 // If there is a call, we have to spill global refs
595 bool spillGlobEffects = (flags & GTF_CALL) ? true : false;
597 if (expr->gtOper == GT_ASG)
599 GenTree* lhs = expr->gtGetOp1();
600 // If we are assigning to a global ref, we have to spill global refs on stack.
601 // TODO-1stClassStructs: Previously, spillGlobEffects was set to true for
602 // GT_INITBLK and GT_COPYBLK, but this is overly conservative, and should be
603 // revisited. (Note that it was NOT set to true for GT_COPYOBJ.)
604 if (!expr->OperIsBlkOp())
606 // If we are assigning to a global ref, we have to spill global refs on stack
607 if ((lhs->gtFlags & GTF_GLOB_REF) != 0)
609 spillGlobEffects = true;
612 else if ((lhs->OperIsBlk() && !lhs->AsBlk()->HasGCPtr()) ||
613 ((lhs->OperGet() == GT_LCL_VAR) &&
614 (lvaTable[lhs->AsLclVarCommon()->gtLclNum].lvStructGcCount == 0)))
616 spillGlobEffects = true;
620 impSpillSideEffects(spillGlobEffects, chkLevel DEBUGARG("impAppendStmt"));
624 impSpillSpecialSideEff();
628 impAppendStmtCheck(stmt, chkLevel);
630 /* Point 'prev' at the previous node, so that we can walk backwards */
632 stmt->gtPrev = impTreeLast;
634 /* Append the expression statement to the list */
636 impTreeLast->gtNext = stmt;
640 impMarkContiguousSIMDFieldAssignments(stmt);
643 /* Once we set impCurStmtOffs in an appended tree, we are ready to
644 report the following offsets. So reset impCurStmtOffs */
646 if (impTreeLast->gtStmt.gtStmtILoffsx == impCurStmtOffs)
648 impCurStmtOffsSet(BAD_IL_OFFSET);
652 if (impLastILoffsStmt == nullptr)
654 impLastILoffsStmt = stmt;
665 /*****************************************************************************
667 * Insert the given GT_STMT "stmt" before GT_STMT "stmtBefore"
670 inline void Compiler::impInsertStmtBefore(GenTree* stmt, GenTree* stmtBefore)
672 assert(stmt->gtOper == GT_STMT);
673 assert(stmtBefore->gtOper == GT_STMT);
675 GenTree* stmtPrev = stmtBefore->gtPrev;
676 stmt->gtPrev = stmtPrev;
677 stmt->gtNext = stmtBefore;
678 stmtPrev->gtNext = stmt;
679 stmtBefore->gtPrev = stmt;
682 /*****************************************************************************
684 * Append the given expression tree to the current block's tree list.
685 * Return the newly created statement.
688 GenTree* Compiler::impAppendTree(GenTree* tree, unsigned chkLevel, IL_OFFSETX offset)
692 /* Allocate an 'expression statement' node */
694 GenTree* expr = gtNewStmt(tree, offset);
696 /* Append the statement to the current block's stmt list */
698 impAppendStmt(expr, chkLevel);
703 /*****************************************************************************
705 * Insert the given exression tree before GT_STMT "stmtBefore"
708 void Compiler::impInsertTreeBefore(GenTree* tree, IL_OFFSETX offset, GenTree* stmtBefore)
710 assert(stmtBefore->gtOper == GT_STMT);
712 /* Allocate an 'expression statement' node */
714 GenTree* expr = gtNewStmt(tree, offset);
716 /* Append the statement to the current block's stmt list */
718 impInsertStmtBefore(expr, stmtBefore);
721 /*****************************************************************************
723 * Append an assignment of the given value to a temp to the current tree list.
724 * curLevel is the stack level for which the spill to the temp is being done.
727 void Compiler::impAssignTempGen(unsigned tmp,
730 GenTree** pAfterStmt, /* = NULL */
731 IL_OFFSETX ilOffset, /* = BAD_IL_OFFSET */
732 BasicBlock* block /* = NULL */
735 GenTree* asg = gtNewTempAssign(tmp, val);
737 if (!asg->IsNothingNode())
741 GenTree* asgStmt = gtNewStmt(asg, ilOffset);
742 *pAfterStmt = fgInsertStmtAfter(block, *pAfterStmt, asgStmt);
746 impAppendTree(asg, curLevel, impCurStmtOffs);
751 /*****************************************************************************
752 * same as above, but handle the valueclass case too
755 void Compiler::impAssignTempGen(unsigned tmpNum,
757 CORINFO_CLASS_HANDLE structType,
759 GenTree** pAfterStmt, /* = NULL */
760 IL_OFFSETX ilOffset, /* = BAD_IL_OFFSET */
761 BasicBlock* block /* = NULL */
766 if (varTypeIsStruct(val))
768 assert(tmpNum < lvaCount);
769 assert(structType != NO_CLASS_HANDLE);
771 // if the method is non-verifiable the assert is not true
772 // so at least ignore it in the case when verification is turned on
773 // since any block that tries to use the temp would have failed verification.
774 var_types varType = lvaTable[tmpNum].lvType;
775 assert(tiVerificationNeeded || varType == TYP_UNDEF || varTypeIsStruct(varType));
776 lvaSetStruct(tmpNum, structType, false);
778 // Now, set the type of the struct value. Note that lvaSetStruct may modify the type
779 // of the lclVar to a specialized type (e.g. TYP_SIMD), based on the handle (structType)
780 // that has been passed in for the value being assigned to the temp, in which case we
781 // need to set 'val' to that same type.
782 // Note also that if we always normalized the types of any node that might be a struct
783 // type, this would not be necessary - but that requires additional JIT/EE interface
784 // calls that may not actually be required - e.g. if we only access a field of a struct.
786 val->gtType = lvaTable[tmpNum].lvType;
788 GenTree* dst = gtNewLclvNode(tmpNum, val->gtType);
789 asg = impAssignStruct(dst, val, structType, curLevel, pAfterStmt, block);
793 asg = gtNewTempAssign(tmpNum, val);
796 if (!asg->IsNothingNode())
800 GenTree* asgStmt = gtNewStmt(asg, ilOffset);
801 *pAfterStmt = fgInsertStmtAfter(block, *pAfterStmt, asgStmt);
805 impAppendTree(asg, curLevel, impCurStmtOffs);
810 /*****************************************************************************
812 * Pop the given number of values from the stack and return a list node with
814 * The 'prefixTree' argument may optionally contain an argument
815 * list that is prepended to the list returned from this function.
817 * The notion of prepended is a bit misleading in that the list is backwards
818 * from the way I would expect: The first element popped is at the end of
819 * the returned list, and prefixTree is 'before' that, meaning closer to
820 * the end of the list. To get to prefixTree, you have to walk to the
823 * For ARG_ORDER_R2L prefixTree is only used to insert extra arguments, as
824 * such we reverse its meaning such that returnValue has a reversed
825 * prefixTree at the head of the list.
828 GenTreeArgList* Compiler::impPopList(unsigned count, CORINFO_SIG_INFO* sig, GenTreeArgList* prefixTree)
830 assert(sig == nullptr || count == sig->numArgs);
832 CORINFO_CLASS_HANDLE structType;
833 GenTreeArgList* treeList;
835 if (Target::g_tgtArgOrder == Target::ARG_ORDER_R2L)
841 treeList = prefixTree;
846 StackEntry se = impPopStack();
847 typeInfo ti = se.seTypeInfo;
848 GenTree* temp = se.val;
850 if (varTypeIsStruct(temp))
852 // Morph trees that aren't already OBJs or MKREFANY to be OBJs
853 assert(ti.IsType(TI_STRUCT));
854 structType = ti.GetClassHandleForValueClass();
858 printf("Calling impNormStructVal on:\n");
862 temp = impNormStructVal(temp, structType, (unsigned)CHECK_SPILL_ALL);
866 printf("resulting tree:\n");
872 /* NOTE: we defer bashing the type for I_IMPL to fgMorphArgs */
873 treeList = gtNewListNode(temp, treeList);
878 if (sig->retTypeSigClass != nullptr && sig->retType != CORINFO_TYPE_CLASS &&
879 sig->retType != CORINFO_TYPE_BYREF && sig->retType != CORINFO_TYPE_PTR && sig->retType != CORINFO_TYPE_VAR)
881 // Make sure that all valuetypes (including enums) that we push are loaded.
882 // This is to guarantee that if a GC is triggerred from the prestub of this methods,
883 // all valuetypes in the method signature are already loaded.
884 // We need to be able to find the size of the valuetypes, but we cannot
885 // do a class-load from within GC.
886 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(sig->retTypeSigClass);
889 CORINFO_ARG_LIST_HANDLE argLst = sig->args;
890 CORINFO_CLASS_HANDLE argClass;
891 CORINFO_CLASS_HANDLE argRealClass;
892 GenTreeArgList* args;
894 for (args = treeList, count = sig->numArgs; count > 0; args = args->Rest(), count--)
896 PREFIX_ASSUME(args != nullptr);
898 CorInfoType corType = strip(info.compCompHnd->getArgType(sig, argLst, &argClass));
900 // insert implied casts (from float to double or double to float)
902 if (corType == CORINFO_TYPE_DOUBLE && args->Current()->TypeGet() == TYP_FLOAT)
904 args->Current() = gtNewCastNode(TYP_DOUBLE, args->Current(), false, TYP_DOUBLE);
906 else if (corType == CORINFO_TYPE_FLOAT && args->Current()->TypeGet() == TYP_DOUBLE)
908 args->Current() = gtNewCastNode(TYP_FLOAT, args->Current(), false, TYP_FLOAT);
911 // insert any widening or narrowing casts for backwards compatibility
913 args->Current() = impImplicitIorI4Cast(args->Current(), JITtype2varType(corType));
915 if (corType != CORINFO_TYPE_CLASS && corType != CORINFO_TYPE_BYREF && corType != CORINFO_TYPE_PTR &&
916 corType != CORINFO_TYPE_VAR && (argRealClass = info.compCompHnd->getArgClass(sig, argLst)) != nullptr)
918 // Everett MC++ could generate IL with a mismatched valuetypes. It used to work with Everett JIT,
919 // but it stopped working in Whidbey when we have started passing simple valuetypes as underlying
921 // We will try to adjust for this case here to avoid breaking customers code (see VSW 485789 for
923 if (corType == CORINFO_TYPE_VALUECLASS && !varTypeIsStruct(args->Current()))
925 args->Current() = impNormStructVal(args->Current(), argRealClass, (unsigned)CHECK_SPILL_ALL, true);
928 // Make sure that all valuetypes (including enums) that we push are loaded.
929 // This is to guarantee that if a GC is triggered from the prestub of this methods,
930 // all valuetypes in the method signature are already loaded.
931 // We need to be able to find the size of the valuetypes, but we cannot
932 // do a class-load from within GC.
933 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(argRealClass);
936 argLst = info.compCompHnd->getArgNext(argLst);
940 if (Target::g_tgtArgOrder == Target::ARG_ORDER_R2L)
942 // Prepend the prefixTree
944 // Simple in-place reversal to place treeList
945 // at the end of a reversed prefixTree
946 while (prefixTree != nullptr)
948 GenTreeArgList* next = prefixTree->Rest();
949 prefixTree->Rest() = treeList;
950 treeList = prefixTree;
957 /*****************************************************************************
959 * Pop the given number of values from the stack in reverse order (STDCALL/CDECL etc.)
960 * The first "skipReverseCount" items are not reversed.
963 GenTreeArgList* Compiler::impPopRevList(unsigned count, CORINFO_SIG_INFO* sig, unsigned skipReverseCount)
966 assert(skipReverseCount <= count);
968 GenTreeArgList* list = impPopList(count, sig);
971 if (list == nullptr || skipReverseCount == count)
976 GenTreeArgList* ptr = nullptr; // Initialized to the first node that needs to be reversed
977 GenTreeArgList* lastSkipNode = nullptr; // Will be set to the last node that does not need to be reversed
979 if (skipReverseCount == 0)
986 // Get to the first node that needs to be reversed
987 for (unsigned i = 0; i < skipReverseCount - 1; i++)
989 lastSkipNode = lastSkipNode->Rest();
992 PREFIX_ASSUME(lastSkipNode != nullptr);
993 ptr = lastSkipNode->Rest();
996 GenTreeArgList* reversedList = nullptr;
1000 GenTreeArgList* tmp = ptr->Rest();
1001 ptr->Rest() = reversedList;
1004 } while (ptr != nullptr);
1006 if (skipReverseCount)
1008 lastSkipNode->Rest() = reversedList;
1013 return reversedList;
1017 /*****************************************************************************
1018 Assign (copy) the structure from 'src' to 'dest'. The structure is a value
1019 class of type 'clsHnd'. It returns the tree that should be appended to the
1020 statement list that represents the assignment.
1021 Temp assignments may be appended to impTreeList if spilling is necessary.
1022 curLevel is the stack level for which a spill may be being done.
1025 GenTree* Compiler::impAssignStruct(GenTree* dest,
1027 CORINFO_CLASS_HANDLE structHnd,
1029 GenTree** pAfterStmt, /* = NULL */
1030 BasicBlock* block /* = NULL */
1033 assert(varTypeIsStruct(dest));
1035 while (dest->gtOper == GT_COMMA)
1037 assert(varTypeIsStruct(dest->gtOp.gtOp2)); // Second thing is the struct
1039 // Append all the op1 of GT_COMMA trees before we evaluate op2 of the GT_COMMA tree.
1042 *pAfterStmt = fgInsertStmtAfter(block, *pAfterStmt, gtNewStmt(dest->gtOp.gtOp1, impCurStmtOffs));
1046 impAppendTree(dest->gtOp.gtOp1, curLevel, impCurStmtOffs); // do the side effect
1049 // set dest to the second thing
1050 dest = dest->gtOp.gtOp2;
1053 assert(dest->gtOper == GT_LCL_VAR || dest->gtOper == GT_RETURN || dest->gtOper == GT_FIELD ||
1054 dest->gtOper == GT_IND || dest->gtOper == GT_OBJ || dest->gtOper == GT_INDEX);
1056 if (dest->OperGet() == GT_LCL_VAR && src->OperGet() == GT_LCL_VAR &&
1057 src->gtLclVarCommon.gtLclNum == dest->gtLclVarCommon.gtLclNum)
1060 return gtNewNothingNode();
1063 // TODO-1stClassStructs: Avoid creating an address if it is not needed,
1064 // or re-creating a Blk node if it is.
1067 if (dest->gtOper == GT_IND || dest->OperIsBlk())
1069 destAddr = dest->gtOp.gtOp1;
1073 destAddr = gtNewOperNode(GT_ADDR, TYP_BYREF, dest);
1076 return (impAssignStructPtr(destAddr, src, structHnd, curLevel, pAfterStmt, block));
1079 /*****************************************************************************/
1081 GenTree* Compiler::impAssignStructPtr(GenTree* destAddr,
1083 CORINFO_CLASS_HANDLE structHnd,
1085 GenTree** pAfterStmt, /* = NULL */
1086 BasicBlock* block /* = NULL */
1090 GenTree* dest = nullptr;
1091 unsigned destFlags = 0;
1093 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1094 assert(varTypeIsStruct(src) || (src->gtOper == GT_ADDR && src->TypeGet() == TYP_BYREF));
1095 // TODO-ARM-BUG: Does ARM need this?
1096 // TODO-ARM64-BUG: Does ARM64 need this?
1097 assert(src->gtOper == GT_LCL_VAR || src->gtOper == GT_FIELD || src->gtOper == GT_IND || src->gtOper == GT_OBJ ||
1098 src->gtOper == GT_CALL || src->gtOper == GT_MKREFANY || src->gtOper == GT_RET_EXPR ||
1099 src->gtOper == GT_COMMA || src->gtOper == GT_ADDR ||
1100 (src->TypeGet() != TYP_STRUCT &&
1101 (GenTree::OperIsSIMD(src->gtOper) || src->OperIsSimdHWIntrinsic() || src->gtOper == GT_LCL_FLD)));
1102 #else // !defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1103 assert(varTypeIsStruct(src));
1105 assert(src->gtOper == GT_LCL_VAR || src->gtOper == GT_FIELD || src->gtOper == GT_IND || src->gtOper == GT_OBJ ||
1106 src->gtOper == GT_CALL || src->gtOper == GT_MKREFANY || src->gtOper == GT_RET_EXPR ||
1107 src->gtOper == GT_COMMA ||
1108 (src->TypeGet() != TYP_STRUCT &&
1109 (GenTree::OperIsSIMD(src->gtOper) || src->OperIsSimdHWIntrinsic() || src->gtOper == GT_LCL_FLD)));
1110 #endif // !defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1111 if (destAddr->OperGet() == GT_ADDR)
1113 GenTree* destNode = destAddr->gtGetOp1();
1114 // If the actual destination is a local (for non-LEGACY_BACKEND), or already a block node, or is a node that
1115 // will be morphed, don't insert an OBJ(ADDR).
1116 if (destNode->gtOper == GT_INDEX || destNode->OperIsBlk()
1117 #ifndef LEGACY_BACKEND
1118 || ((destNode->OperGet() == GT_LCL_VAR) && (destNode->TypeGet() == src->TypeGet()))
1119 #endif // !LEGACY_BACKEND
1124 destType = destNode->TypeGet();
1128 destType = src->TypeGet();
1131 var_types asgType = src->TypeGet();
1133 if (src->gtOper == GT_CALL)
1135 if (src->AsCall()->TreatAsHasRetBufArg(this))
1137 // Case of call returning a struct via hidden retbuf arg
1139 // insert the return value buffer into the argument list as first byref parameter
1140 src->gtCall.gtCallArgs = gtNewListNode(destAddr, src->gtCall.gtCallArgs);
1142 // now returns void, not a struct
1143 src->gtType = TYP_VOID;
1145 // return the morphed call node
1150 // Case of call returning a struct in one or more registers.
1152 var_types returnType = (var_types)src->gtCall.gtReturnType;
1154 // We won't use a return buffer, so change the type of src->gtType to 'returnType'
1155 src->gtType = genActualType(returnType);
1157 // First we try to change this to "LclVar/LclFld = call"
1159 if ((destAddr->gtOper == GT_ADDR) && (destAddr->gtOp.gtOp1->gtOper == GT_LCL_VAR))
1161 // If it is a multi-reg struct return, don't change the oper to GT_LCL_FLD.
1162 // That is, the IR will be of the form lclVar = call for multi-reg return
1164 GenTree* lcl = destAddr->gtOp.gtOp1;
1165 if (src->AsCall()->HasMultiRegRetVal())
1167 // Mark the struct LclVar as used in a MultiReg return context
1168 // which currently makes it non promotable.
1169 // TODO-1stClassStructs: Eliminate this pessimization when we can more generally
1170 // handle multireg returns.
1171 lcl->gtFlags |= GTF_DONT_CSE;
1172 lvaTable[lcl->gtLclVarCommon.gtLclNum].lvIsMultiRegRet = true;
1174 else // The call result is not a multireg return
1176 // We change this to a GT_LCL_FLD (from a GT_ADDR of a GT_LCL_VAR)
1177 lcl->ChangeOper(GT_LCL_FLD);
1178 fgLclFldAssign(lcl->gtLclVarCommon.gtLclNum);
1179 lcl->gtType = src->gtType;
1180 asgType = src->gtType;
1185 #if defined(_TARGET_ARM_)
1186 // TODO-Cleanup: This should have been taken care of in the above HasMultiRegRetVal() case,
1187 // but that method has not been updadted to include ARM.
1188 impMarkLclDstNotPromotable(lcl->gtLclVarCommon.gtLclNum, src, structHnd);
1189 lcl->gtFlags |= GTF_DONT_CSE;
1190 #elif defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1191 // Not allowed for FEATURE_CORCLR which is the only SKU available for System V OSs.
1192 assert(!src->gtCall.IsVarargs() && "varargs not allowed for System V OSs.");
1194 // Make the struct non promotable. The eightbytes could contain multiple fields.
1195 // TODO-1stClassStructs: Eliminate this pessimization when we can more generally
1196 // handle multireg returns.
1197 // TODO-Cleanup: Why is this needed here? This seems that it will set this even for
1198 // non-multireg returns.
1199 lcl->gtFlags |= GTF_DONT_CSE;
1200 lvaTable[lcl->gtLclVarCommon.gtLclNum].lvIsMultiRegRet = true;
1203 else // we don't have a GT_ADDR of a GT_LCL_VAR
1205 // !!! The destination could be on stack. !!!
1206 // This flag will let us choose the correct write barrier.
1207 asgType = returnType;
1208 destFlags = GTF_IND_TGTANYWHERE;
1212 else if (src->gtOper == GT_RET_EXPR)
1214 GenTreeCall* call = src->gtRetExpr.gtInlineCandidate->AsCall();
1215 noway_assert(call->gtOper == GT_CALL);
1217 if (call->HasRetBufArg())
1219 // insert the return value buffer into the argument list as first byref parameter
1220 call->gtCallArgs = gtNewListNode(destAddr, call->gtCallArgs);
1222 // now returns void, not a struct
1223 src->gtType = TYP_VOID;
1224 call->gtType = TYP_VOID;
1226 // We already have appended the write to 'dest' GT_CALL's args
1227 // So now we just return an empty node (pruning the GT_RET_EXPR)
1232 // Case of inline method returning a struct in one or more registers.
1234 var_types returnType = (var_types)call->gtReturnType;
1236 // We won't need a return buffer
1237 asgType = returnType;
1238 src->gtType = genActualType(returnType);
1239 call->gtType = src->gtType;
1241 // If we've changed the type, and it no longer matches a local destination,
1242 // we must use an indirection.
1243 if ((dest != nullptr) && (dest->OperGet() == GT_LCL_VAR) && (dest->TypeGet() != asgType))
1248 // !!! The destination could be on stack. !!!
1249 // This flag will let us choose the correct write barrier.
1250 destFlags = GTF_IND_TGTANYWHERE;
1253 else if (src->OperIsBlk())
1255 asgType = impNormStructType(structHnd);
1256 if (src->gtOper == GT_OBJ)
1258 assert(src->gtObj.gtClass == structHnd);
1261 else if (src->gtOper == GT_INDEX)
1263 asgType = impNormStructType(structHnd);
1264 assert(src->gtIndex.gtStructElemClass == structHnd);
1266 else if (src->gtOper == GT_MKREFANY)
1268 // Since we are assigning the result of a GT_MKREFANY,
1269 // "destAddr" must point to a refany.
1271 GenTree* destAddrClone;
1273 impCloneExpr(destAddr, &destAddrClone, structHnd, curLevel, pAfterStmt DEBUGARG("MKREFANY assignment"));
1275 assert(offsetof(CORINFO_RefAny, dataPtr) == 0);
1276 assert(destAddr->gtType == TYP_I_IMPL || destAddr->gtType == TYP_BYREF);
1277 GetZeroOffsetFieldMap()->Set(destAddr, GetFieldSeqStore()->CreateSingleton(GetRefanyDataField()));
1278 GenTree* ptrSlot = gtNewOperNode(GT_IND, TYP_I_IMPL, destAddr);
1279 GenTreeIntCon* typeFieldOffset = gtNewIconNode(offsetof(CORINFO_RefAny, type), TYP_I_IMPL);
1280 typeFieldOffset->gtFieldSeq = GetFieldSeqStore()->CreateSingleton(GetRefanyTypeField());
1282 gtNewOperNode(GT_IND, TYP_I_IMPL, gtNewOperNode(GT_ADD, destAddr->gtType, destAddrClone, typeFieldOffset));
1284 // append the assign of the pointer value
1285 GenTree* asg = gtNewAssignNode(ptrSlot, src->gtOp.gtOp1);
1288 *pAfterStmt = fgInsertStmtAfter(block, *pAfterStmt, gtNewStmt(asg, impCurStmtOffs));
1292 impAppendTree(asg, curLevel, impCurStmtOffs);
1295 // return the assign of the type value, to be appended
1296 return gtNewAssignNode(typeSlot, src->gtOp.gtOp2);
1298 else if (src->gtOper == GT_COMMA)
1300 // The second thing is the struct or its address.
1301 assert(varTypeIsStruct(src->gtOp.gtOp2) || src->gtOp.gtOp2->gtType == TYP_BYREF);
1304 *pAfterStmt = fgInsertStmtAfter(block, *pAfterStmt, gtNewStmt(src->gtOp.gtOp1, impCurStmtOffs));
1308 impAppendTree(src->gtOp.gtOp1, curLevel, impCurStmtOffs); // do the side effect
1311 // Evaluate the second thing using recursion.
1312 return impAssignStructPtr(destAddr, src->gtOp.gtOp2, structHnd, curLevel, pAfterStmt, block);
1314 else if (src->IsLocal())
1316 asgType = src->TypeGet();
1318 else if (asgType == TYP_STRUCT)
1320 asgType = impNormStructType(structHnd);
1321 src->gtType = asgType;
1322 #ifdef LEGACY_BACKEND
1323 if (asgType == TYP_STRUCT)
1325 GenTree* srcAddr = gtNewOperNode(GT_ADDR, TYP_BYREF, src);
1326 src = gtNewOperNode(GT_IND, TYP_STRUCT, srcAddr);
1330 if (dest == nullptr)
1332 // TODO-1stClassStructs: We shouldn't really need a block node as the destination
1333 // if this is a known struct type.
1334 if (asgType == TYP_STRUCT)
1336 dest = gtNewObjNode(structHnd, destAddr);
1337 gtSetObjGcInfo(dest->AsObj());
1338 // Although an obj as a call argument was always assumed to be a globRef
1339 // (which is itself overly conservative), that is not true of the operands
1340 // of a block assignment.
1341 dest->gtFlags &= ~GTF_GLOB_REF;
1342 dest->gtFlags |= (destAddr->gtFlags & GTF_GLOB_REF);
1344 else if (varTypeIsStruct(asgType))
1346 dest = new (this, GT_BLK) GenTreeBlk(GT_BLK, asgType, destAddr, genTypeSize(asgType));
1350 dest = gtNewOperNode(GT_IND, asgType, destAddr);
1355 dest->gtType = asgType;
1358 dest->gtFlags |= destFlags;
1359 destFlags = dest->gtFlags;
1361 // return an assignment node, to be appended
1362 GenTree* asgNode = gtNewAssignNode(dest, src);
1363 gtBlockOpInit(asgNode, dest, src, false);
1365 // TODO-1stClassStructs: Clean up the settings of GTF_DONT_CSE on the lhs
1367 if ((destFlags & GTF_DONT_CSE) == 0)
1369 dest->gtFlags &= ~(GTF_DONT_CSE);
1374 /*****************************************************************************
1375 Given a struct value, and the class handle for that structure, return
1376 the expression for the address for that structure value.
1378 willDeref - does the caller guarantee to dereference the pointer.
1381 GenTree* Compiler::impGetStructAddr(GenTree* structVal,
1382 CORINFO_CLASS_HANDLE structHnd,
1386 assert(varTypeIsStruct(structVal) || eeIsValueClass(structHnd));
1388 var_types type = structVal->TypeGet();
1390 genTreeOps oper = structVal->gtOper;
1392 if (oper == GT_OBJ && willDeref)
1394 assert(structVal->gtObj.gtClass == structHnd);
1395 return (structVal->gtObj.Addr());
1397 else if (oper == GT_CALL || oper == GT_RET_EXPR || oper == GT_OBJ || oper == GT_MKREFANY ||
1398 structVal->OperIsSimdHWIntrinsic())
1400 unsigned tmpNum = lvaGrabTemp(true DEBUGARG("struct address for call/obj"));
1402 impAssignTempGen(tmpNum, structVal, structHnd, curLevel);
1404 // The 'return value' is now the temp itself
1406 type = genActualType(lvaTable[tmpNum].TypeGet());
1407 GenTree* temp = gtNewLclvNode(tmpNum, type);
1408 temp = gtNewOperNode(GT_ADDR, TYP_BYREF, temp);
1411 else if (oper == GT_COMMA)
1413 assert(structVal->gtOp.gtOp2->gtType == type); // Second thing is the struct
1415 GenTree* oldTreeLast = impTreeLast;
1416 structVal->gtOp.gtOp2 = impGetStructAddr(structVal->gtOp.gtOp2, structHnd, curLevel, willDeref);
1417 structVal->gtType = TYP_BYREF;
1419 if (oldTreeLast != impTreeLast)
1421 // Some temp assignment statement was placed on the statement list
1422 // for Op2, but that would be out of order with op1, so we need to
1423 // spill op1 onto the statement list after whatever was last
1424 // before we recursed on Op2 (i.e. before whatever Op2 appended).
1425 impInsertTreeBefore(structVal->gtOp.gtOp1, impCurStmtOffs, oldTreeLast->gtNext);
1426 structVal->gtOp.gtOp1 = gtNewNothingNode();
1432 return (gtNewOperNode(GT_ADDR, TYP_BYREF, structVal));
1435 //------------------------------------------------------------------------
1436 // impNormStructType: Given a (known to be) struct class handle structHnd, normalize its type,
1437 // and optionally determine the GC layout of the struct.
1440 // structHnd - The class handle for the struct type of interest.
1441 // gcLayout - (optional, default nullptr) - a BYTE pointer, allocated by the caller,
1442 // into which the gcLayout will be written.
1443 // pNumGCVars - (optional, default nullptr) - if non-null, a pointer to an unsigned,
1444 // which will be set to the number of GC fields in the struct.
1445 // pSimdBaseType - (optional, default nullptr) - if non-null, and the struct is a SIMD
1446 // type, set to the SIMD base type
1449 // The JIT type for the struct (e.g. TYP_STRUCT, or TYP_SIMD*).
1450 // The gcLayout will be returned using the pointers provided by the caller, if non-null.
1451 // It may also modify the compFloatingPointUsed flag if the type is a SIMD type.
1454 // The caller must set gcLayout to nullptr OR ensure that it is large enough
1455 // (see ICorStaticInfo::getClassGClayout in corinfo.h).
1458 // Normalizing the type involves examining the struct type to determine if it should
1459 // be modified to one that is handled specially by the JIT, possibly being a candidate
1460 // for full enregistration, e.g. TYP_SIMD16.
1462 var_types Compiler::impNormStructType(CORINFO_CLASS_HANDLE structHnd,
1464 unsigned* pNumGCVars,
1465 var_types* pSimdBaseType)
1467 assert(structHnd != NO_CLASS_HANDLE);
1469 const DWORD structFlags = info.compCompHnd->getClassAttribs(structHnd);
1470 var_types structType = TYP_STRUCT;
1472 // On coreclr the check for GC includes a "may" to account for the special
1473 // ByRef like span structs. The added check for "CONTAINS_STACK_PTR" is the particular bit.
1474 // When this is set the struct will contain a ByRef that could be a GC pointer or a native
1476 const bool mayContainGCPtrs =
1477 ((structFlags & CORINFO_FLG_CONTAINS_STACK_PTR) != 0 || ((structFlags & CORINFO_FLG_CONTAINS_GC_PTR) != 0));
1480 // Check to see if this is a SIMD type.
1481 if (featureSIMD && !mayContainGCPtrs)
1483 unsigned originalSize = info.compCompHnd->getClassSize(structHnd);
1485 if ((originalSize >= minSIMDStructBytes()) && (originalSize <= maxSIMDStructBytes()))
1487 unsigned int sizeBytes;
1488 var_types simdBaseType = getBaseTypeAndSizeOfSIMDType(structHnd, &sizeBytes);
1489 if (simdBaseType != TYP_UNKNOWN)
1491 assert(sizeBytes == originalSize);
1492 structType = getSIMDTypeForSize(sizeBytes);
1493 if (pSimdBaseType != nullptr)
1495 *pSimdBaseType = simdBaseType;
1497 // Also indicate that we use floating point registers.
1498 compFloatingPointUsed = true;
1502 #endif // FEATURE_SIMD
1504 // Fetch GC layout info if requested
1505 if (gcLayout != nullptr)
1507 unsigned numGCVars = info.compCompHnd->getClassGClayout(structHnd, gcLayout);
1509 // Verify that the quick test up above via the class attributes gave a
1510 // safe view of the type's GCness.
1512 // Note there are cases where mayContainGCPtrs is true but getClassGClayout
1513 // does not report any gc fields.
1515 assert(mayContainGCPtrs || (numGCVars == 0));
1517 if (pNumGCVars != nullptr)
1519 *pNumGCVars = numGCVars;
1524 // Can't safely ask for number of GC pointers without also
1525 // asking for layout.
1526 assert(pNumGCVars == nullptr);
1532 //****************************************************************************
1533 // Given TYP_STRUCT value 'structVal', make sure it is 'canonical', that is
1534 // it is either an OBJ or a MKREFANY node, or a node (e.g. GT_INDEX) that will be morphed.
1536 GenTree* Compiler::impNormStructVal(GenTree* structVal,
1537 CORINFO_CLASS_HANDLE structHnd,
1539 bool forceNormalization /*=false*/)
1541 assert(forceNormalization || varTypeIsStruct(structVal));
1542 assert(structHnd != NO_CLASS_HANDLE);
1543 var_types structType = structVal->TypeGet();
1544 bool makeTemp = false;
1545 if (structType == TYP_STRUCT)
1547 structType = impNormStructType(structHnd);
1549 bool alreadyNormalized = false;
1550 GenTreeLclVarCommon* structLcl = nullptr;
1552 genTreeOps oper = structVal->OperGet();
1555 // GT_RETURN and GT_MKREFANY don't capture the handle.
1559 alreadyNormalized = true;
1563 structVal->gtCall.gtRetClsHnd = structHnd;
1568 structVal->gtRetExpr.gtRetClsHnd = structHnd;
1573 structVal->gtArgPlace.gtArgPlaceClsHnd = structHnd;
1577 // This will be transformed to an OBJ later.
1578 alreadyNormalized = true;
1579 structVal->gtIndex.gtStructElemClass = structHnd;
1580 structVal->gtIndex.gtIndElemSize = info.compCompHnd->getClassSize(structHnd);
1584 // Wrap it in a GT_OBJ.
1585 structVal->gtType = structType;
1586 structVal = gtNewObjNode(structHnd, gtNewOperNode(GT_ADDR, TYP_BYREF, structVal));
1591 structLcl = structVal->AsLclVarCommon();
1592 // Wrap it in a GT_OBJ.
1593 structVal = gtNewObjNode(structHnd, gtNewOperNode(GT_ADDR, TYP_BYREF, structVal));
1600 // These should already have the appropriate type.
1601 assert(structVal->gtType == structType);
1602 alreadyNormalized = true;
1606 assert(structVal->gtType == structType);
1607 structVal = gtNewObjNode(structHnd, structVal->gtGetOp1());
1608 alreadyNormalized = true;
1613 assert(varTypeIsSIMD(structVal) && (structVal->gtType == structType));
1615 #endif // FEATURE_SIMD
1616 #ifdef FEATURE_HW_INTRINSICS
1617 case GT_HWIntrinsic:
1618 assert(varTypeIsSIMD(structVal) && (structVal->gtType == structType));
1624 // The second thing could either be a block node or a GT_FIELD or a GT_SIMD or a GT_COMMA node.
1625 GenTree* blockNode = structVal->gtOp.gtOp2;
1626 assert(blockNode->gtType == structType);
1628 // Is this GT_COMMA(op1, GT_COMMA())?
1629 GenTree* parent = structVal;
1630 if (blockNode->OperGet() == GT_COMMA)
1632 // Find the last node in the comma chain.
1635 assert(blockNode->gtType == structType);
1637 blockNode = blockNode->gtOp.gtOp2;
1638 } while (blockNode->OperGet() == GT_COMMA);
1641 if (blockNode->OperGet() == GT_FIELD)
1643 // If we have a GT_FIELD then wrap it in a GT_OBJ.
1644 blockNode = gtNewObjNode(structHnd, gtNewOperNode(GT_ADDR, TYP_BYREF, blockNode));
1648 if (blockNode->OperIsSIMDorSimdHWintrinsic())
1650 parent->gtOp.gtOp2 = impNormStructVal(blockNode, structHnd, curLevel, forceNormalization);
1651 alreadyNormalized = true;
1656 noway_assert(blockNode->OperIsBlk());
1658 // Sink the GT_COMMA below the blockNode addr.
1659 // That is GT_COMMA(op1, op2=blockNode) is tranformed into
1660 // blockNode(GT_COMMA(TYP_BYREF, op1, op2's op1)).
1662 // In case of a chained GT_COMMA case, we sink the last
1663 // GT_COMMA below the blockNode addr.
1664 GenTree* blockNodeAddr = blockNode->gtOp.gtOp1;
1665 assert(blockNodeAddr->gtType == TYP_BYREF);
1666 GenTree* commaNode = parent;
1667 commaNode->gtType = TYP_BYREF;
1668 commaNode->gtOp.gtOp2 = blockNodeAddr;
1669 blockNode->gtOp.gtOp1 = commaNode;
1670 if (parent == structVal)
1672 structVal = blockNode;
1674 alreadyNormalized = true;
1680 noway_assert(!"Unexpected node in impNormStructVal()");
1683 structVal->gtType = structType;
1684 GenTree* structObj = structVal;
1686 if (!alreadyNormalized || forceNormalization)
1690 unsigned tmpNum = lvaGrabTemp(true DEBUGARG("struct address for call/obj"));
1692 impAssignTempGen(tmpNum, structVal, structHnd, curLevel);
1694 // The structVal is now the temp itself
1696 structLcl = gtNewLclvNode(tmpNum, structType)->AsLclVarCommon();
1697 // TODO-1stClassStructs: Avoid always wrapping in GT_OBJ.
1698 structObj = gtNewObjNode(structHnd, gtNewOperNode(GT_ADDR, TYP_BYREF, structLcl));
1700 else if (varTypeIsStruct(structType) && !structVal->OperIsBlk())
1702 // Wrap it in a GT_OBJ
1703 structObj = gtNewObjNode(structHnd, gtNewOperNode(GT_ADDR, TYP_BYREF, structVal));
1707 if (structLcl != nullptr)
1709 // A OBJ on a ADDR(LCL_VAR) can never raise an exception
1710 // so we don't set GTF_EXCEPT here.
1711 if (!lvaIsImplicitByRefLocal(structLcl->gtLclNum))
1713 structObj->gtFlags &= ~GTF_GLOB_REF;
1718 // In general a OBJ is an indirection and could raise an exception.
1719 structObj->gtFlags |= GTF_EXCEPT;
1724 /******************************************************************************/
1725 // Given a type token, generate code that will evaluate to the correct
1726 // handle representation of that token (type handle, field handle, or method handle)
1728 // For most cases, the handle is determined at compile-time, and the code
1729 // generated is simply an embedded handle.
1731 // Run-time lookup is required if the enclosing method is shared between instantiations
1732 // and the token refers to formal type parameters whose instantiation is not known
1735 GenTree* Compiler::impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
1736 BOOL* pRuntimeLookup /* = NULL */,
1737 BOOL mustRestoreHandle /* = FALSE */,
1738 BOOL importParent /* = FALSE */)
1740 assert(!fgGlobalMorph);
1742 CORINFO_GENERICHANDLE_RESULT embedInfo;
1743 info.compCompHnd->embedGenericHandle(pResolvedToken, importParent, &embedInfo);
1747 *pRuntimeLookup = embedInfo.lookup.lookupKind.needsRuntimeLookup;
1750 if (mustRestoreHandle && !embedInfo.lookup.lookupKind.needsRuntimeLookup)
1752 switch (embedInfo.handleType)
1754 case CORINFO_HANDLETYPE_CLASS:
1755 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun((CORINFO_CLASS_HANDLE)embedInfo.compileTimeHandle);
1758 case CORINFO_HANDLETYPE_METHOD:
1759 info.compCompHnd->methodMustBeLoadedBeforeCodeIsRun((CORINFO_METHOD_HANDLE)embedInfo.compileTimeHandle);
1762 case CORINFO_HANDLETYPE_FIELD:
1763 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(
1764 info.compCompHnd->getFieldClass((CORINFO_FIELD_HANDLE)embedInfo.compileTimeHandle));
1772 // Generate the full lookup tree. May be null if we're abandoning an inline attempt.
1773 GenTree* result = impLookupToTree(pResolvedToken, &embedInfo.lookup, gtTokenToIconFlags(pResolvedToken->token),
1774 embedInfo.compileTimeHandle);
1776 // If we have a result and it requires runtime lookup, wrap it in a runtime lookup node.
1777 if ((result != nullptr) && embedInfo.lookup.lookupKind.needsRuntimeLookup)
1779 result = gtNewRuntimeLookup(embedInfo.compileTimeHandle, embedInfo.handleType, result);
1785 GenTree* Compiler::impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
1786 CORINFO_LOOKUP* pLookup,
1787 unsigned handleFlags,
1788 void* compileTimeHandle)
1790 if (!pLookup->lookupKind.needsRuntimeLookup)
1792 // No runtime lookup is required.
1793 // Access is direct or memory-indirect (of a fixed address) reference
1795 CORINFO_GENERIC_HANDLE handle = nullptr;
1796 void* pIndirection = nullptr;
1797 assert(pLookup->constLookup.accessType != IAT_PPVALUE);
1799 if (pLookup->constLookup.accessType == IAT_VALUE)
1801 handle = pLookup->constLookup.handle;
1803 else if (pLookup->constLookup.accessType == IAT_PVALUE)
1805 pIndirection = pLookup->constLookup.addr;
1807 return gtNewIconEmbHndNode(handle, pIndirection, handleFlags, compileTimeHandle);
1809 else if (compIsForInlining())
1811 // Don't import runtime lookups when inlining
1812 // Inlining has to be aborted in such a case
1813 compInlineResult->NoteFatal(InlineObservation::CALLSITE_GENERIC_DICTIONARY_LOOKUP);
1818 // Need to use dictionary-based access which depends on the typeContext
1819 // which is only available at runtime, not at compile-time.
1821 return impRuntimeLookupToTree(pResolvedToken, pLookup, compileTimeHandle);
1825 #ifdef FEATURE_READYTORUN_COMPILER
1826 GenTree* Compiler::impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup,
1827 unsigned handleFlags,
1828 void* compileTimeHandle)
1830 CORINFO_GENERIC_HANDLE handle = nullptr;
1831 void* pIndirection = nullptr;
1832 assert(pLookup->accessType != IAT_PPVALUE);
1834 if (pLookup->accessType == IAT_VALUE)
1836 handle = pLookup->handle;
1838 else if (pLookup->accessType == IAT_PVALUE)
1840 pIndirection = pLookup->addr;
1842 return gtNewIconEmbHndNode(handle, pIndirection, handleFlags, compileTimeHandle);
1845 GenTreeCall* Compiler::impReadyToRunHelperToTree(
1846 CORINFO_RESOLVED_TOKEN* pResolvedToken,
1847 CorInfoHelpFunc helper,
1849 GenTreeArgList* args /* =NULL*/,
1850 CORINFO_LOOKUP_KIND* pGenericLookupKind /* =NULL. Only used with generics */)
1852 CORINFO_CONST_LOOKUP lookup;
1853 if (!info.compCompHnd->getReadyToRunHelper(pResolvedToken, pGenericLookupKind, helper, &lookup))
1858 GenTreeCall* op1 = gtNewHelperCallNode(helper, type, args);
1860 op1->setEntryPoint(lookup);
1866 GenTree* Compiler::impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo)
1868 GenTree* op1 = nullptr;
1870 switch (pCallInfo->kind)
1873 op1 = new (this, GT_FTN_ADDR) GenTreeFptrVal(TYP_I_IMPL, pCallInfo->hMethod);
1875 #ifdef FEATURE_READYTORUN_COMPILER
1876 if (opts.IsReadyToRun())
1878 op1->gtFptrVal.gtEntryPoint = pCallInfo->codePointerLookup.constLookup;
1882 op1->gtFptrVal.gtEntryPoint.addr = nullptr;
1883 op1->gtFptrVal.gtEntryPoint.accessType = IAT_VALUE;
1888 case CORINFO_CALL_CODE_POINTER:
1889 if (compIsForInlining())
1891 // Don't import runtime lookups when inlining
1892 // Inlining has to be aborted in such a case
1893 compInlineResult->NoteFatal(InlineObservation::CALLSITE_GENERIC_DICTIONARY_LOOKUP);
1897 op1 = impLookupToTree(pResolvedToken, &pCallInfo->codePointerLookup, GTF_ICON_FTN_ADDR, pCallInfo->hMethod);
1901 noway_assert(!"unknown call kind");
1908 //------------------------------------------------------------------------
1909 // getRuntimeContextTree: find pointer to context for runtime lookup.
1912 // kind - lookup kind.
1915 // Return GenTree pointer to generic shared context.
1918 // Reports about generic context using.
1920 GenTree* Compiler::getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind)
1922 GenTree* ctxTree = nullptr;
1924 // Collectible types requires that for shared generic code, if we use the generic context parameter
1925 // that we report it. (This is a conservative approach, we could detect some cases particularly when the
1926 // context parameter is this that we don't need the eager reporting logic.)
1927 lvaGenericsContextUseCount++;
1929 if (kind == CORINFO_LOOKUP_THISOBJ)
1932 ctxTree = gtNewLclvNode(info.compThisArg, TYP_REF);
1934 // Vtable pointer of this object
1935 ctxTree = gtNewOperNode(GT_IND, TYP_I_IMPL, ctxTree);
1936 ctxTree->gtFlags |= GTF_EXCEPT; // Null-pointer exception
1937 ctxTree->gtFlags |= GTF_IND_INVARIANT;
1941 assert(kind == CORINFO_LOOKUP_METHODPARAM || kind == CORINFO_LOOKUP_CLASSPARAM);
1943 ctxTree = gtNewLclvNode(info.compTypeCtxtArg, TYP_I_IMPL); // Exact method descriptor as passed in as last arg
1948 /*****************************************************************************/
1949 /* Import a dictionary lookup to access a handle in code shared between
1950 generic instantiations.
1951 The lookup depends on the typeContext which is only available at
1952 runtime, and not at compile-time.
1953 pLookup->token1 and pLookup->token2 specify the handle that is needed.
1956 1. pLookup->indirections == CORINFO_USEHELPER : Call a helper passing it the
1957 instantiation-specific handle, and the tokens to lookup the handle.
1958 2. pLookup->indirections != CORINFO_USEHELPER :
1959 2a. pLookup->testForNull == false : Dereference the instantiation-specific handle
1961 2b. pLookup->testForNull == true : Dereference the instantiation-specific handle.
1962 If it is non-NULL, it is the handle required. Else, call a helper
1963 to lookup the handle.
1966 GenTree* Compiler::impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
1967 CORINFO_LOOKUP* pLookup,
1968 void* compileTimeHandle)
1971 // This method can only be called from the importer instance of the Compiler.
1972 // In other word, it cannot be called by the instance of the Compiler for the inlinee.
1973 assert(!compIsForInlining());
1975 GenTree* ctxTree = getRuntimeContextTree(pLookup->lookupKind.runtimeLookupKind);
1977 CORINFO_RUNTIME_LOOKUP* pRuntimeLookup = &pLookup->runtimeLookup;
1978 // It's available only via the run-time helper function
1979 if (pRuntimeLookup->indirections == CORINFO_USEHELPER)
1981 #ifdef FEATURE_READYTORUN_COMPILER
1982 if (opts.IsReadyToRun())
1984 return impReadyToRunHelperToTree(pResolvedToken, CORINFO_HELP_READYTORUN_GENERIC_HANDLE, TYP_I_IMPL,
1985 gtNewArgList(ctxTree), &pLookup->lookupKind);
1989 gtNewIconEmbHndNode(pRuntimeLookup->signature, nullptr, GTF_ICON_TOKEN_HDL, compileTimeHandle);
1990 GenTreeArgList* helperArgs = gtNewArgList(ctxTree, argNode);
1992 return gtNewHelperCallNode(pRuntimeLookup->helper, TYP_I_IMPL, helperArgs);
1996 GenTree* slotPtrTree = ctxTree;
1998 if (pRuntimeLookup->testForNull)
2000 slotPtrTree = impCloneExpr(ctxTree, &ctxTree, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
2001 nullptr DEBUGARG("impRuntimeLookup slot"));
2004 GenTree* indOffTree = nullptr;
2006 // Applied repeated indirections
2007 for (WORD i = 0; i < pRuntimeLookup->indirections; i++)
2009 if ((i == 1 && pRuntimeLookup->indirectFirstOffset) || (i == 2 && pRuntimeLookup->indirectSecondOffset))
2011 indOffTree = impCloneExpr(slotPtrTree, &slotPtrTree, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
2012 nullptr DEBUGARG("impRuntimeLookup indirectOffset"));
2017 slotPtrTree = gtNewOperNode(GT_IND, TYP_I_IMPL, slotPtrTree);
2018 slotPtrTree->gtFlags |= GTF_IND_NONFAULTING;
2019 slotPtrTree->gtFlags |= GTF_IND_INVARIANT;
2022 if ((i == 1 && pRuntimeLookup->indirectFirstOffset) || (i == 2 && pRuntimeLookup->indirectSecondOffset))
2024 slotPtrTree = gtNewOperNode(GT_ADD, TYP_I_IMPL, indOffTree, slotPtrTree);
2027 if (pRuntimeLookup->offsets[i] != 0)
2030 gtNewOperNode(GT_ADD, TYP_I_IMPL, slotPtrTree, gtNewIconNode(pRuntimeLookup->offsets[i], TYP_I_IMPL));
2034 // No null test required
2035 if (!pRuntimeLookup->testForNull)
2037 if (pRuntimeLookup->indirections == 0)
2042 slotPtrTree = gtNewOperNode(GT_IND, TYP_I_IMPL, slotPtrTree);
2043 slotPtrTree->gtFlags |= GTF_IND_NONFAULTING;
2045 if (!pRuntimeLookup->testForFixup)
2050 impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("bubbling QMark0"));
2052 unsigned slotLclNum = lvaGrabTemp(true DEBUGARG("impRuntimeLookup test"));
2053 impAssignTempGen(slotLclNum, slotPtrTree, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL, nullptr, impCurStmtOffs);
2055 GenTree* slot = gtNewLclvNode(slotLclNum, TYP_I_IMPL);
2056 // downcast the pointer to a TYP_INT on 64-bit targets
2057 slot = impImplicitIorI4Cast(slot, TYP_INT);
2058 // Use a GT_AND to check for the lowest bit and indirect if it is set
2059 GenTree* test = gtNewOperNode(GT_AND, TYP_INT, slot, gtNewIconNode(1));
2060 GenTree* relop = gtNewOperNode(GT_EQ, TYP_INT, test, gtNewIconNode(0));
2061 relop->gtFlags |= GTF_RELOP_QMARK;
2063 // slot = GT_IND(slot - 1)
2064 slot = gtNewLclvNode(slotLclNum, TYP_I_IMPL);
2065 GenTree* add = gtNewOperNode(GT_ADD, TYP_I_IMPL, slot, gtNewIconNode(-1, TYP_I_IMPL));
2066 GenTree* indir = gtNewOperNode(GT_IND, TYP_I_IMPL, add);
2067 indir->gtFlags |= GTF_IND_NONFAULTING;
2068 indir->gtFlags |= GTF_IND_INVARIANT;
2070 slot = gtNewLclvNode(slotLclNum, TYP_I_IMPL);
2071 GenTree* asg = gtNewAssignNode(slot, indir);
2072 GenTree* colon = new (this, GT_COLON) GenTreeColon(TYP_VOID, gtNewNothingNode(), asg);
2073 GenTree* qmark = gtNewQmarkNode(TYP_VOID, relop, colon);
2074 impAppendTree(qmark, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
2076 return gtNewLclvNode(slotLclNum, TYP_I_IMPL);
2079 assert(pRuntimeLookup->indirections != 0);
2081 impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("bubbling QMark1"));
2083 // Extract the handle
2084 GenTree* handle = gtNewOperNode(GT_IND, TYP_I_IMPL, slotPtrTree);
2085 handle->gtFlags |= GTF_IND_NONFAULTING;
2087 GenTree* handleCopy = impCloneExpr(handle, &handle, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
2088 nullptr DEBUGARG("impRuntimeLookup typehandle"));
2091 GenTree* argNode = gtNewIconEmbHndNode(pRuntimeLookup->signature, nullptr, GTF_ICON_TOKEN_HDL, compileTimeHandle);
2093 GenTreeArgList* helperArgs = gtNewArgList(ctxTree, argNode);
2094 GenTree* helperCall = gtNewHelperCallNode(pRuntimeLookup->helper, TYP_I_IMPL, helperArgs);
2096 // Check for null and possibly call helper
2097 GenTree* relop = gtNewOperNode(GT_NE, TYP_INT, handle, gtNewIconNode(0, TYP_I_IMPL));
2098 relop->gtFlags |= GTF_RELOP_QMARK;
2100 GenTree* colon = new (this, GT_COLON) GenTreeColon(TYP_I_IMPL,
2101 gtNewNothingNode(), // do nothing if nonnull
2104 GenTree* qmark = gtNewQmarkNode(TYP_I_IMPL, relop, colon);
2107 if (handleCopy->IsLocal())
2109 tmp = handleCopy->gtLclVarCommon.gtLclNum;
2113 tmp = lvaGrabTemp(true DEBUGARG("spilling QMark1"));
2116 impAssignTempGen(tmp, qmark, (unsigned)CHECK_SPILL_NONE);
2117 return gtNewLclvNode(tmp, TYP_I_IMPL);
2120 /******************************************************************************
2121 * Spills the stack at verCurrentState.esStack[level] and replaces it with a temp.
2122 * If tnum!=BAD_VAR_NUM, the temp var used to replace the tree is tnum,
2123 * else, grab a new temp.
2124 * For structs (which can be pushed on the stack using obj, etc),
2125 * special handling is needed
2128 struct RecursiveGuard
2133 m_pAddress = nullptr;
2140 *m_pAddress = false;
2144 void Init(bool* pAddress, bool bInitialize)
2146 assert(pAddress && *pAddress == false && "Recursive guard violation");
2147 m_pAddress = pAddress;
2159 bool Compiler::impSpillStackEntry(unsigned level,
2163 bool bAssertOnRecursion,
2170 RecursiveGuard guard;
2171 guard.Init(&impNestedStackSpill, bAssertOnRecursion);
2174 GenTree* tree = verCurrentState.esStack[level].val;
2176 /* Allocate a temp if we haven't been asked to use a particular one */
2178 if (tiVerificationNeeded)
2180 // Ignore bad temp requests (they will happen with bad code and will be
2181 // catched when importing the destblock)
2182 if ((tnum != BAD_VAR_NUM && tnum >= lvaCount) && verNeedsVerification())
2189 if (tnum != BAD_VAR_NUM && (tnum >= lvaCount))
2195 bool isNewTemp = false;
2197 if (tnum == BAD_VAR_NUM)
2199 tnum = lvaGrabTemp(true DEBUGARG(reason));
2202 else if (tiVerificationNeeded && lvaTable[tnum].TypeGet() != TYP_UNDEF)
2204 // if verification is needed and tnum's type is incompatible with
2205 // type on that stack, we grab a new temp. This is safe since
2206 // we will throw a verification exception in the dest block.
2208 var_types valTyp = tree->TypeGet();
2209 var_types dstTyp = lvaTable[tnum].TypeGet();
2211 // if the two types are different, we return. This will only happen with bad code and will
2212 // be catched when importing the destblock. We still allow int/byrefs and float/double differences.
2213 if ((genActualType(valTyp) != genActualType(dstTyp)) &&
2215 #ifndef _TARGET_64BIT_
2216 (valTyp == TYP_I_IMPL && dstTyp == TYP_BYREF) || (valTyp == TYP_BYREF && dstTyp == TYP_I_IMPL) ||
2217 #endif // !_TARGET_64BIT_
2218 (varTypeIsFloating(dstTyp) && varTypeIsFloating(valTyp))))
2220 if (verNeedsVerification())
2227 /* Assign the spilled entry to the temp */
2228 impAssignTempGen(tnum, tree, verCurrentState.esStack[level].seTypeInfo.GetClassHandle(), level);
2230 // If temp is newly introduced and a ref type, grab what type info we can.
2231 if (isNewTemp && (lvaTable[tnum].lvType == TYP_REF))
2233 CORINFO_CLASS_HANDLE stkHnd = verCurrentState.esStack[level].seTypeInfo.GetClassHandle();
2234 lvaSetClass(tnum, tree, stkHnd);
2237 // The tree type may be modified by impAssignTempGen, so use the type of the lclVar.
2238 var_types type = genActualType(lvaTable[tnum].TypeGet());
2239 GenTree* temp = gtNewLclvNode(tnum, type);
2240 verCurrentState.esStack[level].val = temp;
2245 /*****************************************************************************
2247 * Ensure that the stack has only spilled values
2250 void Compiler::impSpillStackEnsure(bool spillLeaves)
2252 assert(!spillLeaves || opts.compDbgCode);
2254 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
2256 GenTree* tree = verCurrentState.esStack[level].val;
2258 if (!spillLeaves && tree->OperIsLeaf())
2263 // Temps introduced by the importer itself don't need to be spilled
2265 bool isTempLcl = (tree->OperGet() == GT_LCL_VAR) && (tree->gtLclVarCommon.gtLclNum >= info.compLocalsCount);
2272 impSpillStackEntry(level, BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impSpillStackEnsure"));
2276 void Compiler::impSpillEvalStack()
2278 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
2280 impSpillStackEntry(level, BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impSpillEvalStack"));
2284 /*****************************************************************************
2286 * If the stack contains any trees with side effects in them, assign those
2287 * trees to temps and append the assignments to the statement list.
2288 * On return the stack is guaranteed to be empty.
2291 inline void Compiler::impEvalSideEffects()
2293 impSpillSideEffects(false, (unsigned)CHECK_SPILL_ALL DEBUGARG("impEvalSideEffects"));
2294 verCurrentState.esStackDepth = 0;
2297 /*****************************************************************************
2299 * If the stack contains any trees with side effects in them, assign those
2300 * trees to temps and replace them on the stack with refs to their temps.
2301 * [0..chkLevel) is the portion of the stack which will be checked and spilled.
2304 inline void Compiler::impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason))
2306 assert(chkLevel != (unsigned)CHECK_SPILL_NONE);
2308 /* Before we make any appends to the tree list we must spill the
2309 * "special" side effects (GTF_ORDER_SIDEEFF on a GT_CATCH_ARG) */
2311 impSpillSpecialSideEff();
2313 if (chkLevel == (unsigned)CHECK_SPILL_ALL)
2315 chkLevel = verCurrentState.esStackDepth;
2318 assert(chkLevel <= verCurrentState.esStackDepth);
2320 unsigned spillFlags = spillGlobEffects ? GTF_GLOB_EFFECT : GTF_SIDE_EFFECT;
2322 for (unsigned i = 0; i < chkLevel; i++)
2324 GenTree* tree = verCurrentState.esStack[i].val;
2326 GenTree* lclVarTree;
2328 if ((tree->gtFlags & spillFlags) != 0 ||
2329 (spillGlobEffects && // Only consider the following when spillGlobEffects == TRUE
2330 !impIsAddressInLocal(tree, &lclVarTree) && // No need to spill the GT_ADDR node on a local.
2331 gtHasLocalsWithAddrOp(tree))) // Spill if we still see GT_LCL_VAR that contains lvHasLdAddrOp or
2332 // lvAddrTaken flag.
2334 impSpillStackEntry(i, BAD_VAR_NUM DEBUGARG(false) DEBUGARG(reason));
2339 /*****************************************************************************
2341 * If the stack contains any trees with special side effects in them, assign
2342 * those trees to temps and replace them on the stack with refs to their temps.
2345 inline void Compiler::impSpillSpecialSideEff()
2347 // Only exception objects need to be carefully handled
2349 if (!compCurBB->bbCatchTyp)
2354 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
2356 GenTree* tree = verCurrentState.esStack[level].val;
2357 // Make sure if we have an exception object in the sub tree we spill ourselves.
2358 if (gtHasCatchArg(tree))
2360 impSpillStackEntry(level, BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impSpillSpecialSideEff"));
2365 /*****************************************************************************
2367 * Spill all stack references to value classes (TYP_STRUCT nodes)
2370 void Compiler::impSpillValueClasses()
2372 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
2374 GenTree* tree = verCurrentState.esStack[level].val;
2376 if (fgWalkTreePre(&tree, impFindValueClasses) == WALK_ABORT)
2378 // Tree walk was aborted, which means that we found a
2379 // value class on the stack. Need to spill that
2382 impSpillStackEntry(level, BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impSpillValueClasses"));
2387 /*****************************************************************************
2389 * Callback that checks if a tree node is TYP_STRUCT
2392 Compiler::fgWalkResult Compiler::impFindValueClasses(GenTree** pTree, fgWalkData* data)
2394 fgWalkResult walkResult = WALK_CONTINUE;
2396 if ((*pTree)->gtType == TYP_STRUCT)
2398 // Abort the walk and indicate that we found a value class
2400 walkResult = WALK_ABORT;
2406 /*****************************************************************************
2408 * If the stack contains any trees with references to local #lclNum, assign
2409 * those trees to temps and replace their place on the stack with refs to
2413 void Compiler::impSpillLclRefs(ssize_t lclNum)
2415 /* Before we make any appends to the tree list we must spill the
2416 * "special" side effects (GTF_ORDER_SIDEEFF) - GT_CATCH_ARG */
2418 impSpillSpecialSideEff();
2420 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
2422 GenTree* tree = verCurrentState.esStack[level].val;
2424 /* If the tree may throw an exception, and the block has a handler,
2425 then we need to spill assignments to the local if the local is
2426 live on entry to the handler.
2427 Just spill 'em all without considering the liveness */
2429 bool xcptnCaught = ehBlockHasExnFlowDsc(compCurBB) && (tree->gtFlags & (GTF_CALL | GTF_EXCEPT));
2431 /* Skip the tree if it doesn't have an affected reference,
2432 unless xcptnCaught */
2434 if (xcptnCaught || gtHasRef(tree, lclNum, false))
2436 impSpillStackEntry(level, BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impSpillLclRefs"));
2441 /*****************************************************************************
2443 * Push catch arg onto the stack.
2444 * If there are jumps to the beginning of the handler, insert basic block
2445 * and spill catch arg to a temp. Update the handler block if necessary.
2447 * Returns the basic block of the actual handler.
2450 BasicBlock* Compiler::impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter)
2452 // Do not inject the basic block twice on reimport. This should be
2453 // hit only under JIT stress. See if the block is the one we injected.
2454 // Note that EH canonicalization can inject internal blocks here. We might
2455 // be able to re-use such a block (but we don't, right now).
2456 if ((hndBlk->bbFlags & (BBF_IMPORTED | BBF_INTERNAL | BBF_DONT_REMOVE | BBF_HAS_LABEL | BBF_JMP_TARGET)) ==
2457 (BBF_IMPORTED | BBF_INTERNAL | BBF_DONT_REMOVE | BBF_HAS_LABEL | BBF_JMP_TARGET))
2459 GenTree* tree = hndBlk->bbTreeList;
2461 if (tree != nullptr && tree->gtOper == GT_STMT)
2463 tree = tree->gtStmt.gtStmtExpr;
2464 assert(tree != nullptr);
2466 if ((tree->gtOper == GT_ASG) && (tree->gtOp.gtOp1->gtOper == GT_LCL_VAR) &&
2467 (tree->gtOp.gtOp2->gtOper == GT_CATCH_ARG))
2469 tree = gtNewLclvNode(tree->gtOp.gtOp1->gtLclVarCommon.gtLclNum, TYP_REF);
2471 impPushOnStack(tree, typeInfo(TI_REF, clsHnd));
2473 return hndBlk->bbNext;
2477 // If we get here, it must have been some other kind of internal block. It's possible that
2478 // someone prepended something to our injected block, but that's unlikely.
2481 /* Push the exception address value on the stack */
2482 GenTree* arg = new (this, GT_CATCH_ARG) GenTree(GT_CATCH_ARG, TYP_REF);
2484 /* Mark the node as having a side-effect - i.e. cannot be
2485 * moved around since it is tied to a fixed location (EAX) */
2486 arg->gtFlags |= GTF_ORDER_SIDEEFF;
2488 #if defined(JIT32_GCENCODER)
2489 const bool forceInsertNewBlock = isSingleBlockFilter || compStressCompile(STRESS_CATCH_ARG, 5);
2491 const bool forceInsertNewBlock = compStressCompile(STRESS_CATCH_ARG, 5);
2492 #endif // defined(JIT32_GCENCODER)
2494 /* Spill GT_CATCH_ARG to a temp if there are jumps to the beginning of the handler */
2495 if (hndBlk->bbRefs > 1 || forceInsertNewBlock)
2497 if (hndBlk->bbRefs == 1)
2502 /* Create extra basic block for the spill */
2503 BasicBlock* newBlk = fgNewBBbefore(BBJ_NONE, hndBlk, /* extendRegion */ true);
2504 newBlk->bbFlags |= BBF_IMPORTED | BBF_DONT_REMOVE | BBF_HAS_LABEL | BBF_JMP_TARGET;
2505 newBlk->setBBWeight(hndBlk->bbWeight);
2506 newBlk->bbCodeOffs = hndBlk->bbCodeOffs;
2508 /* Account for the new link we are about to create */
2511 /* Spill into a temp */
2512 unsigned tempNum = lvaGrabTemp(false DEBUGARG("SpillCatchArg"));
2513 lvaTable[tempNum].lvType = TYP_REF;
2514 arg = gtNewTempAssign(tempNum, arg);
2516 hndBlk->bbStkTempsIn = tempNum;
2518 /* Report the debug info. impImportBlockCode won't treat
2519 * the actual handler as exception block and thus won't do it for us. */
2520 if (info.compStmtOffsetsImplicit & ICorDebugInfo::CALL_SITE_BOUNDARIES)
2522 impCurStmtOffs = newBlk->bbCodeOffs | IL_OFFSETX_STKBIT;
2523 arg = gtNewStmt(arg, impCurStmtOffs);
2526 fgInsertStmtAtEnd(newBlk, arg);
2528 arg = gtNewLclvNode(tempNum, TYP_REF);
2531 impPushOnStack(arg, typeInfo(TI_REF, clsHnd));
2536 /*****************************************************************************
2538 * Given a tree, clone it. *pClone is set to the cloned tree.
2539 * Returns the original tree if the cloning was easy,
2540 * else returns the temp to which the tree had to be spilled to.
2541 * If the tree has side-effects, it will be spilled to a temp.
2544 GenTree* Compiler::impCloneExpr(GenTree* tree,
2546 CORINFO_CLASS_HANDLE structHnd,
2548 GenTree** pAfterStmt DEBUGARG(const char* reason))
2550 if (!(tree->gtFlags & GTF_GLOB_EFFECT))
2552 GenTree* clone = gtClone(tree, true);
2561 /* Store the operand in a temp and return the temp */
2563 unsigned temp = lvaGrabTemp(true DEBUGARG(reason));
2565 // impAssignTempGen() may change tree->gtType to TYP_VOID for calls which
2566 // return a struct type. It also may modify the struct type to a more
2567 // specialized type (e.g. a SIMD type). So we will get the type from
2568 // the lclVar AFTER calling impAssignTempGen().
2570 impAssignTempGen(temp, tree, structHnd, curLevel, pAfterStmt, impCurStmtOffs);
2571 var_types type = genActualType(lvaTable[temp].TypeGet());
2573 *pClone = gtNewLclvNode(temp, type);
2574 return gtNewLclvNode(temp, type);
2577 /*****************************************************************************
2578 * Remember the IL offset (including stack-empty info) for the trees we will
2582 inline void Compiler::impCurStmtOffsSet(IL_OFFSET offs)
2584 if (compIsForInlining())
2586 GenTree* callStmt = impInlineInfo->iciStmt;
2587 assert(callStmt->gtOper == GT_STMT);
2588 impCurStmtOffs = callStmt->gtStmt.gtStmtILoffsx;
2592 assert(offs == BAD_IL_OFFSET || (offs & IL_OFFSETX_BITS) == 0);
2593 IL_OFFSETX stkBit = (verCurrentState.esStackDepth > 0) ? IL_OFFSETX_STKBIT : 0;
2594 impCurStmtOffs = offs | stkBit;
2598 /*****************************************************************************
2599 * Returns current IL offset with stack-empty and call-instruction info incorporated
2601 inline IL_OFFSETX Compiler::impCurILOffset(IL_OFFSET offs, bool callInstruction)
2603 if (compIsForInlining())
2605 return BAD_IL_OFFSET;
2609 assert(offs == BAD_IL_OFFSET || (offs & IL_OFFSETX_BITS) == 0);
2610 IL_OFFSETX stkBit = (verCurrentState.esStackDepth > 0) ? IL_OFFSETX_STKBIT : 0;
2611 IL_OFFSETX callInstructionBit = callInstruction ? IL_OFFSETX_CALLINSTRUCTIONBIT : 0;
2612 return offs | stkBit | callInstructionBit;
2616 //------------------------------------------------------------------------
2617 // impCanSpillNow: check is it possible to spill all values from eeStack to local variables.
2620 // prevOpcode - last importer opcode
2623 // true if it is legal, false if it could be a sequence that we do not want to divide.
2624 bool Compiler::impCanSpillNow(OPCODE prevOpcode)
2626 // Don't spill after ldtoken, newarr and newobj, because it could be a part of the InitializeArray sequence.
2627 // Avoid breaking up to guarantee that impInitializeArrayIntrinsic can succeed.
2628 return (prevOpcode != CEE_LDTOKEN) && (prevOpcode != CEE_NEWARR) && (prevOpcode != CEE_NEWOBJ);
2631 /*****************************************************************************
2633 * Remember the instr offset for the statements
2635 * When we do impAppendTree(tree), we can't set tree->gtStmtLastILoffs to
2636 * impCurOpcOffs, if the append was done because of a partial stack spill,
2637 * as some of the trees corresponding to code up to impCurOpcOffs might
2638 * still be sitting on the stack.
2639 * So we delay marking of gtStmtLastILoffs until impNoteLastILoffs().
2640 * This should be called when an opcode finally/explicitly causes
2641 * impAppendTree(tree) to be called (as opposed to being called because of
2642 * a spill caused by the opcode)
2647 void Compiler::impNoteLastILoffs()
2649 if (impLastILoffsStmt == nullptr)
2651 // We should have added a statement for the current basic block
2652 // Is this assert correct ?
2654 assert(impTreeLast);
2655 assert(impTreeLast->gtOper == GT_STMT);
2657 impTreeLast->gtStmt.gtStmtLastILoffs = compIsForInlining() ? BAD_IL_OFFSET : impCurOpcOffs;
2661 impLastILoffsStmt->gtStmt.gtStmtLastILoffs = compIsForInlining() ? BAD_IL_OFFSET : impCurOpcOffs;
2662 impLastILoffsStmt = nullptr;
2668 /*****************************************************************************
2669 * We don't create any GenTree (excluding spills) for a branch.
2670 * For debugging info, we need a placeholder so that we can note
2671 * the IL offset in gtStmt.gtStmtOffs. So append an empty statement.
2674 void Compiler::impNoteBranchOffs()
2676 if (opts.compDbgCode)
2678 impAppendTree(gtNewNothingNode(), (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
2682 /*****************************************************************************
2683 * Locate the next stmt boundary for which we need to record info.
2684 * We will have to spill the stack at such boundaries if it is not
2686 * Returns the next stmt boundary (after the start of the block)
2689 unsigned Compiler::impInitBlockLineInfo()
2691 /* Assume the block does not correspond with any IL offset. This prevents
2692 us from reporting extra offsets. Extra mappings can cause confusing
2693 stepping, especially if the extra mapping is a jump-target, and the
2694 debugger does not ignore extra mappings, but instead rewinds to the
2695 nearest known offset */
2697 impCurStmtOffsSet(BAD_IL_OFFSET);
2699 if (compIsForInlining())
2704 IL_OFFSET blockOffs = compCurBB->bbCodeOffs;
2706 if ((verCurrentState.esStackDepth == 0) && (info.compStmtOffsetsImplicit & ICorDebugInfo::STACK_EMPTY_BOUNDARIES))
2708 impCurStmtOffsSet(blockOffs);
2711 if (false && (info.compStmtOffsetsImplicit & ICorDebugInfo::CALL_SITE_BOUNDARIES))
2713 impCurStmtOffsSet(blockOffs);
2716 /* Always report IL offset 0 or some tests get confused.
2717 Probably a good idea anyways */
2721 impCurStmtOffsSet(blockOffs);
2724 if (!info.compStmtOffsetsCount)
2729 /* Find the lowest explicit stmt boundary within the block */
2731 /* Start looking at an entry that is based on our instr offset */
2733 unsigned index = (info.compStmtOffsetsCount * blockOffs) / info.compILCodeSize;
2735 if (index >= info.compStmtOffsetsCount)
2737 index = info.compStmtOffsetsCount - 1;
2740 /* If we've guessed too far, back up */
2742 while (index > 0 && info.compStmtOffsets[index - 1] >= blockOffs)
2747 /* If we guessed short, advance ahead */
2749 while (info.compStmtOffsets[index] < blockOffs)
2753 if (index == info.compStmtOffsetsCount)
2755 return info.compStmtOffsetsCount;
2759 assert(index < info.compStmtOffsetsCount);
2761 if (info.compStmtOffsets[index] == blockOffs)
2763 /* There is an explicit boundary for the start of this basic block.
2764 So we will start with bbCodeOffs. Else we will wait until we
2765 get to the next explicit boundary */
2767 impCurStmtOffsSet(blockOffs);
2775 /*****************************************************************************/
2777 static inline bool impOpcodeIsCallOpcode(OPCODE opcode)
2791 /*****************************************************************************/
2793 static inline bool impOpcodeIsCallSiteBoundary(OPCODE opcode)
2810 /*****************************************************************************/
2812 // One might think it is worth caching these values, but results indicate
2814 // In addition, caching them causes SuperPMI to be unable to completely
2815 // encapsulate an individual method context.
2816 CORINFO_CLASS_HANDLE Compiler::impGetRefAnyClass()
2818 CORINFO_CLASS_HANDLE refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
2819 assert(refAnyClass != (CORINFO_CLASS_HANDLE) nullptr);
2823 CORINFO_CLASS_HANDLE Compiler::impGetTypeHandleClass()
2825 CORINFO_CLASS_HANDLE typeHandleClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPE_HANDLE);
2826 assert(typeHandleClass != (CORINFO_CLASS_HANDLE) nullptr);
2827 return typeHandleClass;
2830 CORINFO_CLASS_HANDLE Compiler::impGetRuntimeArgumentHandle()
2832 CORINFO_CLASS_HANDLE argIteratorClass = info.compCompHnd->getBuiltinClass(CLASSID_ARGUMENT_HANDLE);
2833 assert(argIteratorClass != (CORINFO_CLASS_HANDLE) nullptr);
2834 return argIteratorClass;
2837 CORINFO_CLASS_HANDLE Compiler::impGetStringClass()
2839 CORINFO_CLASS_HANDLE stringClass = info.compCompHnd->getBuiltinClass(CLASSID_STRING);
2840 assert(stringClass != (CORINFO_CLASS_HANDLE) nullptr);
2844 CORINFO_CLASS_HANDLE Compiler::impGetObjectClass()
2846 CORINFO_CLASS_HANDLE objectClass = info.compCompHnd->getBuiltinClass(CLASSID_SYSTEM_OBJECT);
2847 assert(objectClass != (CORINFO_CLASS_HANDLE) nullptr);
2851 /*****************************************************************************
2852 * "&var" can be used either as TYP_BYREF or TYP_I_IMPL, but we
2853 * set its type to TYP_BYREF when we create it. We know if it can be
2854 * changed to TYP_I_IMPL only at the point where we use it
2858 void Compiler::impBashVarAddrsToI(GenTree* tree1, GenTree* tree2)
2860 if (tree1->IsVarAddr())
2862 tree1->gtType = TYP_I_IMPL;
2865 if (tree2 && tree2->IsVarAddr())
2867 tree2->gtType = TYP_I_IMPL;
2871 /*****************************************************************************
2872 * TYP_INT and TYP_I_IMPL can be used almost interchangeably, but we want
2873 * to make that an explicit cast in our trees, so any implicit casts that
2874 * exist in the IL (at least on 64-bit where TYP_I_IMPL != TYP_INT) are
2875 * turned into explicit casts here.
2876 * We also allow an implicit conversion of a ldnull into a TYP_I_IMPL(0)
2879 GenTree* Compiler::impImplicitIorI4Cast(GenTree* tree, var_types dstTyp)
2881 var_types currType = genActualType(tree->gtType);
2882 var_types wantedType = genActualType(dstTyp);
2884 if (wantedType != currType)
2886 // Automatic upcast for a GT_CNS_INT into TYP_I_IMPL
2887 if ((tree->OperGet() == GT_CNS_INT) && varTypeIsI(dstTyp))
2889 if (!varTypeIsI(tree->gtType) || ((tree->gtType == TYP_REF) && (tree->gtIntCon.gtIconVal == 0)))
2891 tree->gtType = TYP_I_IMPL;
2894 #ifdef _TARGET_64BIT_
2895 else if (varTypeIsI(wantedType) && (currType == TYP_INT))
2897 // Note that this allows TYP_INT to be cast to a TYP_I_IMPL when wantedType is a TYP_BYREF or TYP_REF
2898 tree = gtNewCastNode(TYP_I_IMPL, tree, false, TYP_I_IMPL);
2900 else if ((wantedType == TYP_INT) && varTypeIsI(currType))
2902 // Note that this allows TYP_BYREF or TYP_REF to be cast to a TYP_INT
2903 tree = gtNewCastNode(TYP_INT, tree, false, TYP_INT);
2905 #endif // _TARGET_64BIT_
2911 /*****************************************************************************
2912 * TYP_FLOAT and TYP_DOUBLE can be used almost interchangeably in some cases,
2913 * but we want to make that an explicit cast in our trees, so any implicit casts
2914 * that exist in the IL are turned into explicit casts here.
2917 GenTree* Compiler::impImplicitR4orR8Cast(GenTree* tree, var_types dstTyp)
2919 #ifndef LEGACY_BACKEND
2920 if (varTypeIsFloating(tree) && varTypeIsFloating(dstTyp) && (dstTyp != tree->gtType))
2922 tree = gtNewCastNode(dstTyp, tree, false, dstTyp);
2924 #endif // !LEGACY_BACKEND
2929 //------------------------------------------------------------------------
2930 // impInitializeArrayIntrinsic: Attempts to replace a call to InitializeArray
2931 // with a GT_COPYBLK node.
2934 // sig - The InitializeArray signature.
2937 // A pointer to the newly created GT_COPYBLK node if the replacement succeeds or
2938 // nullptr otherwise.
2941 // The function recognizes the following IL pattern:
2942 // ldc <length> or a list of ldc <lower bound>/<length>
2945 // ldtoken <field handle>
2946 // call InitializeArray
2947 // The lower bounds need not be constant except when the array rank is 1.
2948 // The function recognizes all kinds of arrays thus enabling a small runtime
2949 // such as CoreRT to skip providing an implementation for InitializeArray.
2951 GenTree* Compiler::impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig)
2953 assert(sig->numArgs == 2);
2955 GenTree* fieldTokenNode = impStackTop(0).val;
2956 GenTree* arrayLocalNode = impStackTop(1).val;
2959 // Verify that the field token is known and valid. Note that It's also
2960 // possible for the token to come from reflection, in which case we cannot do
2961 // the optimization and must therefore revert to calling the helper. You can
2962 // see an example of this in bvt\DynIL\initarray2.exe (in Main).
2965 // Check to see if the ldtoken helper call is what we see here.
2966 if (fieldTokenNode->gtOper != GT_CALL || (fieldTokenNode->gtCall.gtCallType != CT_HELPER) ||
2967 (fieldTokenNode->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_FIELDDESC_TO_STUBRUNTIMEFIELD)))
2972 // Strip helper call away
2973 fieldTokenNode = fieldTokenNode->gtCall.gtCallArgs->Current();
2975 if (fieldTokenNode->gtOper == GT_IND)
2977 fieldTokenNode = fieldTokenNode->gtOp.gtOp1;
2980 // Check for constant
2981 if (fieldTokenNode->gtOper != GT_CNS_INT)
2986 CORINFO_FIELD_HANDLE fieldToken = (CORINFO_FIELD_HANDLE)fieldTokenNode->gtIntCon.gtCompileTimeHandle;
2987 if (!fieldTokenNode->IsIconHandle(GTF_ICON_FIELD_HDL) || (fieldToken == nullptr))
2993 // We need to get the number of elements in the array and the size of each element.
2994 // We verify that the newarr statement is exactly what we expect it to be.
2995 // If it's not then we just return NULL and we don't optimize this call
2999 // It is possible the we don't have any statements in the block yet
3001 if (impTreeLast->gtOper != GT_STMT)
3003 assert(impTreeLast->gtOper == GT_BEG_STMTS);
3008 // We start by looking at the last statement, making sure it's an assignment, and
3009 // that the target of the assignment is the array passed to InitializeArray.
3011 GenTree* arrayAssignment = impTreeLast->gtStmt.gtStmtExpr;
3012 if ((arrayAssignment->gtOper != GT_ASG) || (arrayAssignment->gtOp.gtOp1->gtOper != GT_LCL_VAR) ||
3013 (arrayLocalNode->gtOper != GT_LCL_VAR) ||
3014 (arrayAssignment->gtOp.gtOp1->gtLclVarCommon.gtLclNum != arrayLocalNode->gtLclVarCommon.gtLclNum))
3020 // Make sure that the object being assigned is a helper call.
3023 GenTree* newArrayCall = arrayAssignment->gtOp.gtOp2;
3024 if ((newArrayCall->gtOper != GT_CALL) || (newArrayCall->gtCall.gtCallType != CT_HELPER))
3030 // Verify that it is one of the new array helpers.
3033 bool isMDArray = false;
3035 if (newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEWARR_1_DIRECT) &&
3036 newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEWARR_1_OBJ) &&
3037 newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEWARR_1_VC) &&
3038 newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEWARR_1_ALIGN8)
3039 #ifdef FEATURE_READYTORUN_COMPILER
3040 && newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEWARR_1_R2R_DIRECT) &&
3041 newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_READYTORUN_NEWARR_1)
3045 if (newArrayCall->gtCall.gtCallMethHnd != eeFindHelper(CORINFO_HELP_NEW_MDARR_NONVARARG))
3053 CORINFO_CLASS_HANDLE arrayClsHnd = (CORINFO_CLASS_HANDLE)newArrayCall->gtCall.compileTimeHelperArgumentHandle;
3056 // Make sure we found a compile time handle to the array
3065 S_UINT32 numElements;
3069 rank = info.compCompHnd->getArrayRank(arrayClsHnd);
3076 GenTreeArgList* tokenArg = newArrayCall->gtCall.gtCallArgs;
3077 assert(tokenArg != nullptr);
3078 GenTreeArgList* numArgsArg = tokenArg->Rest();
3079 assert(numArgsArg != nullptr);
3080 GenTreeArgList* argsArg = numArgsArg->Rest();
3081 assert(argsArg != nullptr);
3084 // The number of arguments should be a constant between 1 and 64. The rank can't be 0
3085 // so at least one length must be present and the rank can't exceed 32 so there can
3086 // be at most 64 arguments - 32 lengths and 32 lower bounds.
3089 if ((!numArgsArg->Current()->IsCnsIntOrI()) || (numArgsArg->Current()->AsIntCon()->IconValue() < 1) ||
3090 (numArgsArg->Current()->AsIntCon()->IconValue() > 64))
3095 unsigned numArgs = static_cast<unsigned>(numArgsArg->Current()->AsIntCon()->IconValue());
3096 bool lowerBoundsSpecified;
3098 if (numArgs == rank * 2)
3100 lowerBoundsSpecified = true;
3102 else if (numArgs == rank)
3104 lowerBoundsSpecified = false;
3107 // If the rank is 1 and a lower bound isn't specified then the runtime creates
3108 // a SDArray. Note that even if a lower bound is specified it can be 0 and then
3109 // we get a SDArray as well, see the for loop below.
3123 // The rank is known to be at least 1 so we can start with numElements being 1
3124 // to avoid the need to special case the first dimension.
3127 numElements = S_UINT32(1);
3131 static bool IsArgsFieldInit(GenTree* tree, unsigned index, unsigned lvaNewObjArrayArgs)
3133 return (tree->OperGet() == GT_ASG) && IsArgsFieldIndir(tree->gtGetOp1(), index, lvaNewObjArrayArgs) &&
3134 IsArgsAddr(tree->gtGetOp1()->gtGetOp1()->gtGetOp1(), lvaNewObjArrayArgs);
3137 static bool IsArgsFieldIndir(GenTree* tree, unsigned index, unsigned lvaNewObjArrayArgs)
3139 return (tree->OperGet() == GT_IND) && (tree->gtGetOp1()->OperGet() == GT_ADD) &&
3140 (tree->gtGetOp1()->gtGetOp2()->IsIntegralConst(sizeof(INT32) * index)) &&
3141 IsArgsAddr(tree->gtGetOp1()->gtGetOp1(), lvaNewObjArrayArgs);
3144 static bool IsArgsAddr(GenTree* tree, unsigned lvaNewObjArrayArgs)
3146 return (tree->OperGet() == GT_ADDR) && (tree->gtGetOp1()->OperGet() == GT_LCL_VAR) &&
3147 (tree->gtGetOp1()->AsLclVar()->GetLclNum() == lvaNewObjArrayArgs);
3150 static bool IsComma(GenTree* tree)
3152 return (tree != nullptr) && (tree->OperGet() == GT_COMMA);
3156 unsigned argIndex = 0;
3159 for (comma = argsArg->Current(); Match::IsComma(comma); comma = comma->gtGetOp2())
3161 if (lowerBoundsSpecified)
3164 // In general lower bounds can be ignored because they're not needed to
3165 // calculate the total number of elements. But for single dimensional arrays
3166 // we need to know if the lower bound is 0 because in this case the runtime
3167 // creates a SDArray and this affects the way the array data offset is calculated.
3172 GenTree* lowerBoundAssign = comma->gtGetOp1();
3173 assert(Match::IsArgsFieldInit(lowerBoundAssign, argIndex, lvaNewObjArrayArgs));
3174 GenTree* lowerBoundNode = lowerBoundAssign->gtGetOp2();
3176 if (lowerBoundNode->IsIntegralConst(0))
3182 comma = comma->gtGetOp2();
3186 GenTree* lengthNodeAssign = comma->gtGetOp1();
3187 assert(Match::IsArgsFieldInit(lengthNodeAssign, argIndex, lvaNewObjArrayArgs));
3188 GenTree* lengthNode = lengthNodeAssign->gtGetOp2();
3190 if (!lengthNode->IsCnsIntOrI())
3195 numElements *= S_SIZE_T(lengthNode->AsIntCon()->IconValue());
3199 assert((comma != nullptr) && Match::IsArgsAddr(comma, lvaNewObjArrayArgs));
3201 if (argIndex != numArgs)
3209 // Make sure there are exactly two arguments: the array class and
3210 // the number of elements.
3213 GenTree* arrayLengthNode;
3215 GenTreeArgList* args = newArrayCall->gtCall.gtCallArgs;
3216 #ifdef FEATURE_READYTORUN_COMPILER
3217 if (newArrayCall->gtCall.gtCallMethHnd == eeFindHelper(CORINFO_HELP_READYTORUN_NEWARR_1))
3219 // Array length is 1st argument for readytorun helper
3220 arrayLengthNode = args->Current();
3225 // Array length is 2nd argument for regular helper
3226 arrayLengthNode = args->Rest()->Current();
3230 // Make sure that the number of elements look valid.
3232 if (arrayLengthNode->gtOper != GT_CNS_INT)
3237 numElements = S_SIZE_T(arrayLengthNode->gtIntCon.gtIconVal);
3239 if (!info.compCompHnd->isSDArray(arrayClsHnd))
3245 CORINFO_CLASS_HANDLE elemClsHnd;
3246 var_types elementType = JITtype2varType(info.compCompHnd->getChildType(arrayClsHnd, &elemClsHnd));
3249 // Note that genTypeSize will return zero for non primitive types, which is exactly
3250 // what we want (size will then be 0, and we will catch this in the conditional below).
3251 // Note that we don't expect this to fail for valid binaries, so we assert in the
3252 // non-verification case (the verification case should not assert but rather correctly
3253 // handle bad binaries). This assert is not guarding any specific invariant, but rather
3254 // saying that we don't expect this to happen, and if it is hit, we need to investigate
3258 S_UINT32 elemSize(genTypeSize(elementType));
3259 S_UINT32 size = elemSize * S_UINT32(numElements);
3261 if (size.IsOverflow())
3266 if ((size.Value() == 0) || (varTypeIsGC(elementType)))
3268 assert(verNeedsVerification());
3272 void* initData = info.compCompHnd->getArrayInitializationData(fieldToken, size.Value());
3279 // At this point we are ready to commit to implementing the InitializeArray
3280 // intrinsic using a struct assignment. Pop the arguments from the stack and
3281 // return the struct assignment node.
3287 const unsigned blkSize = size.Value();
3288 unsigned dataOffset;
3292 dataOffset = eeGetMDArrayDataOffset(elementType, rank);
3296 dataOffset = eeGetArrayDataOffset(elementType);
3299 GenTree* dst = gtNewOperNode(GT_ADD, TYP_BYREF, arrayLocalNode, gtNewIconNode(dataOffset, TYP_I_IMPL));
3300 GenTree* blk = gtNewBlockVal(dst, blkSize);
3301 GenTree* src = gtNewIndOfIconHandleNode(TYP_STRUCT, (size_t)initData, GTF_ICON_STATIC_HDL, false);
3303 return gtNewBlkOpNode(blk, // dst
3310 //------------------------------------------------------------------------
3311 // impIntrinsic: possibly expand intrinsic call into alternate IR sequence
3314 // newobjThis - for constructor calls, the tree for the newly allocated object
3315 // clsHnd - handle for the intrinsic method's class
3316 // method - handle for the intrinsic method
3317 // sig - signature of the intrinsic method
3318 // methodFlags - CORINFO_FLG_XXX flags of the intrinsic method
3319 // memberRef - the token for the intrinsic method
3320 // readonlyCall - true if call has a readonly prefix
3321 // tailCall - true if call is in tail position
3322 // pConstrainedResolvedToken -- resolved token for constrained call, or nullptr
3323 // if call is not constrained
3324 // constraintCallThisTransform -- this transform to apply for a constrained call
3325 // pIntrinsicID [OUT] -- intrinsic ID (see enumeration in corinfo.h)
3326 // for "traditional" jit intrinsics
3327 // isSpecialIntrinsic [OUT] -- set true if intrinsic expansion is a call
3328 // that is amenable to special downstream optimization opportunities
3331 // IR tree to use in place of the call, or nullptr if the jit should treat
3332 // the intrinsic call like a normal call.
3334 // pIntrinsicID set to non-illegal value if the call is recognized as a
3335 // traditional jit intrinsic, even if the intrinsic is not expaned.
3337 // isSpecial set true if the expansion is subject to special
3338 // optimizations later in the jit processing
3341 // On success the IR tree may be a call to a different method or an inline
3342 // sequence. If it is a call, then the intrinsic processing here is responsible
3343 // for handling all the special cases, as upon return to impImportCall
3344 // expanded intrinsics bypass most of the normal call processing.
3346 // Intrinsics are generally not recognized in minopts and debug codegen.
3348 // However, certain traditional intrinsics are identifed as "must expand"
3349 // if there is no fallback implmentation to invoke; these must be handled
3350 // in all codegen modes.
3352 // New style intrinsics (where the fallback implementation is in IL) are
3353 // identified as "must expand" if they are invoked from within their
3354 // own method bodies.
3357 GenTree* Compiler::impIntrinsic(GenTree* newobjThis,
3358 CORINFO_CLASS_HANDLE clsHnd,
3359 CORINFO_METHOD_HANDLE method,
3360 CORINFO_SIG_INFO* sig,
3361 unsigned methodFlags,
3365 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
3366 CORINFO_THIS_TRANSFORM constraintCallThisTransform,
3367 CorInfoIntrinsics* pIntrinsicID,
3368 bool* isSpecialIntrinsic)
3370 assert((methodFlags & (CORINFO_FLG_INTRINSIC | CORINFO_FLG_JIT_INTRINSIC)) != 0);
3372 bool mustExpand = false;
3373 bool isSpecial = false;
3374 CorInfoIntrinsics intrinsicID = CORINFO_INTRINSIC_Illegal;
3375 NamedIntrinsic ni = NI_Illegal;
3377 if ((methodFlags & CORINFO_FLG_INTRINSIC) != 0)
3379 intrinsicID = info.compCompHnd->getIntrinsicID(method, &mustExpand);
3382 if ((methodFlags & CORINFO_FLG_JIT_INTRINSIC) != 0)
3384 // The recursive calls to Jit intrinsics are must-expand by convention.
3385 mustExpand = mustExpand || gtIsRecursiveCall(method);
3387 if (intrinsicID == CORINFO_INTRINSIC_Illegal)
3389 ni = lookupNamedIntrinsic(method);
3391 #ifdef FEATURE_HW_INTRINSICS
3392 if (ni > NI_HW_INTRINSIC_START && ni < NI_HW_INTRINSIC_END)
3394 return impHWIntrinsic(ni, method, sig, mustExpand);
3396 #endif // FEATURE_HW_INTRINSICS
3400 *pIntrinsicID = intrinsicID;
3402 #ifndef _TARGET_ARM_
3403 genTreeOps interlockedOperator;
3406 if (intrinsicID == CORINFO_INTRINSIC_StubHelpers_GetStubContext)
3408 // must be done regardless of DbgCode and MinOpts
3409 return gtNewLclvNode(lvaStubArgumentVar, TYP_I_IMPL);
3411 #ifdef _TARGET_64BIT_
3412 if (intrinsicID == CORINFO_INTRINSIC_StubHelpers_GetStubContextAddr)
3414 // must be done regardless of DbgCode and MinOpts
3415 return gtNewOperNode(GT_ADDR, TYP_I_IMPL, gtNewLclvNode(lvaStubArgumentVar, TYP_I_IMPL));
3418 assert(intrinsicID != CORINFO_INTRINSIC_StubHelpers_GetStubContextAddr);
3421 GenTree* retNode = nullptr;
3423 // Under debug and minopts, only expand what is required.
3424 if (!mustExpand && (opts.compDbgCode || opts.MinOpts()))
3426 *pIntrinsicID = CORINFO_INTRINSIC_Illegal;
3430 var_types callType = JITtype2varType(sig->retType);
3432 /* First do the intrinsics which are always smaller than a call */
3434 switch (intrinsicID)
3439 case CORINFO_INTRINSIC_Sin:
3440 case CORINFO_INTRINSIC_Cbrt:
3441 case CORINFO_INTRINSIC_Sqrt:
3442 case CORINFO_INTRINSIC_Abs:
3443 case CORINFO_INTRINSIC_Cos:
3444 case CORINFO_INTRINSIC_Round:
3445 case CORINFO_INTRINSIC_Cosh:
3446 case CORINFO_INTRINSIC_Sinh:
3447 case CORINFO_INTRINSIC_Tan:
3448 case CORINFO_INTRINSIC_Tanh:
3449 case CORINFO_INTRINSIC_Asin:
3450 case CORINFO_INTRINSIC_Asinh:
3451 case CORINFO_INTRINSIC_Acos:
3452 case CORINFO_INTRINSIC_Acosh:
3453 case CORINFO_INTRINSIC_Atan:
3454 case CORINFO_INTRINSIC_Atan2:
3455 case CORINFO_INTRINSIC_Atanh:
3456 case CORINFO_INTRINSIC_Log10:
3457 case CORINFO_INTRINSIC_Pow:
3458 case CORINFO_INTRINSIC_Exp:
3459 case CORINFO_INTRINSIC_Ceiling:
3460 case CORINFO_INTRINSIC_Floor:
3461 retNode = impMathIntrinsic(method, sig, callType, intrinsicID, tailCall);
3464 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
3465 // TODO-ARM-CQ: reenable treating Interlocked operation as intrinsic
3466 case CORINFO_INTRINSIC_InterlockedAdd32:
3467 interlockedOperator = GT_LOCKADD;
3468 goto InterlockedBinOpCommon;
3469 case CORINFO_INTRINSIC_InterlockedXAdd32:
3470 interlockedOperator = GT_XADD;
3471 goto InterlockedBinOpCommon;
3472 case CORINFO_INTRINSIC_InterlockedXchg32:
3473 interlockedOperator = GT_XCHG;
3474 goto InterlockedBinOpCommon;
3476 #ifdef _TARGET_64BIT_
3477 case CORINFO_INTRINSIC_InterlockedAdd64:
3478 interlockedOperator = GT_LOCKADD;
3479 goto InterlockedBinOpCommon;
3480 case CORINFO_INTRINSIC_InterlockedXAdd64:
3481 interlockedOperator = GT_XADD;
3482 goto InterlockedBinOpCommon;
3483 case CORINFO_INTRINSIC_InterlockedXchg64:
3484 interlockedOperator = GT_XCHG;
3485 goto InterlockedBinOpCommon;
3486 #endif // _TARGET_AMD64_
3488 InterlockedBinOpCommon:
3489 assert(callType != TYP_STRUCT);
3490 assert(sig->numArgs == 2);
3492 op2 = impPopStack().val;
3493 op1 = impPopStack().val;
3499 // field (for example)
3501 // In the case where the first argument is the address of a local, we might
3502 // want to make this *not* make the var address-taken -- but atomic instructions
3503 // on a local are probably pretty useless anyway, so we probably don't care.
3505 op1 = gtNewOperNode(interlockedOperator, genActualType(callType), op1, op2);
3506 op1->gtFlags |= GTF_GLOB_REF | GTF_ASG;
3509 #endif // defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
3511 case CORINFO_INTRINSIC_MemoryBarrier:
3513 assert(sig->numArgs == 0);
3515 op1 = new (this, GT_MEMORYBARRIER) GenTree(GT_MEMORYBARRIER, TYP_VOID);
3516 op1->gtFlags |= GTF_GLOB_REF | GTF_ASG;
3520 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
3521 // TODO-ARM-CQ: reenable treating InterlockedCmpXchg32 operation as intrinsic
3522 case CORINFO_INTRINSIC_InterlockedCmpXchg32:
3523 #ifdef _TARGET_64BIT_
3524 case CORINFO_INTRINSIC_InterlockedCmpXchg64:
3527 assert(callType != TYP_STRUCT);
3528 assert(sig->numArgs == 3);
3531 op3 = impPopStack().val; // comparand
3532 op2 = impPopStack().val; // value
3533 op1 = impPopStack().val; // location
3535 GenTree* node = new (this, GT_CMPXCHG) GenTreeCmpXchg(genActualType(callType), op1, op2, op3);
3537 node->gtCmpXchg.gtOpLocation->gtFlags |= GTF_DONT_CSE;
3541 #endif // defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
3543 case CORINFO_INTRINSIC_StringLength:
3544 op1 = impPopStack().val;
3545 if (!opts.MinOpts() && !opts.compDbgCode)
3547 GenTreeArrLen* arrLen = gtNewArrLen(TYP_INT, op1, offsetof(CORINFO_String, stringLen));
3552 /* Create the expression "*(str_addr + stringLengthOffset)" */
3553 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1,
3554 gtNewIconNode(offsetof(CORINFO_String, stringLen), TYP_I_IMPL));
3555 op1 = gtNewOperNode(GT_IND, TYP_INT, op1);
3558 // Getting the length of a null string should throw
3559 op1->gtFlags |= GTF_EXCEPT;
3564 case CORINFO_INTRINSIC_StringGetChar:
3565 op2 = impPopStack().val;
3566 op1 = impPopStack().val;
3567 op1 = gtNewIndexRef(TYP_USHORT, op1, op2);
3568 op1->gtFlags |= GTF_INX_STRING_LAYOUT;
3572 case CORINFO_INTRINSIC_InitializeArray:
3573 retNode = impInitializeArrayIntrinsic(sig);
3576 case CORINFO_INTRINSIC_Array_Address:
3577 case CORINFO_INTRINSIC_Array_Get:
3578 case CORINFO_INTRINSIC_Array_Set:
3579 retNode = impArrayAccessIntrinsic(clsHnd, sig, memberRef, readonlyCall, intrinsicID);
3582 case CORINFO_INTRINSIC_GetTypeFromHandle:
3583 op1 = impStackTop(0).val;
3584 if (op1->gtOper == GT_CALL && (op1->gtCall.gtCallType == CT_HELPER) &&
3585 gtIsTypeHandleToRuntimeTypeHelper(op1->AsCall()))
3587 op1 = impPopStack().val;
3588 // Change call to return RuntimeType directly.
3589 op1->gtType = TYP_REF;
3592 // Call the regular function.
3595 case CORINFO_INTRINSIC_RTH_GetValueInternal:
3596 op1 = impStackTop(0).val;
3597 if (op1->gtOper == GT_CALL && (op1->gtCall.gtCallType == CT_HELPER) &&
3598 gtIsTypeHandleToRuntimeTypeHelper(op1->AsCall()))
3601 // Helper-RuntimeTypeHandle -> TreeToGetNativeTypeHandle
3604 // TreeToGetNativeTypeHandle
3606 // Remove call to helper and return the native TypeHandle pointer that was the parameter
3609 op1 = impPopStack().val;
3611 // Get native TypeHandle argument to old helper
3612 op1 = op1->gtCall.gtCallArgs;
3613 assert(op1->OperIsList());
3614 assert(op1->gtOp.gtOp2 == nullptr);
3615 op1 = op1->gtOp.gtOp1;
3618 // Call the regular function.
3621 #ifndef LEGACY_BACKEND
3622 case CORINFO_INTRINSIC_Object_GetType:
3624 JITDUMP("\n impIntrinsic: call to Object.GetType\n");
3625 op1 = impStackTop(0).val;
3627 // If we're calling GetType on a boxed value, just get the type directly.
3628 if (op1->IsBoxedValue())
3630 JITDUMP("Attempting to optimize box(...).getType() to direct type construction\n");
3632 // Try and clean up the box. Obtain the handle we
3633 // were going to pass to the newobj.
3634 GenTree* boxTypeHandle = gtTryRemoveBoxUpstreamEffects(op1, BR_REMOVE_AND_NARROW_WANT_TYPE_HANDLE);
3636 if (boxTypeHandle != nullptr)
3638 // Note we don't need to play the TYP_STRUCT games here like
3639 // do for LDTOKEN since the return value of this operator is Type,
3640 // not RuntimeTypeHandle.
3642 GenTreeArgList* helperArgs = gtNewArgList(boxTypeHandle);
3643 GenTree* runtimeType =
3644 gtNewHelperCallNode(CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE, TYP_REF, helperArgs);
3645 retNode = runtimeType;
3649 // If we have a constrained callvirt with a "box this" transform
3650 // we know we have a value class and hence an exact type.
3652 // If so, instead of boxing and then extracting the type, just
3653 // construct the type directly.
3654 if ((retNode == nullptr) && (pConstrainedResolvedToken != nullptr) &&
3655 (constraintCallThisTransform == CORINFO_BOX_THIS))
3657 // Ensure this is one of the is simple box cases (in particular, rule out nullables).
3658 const CorInfoHelpFunc boxHelper = info.compCompHnd->getBoxHelper(pConstrainedResolvedToken->hClass);
3659 const bool isSafeToOptimize = (boxHelper == CORINFO_HELP_BOX);
3661 if (isSafeToOptimize)
3663 JITDUMP("Optimizing constrained box-this obj.getType() to direct type construction\n");
3665 GenTree* typeHandleOp =
3666 impTokenToHandle(pConstrainedResolvedToken, nullptr, TRUE /* mustRestoreHandle */);
3667 if (typeHandleOp == nullptr)
3669 assert(compDonotInline());
3672 GenTreeArgList* helperArgs = gtNewArgList(typeHandleOp);
3673 GenTree* runtimeType =
3674 gtNewHelperCallNode(CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE, TYP_REF, helperArgs);
3675 retNode = runtimeType;
3680 if (retNode != nullptr)
3682 JITDUMP("Optimized result for call to GetType is\n");
3685 gtDispTree(retNode);
3690 // Else expand as an intrinsic, unless the call is constrained,
3691 // in which case we defer expansion to allow impImportCall do the
3692 // special constraint processing.
3693 if ((retNode == nullptr) && (pConstrainedResolvedToken == nullptr))
3695 JITDUMP("Expanding as special intrinsic\n");
3697 op1 = new (this, GT_INTRINSIC) GenTreeIntrinsic(genActualType(callType), op1, intrinsicID, method);
3699 // Set the CALL flag to indicate that the operator is implemented by a call.
3700 // Set also the EXCEPTION flag because the native implementation of
3701 // CORINFO_INTRINSIC_Object_GetType intrinsic can throw NullReferenceException.
3702 op1->gtFlags |= (GTF_CALL | GTF_EXCEPT);
3704 // Might be further optimizable, so arrange to leave a mark behind
3708 if (retNode == nullptr)
3710 JITDUMP("Leaving as normal call\n");
3711 // Might be further optimizable, so arrange to leave a mark behind
3719 // Implement ByReference Ctor. This wraps the assignment of the ref into a byref-like field
3720 // in a value type. The canonical example of this is Span<T>. In effect this is just a
3721 // substitution. The parameter byref will be assigned into the newly allocated object.
3722 case CORINFO_INTRINSIC_ByReference_Ctor:
3724 // Remove call to constructor and directly assign the byref passed
3725 // to the call to the first slot of the ByReference struct.
3726 op1 = impPopStack().val;
3727 GenTree* thisptr = newobjThis;
3728 CORINFO_FIELD_HANDLE fldHnd = info.compCompHnd->getFieldInClass(clsHnd, 0);
3729 GenTree* field = gtNewFieldRef(TYP_BYREF, fldHnd, thisptr, 0, false);
3730 GenTree* assign = gtNewAssignNode(field, op1);
3731 GenTree* byReferenceStruct = gtCloneExpr(thisptr->gtGetOp1());
3732 assert(byReferenceStruct != nullptr);
3733 impPushOnStack(byReferenceStruct, typeInfo(TI_STRUCT, clsHnd));
3737 // Implement ptr value getter for ByReference struct.
3738 case CORINFO_INTRINSIC_ByReference_Value:
3740 op1 = impPopStack().val;
3741 CORINFO_FIELD_HANDLE fldHnd = info.compCompHnd->getFieldInClass(clsHnd, 0);
3742 GenTree* field = gtNewFieldRef(TYP_BYREF, fldHnd, op1, 0, false);
3746 case CORINFO_INTRINSIC_Span_GetItem:
3747 case CORINFO_INTRINSIC_ReadOnlySpan_GetItem:
3749 // Have index, stack pointer-to Span<T> s on the stack. Expand to:
3753 // BoundsCheck(index, s->_length)
3754 // s->_pointer + index * sizeof(T)
3756 // For ReadOnlySpan<T> -- same expansion, as it now returns a readonly ref
3758 // Signature should show one class type parameter, which
3759 // we need to examine.
3760 assert(sig->sigInst.classInstCount == 1);
3761 CORINFO_CLASS_HANDLE spanElemHnd = sig->sigInst.classInst[0];
3762 const unsigned elemSize = info.compCompHnd->getClassSize(spanElemHnd);
3763 assert(elemSize > 0);
3765 const bool isReadOnly = (intrinsicID == CORINFO_INTRINSIC_ReadOnlySpan_GetItem);
3767 JITDUMP("\nimpIntrinsic: Expanding %sSpan<T>.get_Item, T=%s, sizeof(T)=%u\n", isReadOnly ? "ReadOnly" : "",
3768 info.compCompHnd->getClassName(spanElemHnd), elemSize);
3770 GenTree* index = impPopStack().val;
3771 GenTree* ptrToSpan = impPopStack().val;
3772 GenTree* indexClone = nullptr;
3773 GenTree* ptrToSpanClone = nullptr;
3778 printf("with ptr-to-span\n");
3779 gtDispTree(ptrToSpan);
3780 printf("and index\n");
3783 #endif // defined(DEBUG)
3785 // We need to use both index and ptr-to-span twice, so clone or spill.
3786 index = impCloneExpr(index, &indexClone, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
3787 nullptr DEBUGARG("Span.get_Item index"));
3788 ptrToSpan = impCloneExpr(ptrToSpan, &ptrToSpanClone, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
3789 nullptr DEBUGARG("Span.get_Item ptrToSpan"));
3792 CORINFO_FIELD_HANDLE lengthHnd = info.compCompHnd->getFieldInClass(clsHnd, 1);
3793 const unsigned lengthOffset = info.compCompHnd->getFieldOffset(lengthHnd);
3794 GenTree* length = gtNewFieldRef(TYP_INT, lengthHnd, ptrToSpan, lengthOffset, false);
3795 GenTree* boundsCheck = new (this, GT_ARR_BOUNDS_CHECK)
3796 GenTreeBoundsChk(GT_ARR_BOUNDS_CHECK, TYP_VOID, index, length, SCK_RNGCHK_FAIL);
3799 GenTree* indexIntPtr = impImplicitIorI4Cast(indexClone, TYP_I_IMPL);
3800 GenTree* sizeofNode = gtNewIconNode(elemSize);
3801 GenTree* mulNode = gtNewOperNode(GT_MUL, TYP_I_IMPL, indexIntPtr, sizeofNode);
3802 CORINFO_FIELD_HANDLE ptrHnd = info.compCompHnd->getFieldInClass(clsHnd, 0);
3803 const unsigned ptrOffset = info.compCompHnd->getFieldOffset(ptrHnd);
3804 GenTree* data = gtNewFieldRef(TYP_BYREF, ptrHnd, ptrToSpanClone, ptrOffset, false);
3805 GenTree* result = gtNewOperNode(GT_ADD, TYP_BYREF, data, mulNode);
3808 var_types resultType = JITtype2varType(sig->retType);
3809 assert(resultType == result->TypeGet());
3810 retNode = gtNewOperNode(GT_COMMA, resultType, boundsCheck, result);
3815 case CORINFO_INTRINSIC_GetRawHandle:
3817 noway_assert(IsTargetAbi(CORINFO_CORERT_ABI)); // Only CoreRT supports it.
3818 CORINFO_RESOLVED_TOKEN resolvedToken;
3819 resolvedToken.tokenContext = MAKE_METHODCONTEXT(info.compMethodHnd);
3820 resolvedToken.tokenScope = info.compScopeHnd;
3821 resolvedToken.token = memberRef;
3822 resolvedToken.tokenType = CORINFO_TOKENKIND_Method;
3824 CORINFO_GENERICHANDLE_RESULT embedInfo;
3825 info.compCompHnd->expandRawHandleIntrinsic(&resolvedToken, &embedInfo);
3827 GenTree* rawHandle = impLookupToTree(&resolvedToken, &embedInfo.lookup, gtTokenToIconFlags(memberRef),
3828 embedInfo.compileTimeHandle);
3829 if (rawHandle == nullptr)
3834 noway_assert(genTypeSize(rawHandle->TypeGet()) == genTypeSize(TYP_I_IMPL));
3836 unsigned rawHandleSlot = lvaGrabTemp(true DEBUGARG("rawHandle"));
3837 impAssignTempGen(rawHandleSlot, rawHandle, clsHnd, (unsigned)CHECK_SPILL_NONE);
3839 GenTree* lclVar = gtNewLclvNode(rawHandleSlot, TYP_I_IMPL);
3840 GenTree* lclVarAddr = gtNewOperNode(GT_ADDR, TYP_I_IMPL, lclVar);
3841 var_types resultType = JITtype2varType(sig->retType);
3842 retNode = gtNewOperNode(GT_IND, resultType, lclVarAddr);
3847 case CORINFO_INTRINSIC_TypeEQ:
3848 case CORINFO_INTRINSIC_TypeNEQ:
3850 JITDUMP("Importing Type.op_*Equality intrinsic\n");
3851 op1 = impStackTop(1).val;
3852 op2 = impStackTop(0).val;
3853 GenTree* optTree = gtFoldTypeEqualityCall(intrinsicID, op1, op2);
3854 if (optTree != nullptr)
3856 // Success, clean up the evaluation stack.
3860 // See if we can optimize even further, to a handle compare.
3861 optTree = gtFoldTypeCompare(optTree);
3863 // See if we can now fold a handle compare to a constant.
3864 optTree = gtFoldExpr(optTree);
3870 // Retry optimizing these later
3876 case CORINFO_INTRINSIC_GetCurrentManagedThread:
3877 case CORINFO_INTRINSIC_GetManagedThreadId:
3879 // Retry optimizing these during morph
3885 /* Unknown intrinsic */
3886 intrinsicID = CORINFO_INTRINSIC_Illegal;
3890 // Look for new-style jit intrinsics by name
3891 if (ni != NI_Illegal)
3893 assert(retNode == nullptr);
3896 case NI_System_Enum_HasFlag:
3898 GenTree* thisOp = impStackTop(1).val;
3899 GenTree* flagOp = impStackTop(0).val;
3900 GenTree* optTree = gtOptimizeEnumHasFlag(thisOp, flagOp);
3902 if (optTree != nullptr)
3904 // Optimization successful. Pop the stack for real.
3911 // Retry optimizing this during morph.
3918 case NI_MathF_Round:
3921 // Math.Round and MathF.Round used to be a traditional JIT intrinsic. In order
3922 // to simplify the transition, we will just treat it as if it was still the
3923 // old intrinsic, CORINFO_INTRINSIC_Round. This should end up flowing properly
3926 retNode = impMathIntrinsic(method, sig, callType, CORINFO_INTRINSIC_Round, tailCall);
3930 case NI_System_Collections_Generic_EqualityComparer_get_Default:
3932 // Flag for later handling during devirtualization.
3944 if (retNode == nullptr)
3946 NO_WAY("JIT must expand the intrinsic!");
3950 // Optionally report if this intrinsic is special
3951 // (that is, potentially re-optimizable during morph).
3952 if (isSpecialIntrinsic != nullptr)
3954 *isSpecialIntrinsic = isSpecial;
3960 GenTree* Compiler::impMathIntrinsic(CORINFO_METHOD_HANDLE method,
3961 CORINFO_SIG_INFO* sig,
3963 CorInfoIntrinsics intrinsicID,
3969 assert(callType != TYP_STRUCT);
3970 assert((intrinsicID == CORINFO_INTRINSIC_Sin) || intrinsicID == CORINFO_INTRINSIC_Cbrt ||
3971 (intrinsicID == CORINFO_INTRINSIC_Sqrt) || (intrinsicID == CORINFO_INTRINSIC_Abs) ||
3972 (intrinsicID == CORINFO_INTRINSIC_Cos) || (intrinsicID == CORINFO_INTRINSIC_Round) ||
3973 (intrinsicID == CORINFO_INTRINSIC_Cosh) || (intrinsicID == CORINFO_INTRINSIC_Sinh) ||
3974 (intrinsicID == CORINFO_INTRINSIC_Tan) || (intrinsicID == CORINFO_INTRINSIC_Tanh) ||
3975 (intrinsicID == CORINFO_INTRINSIC_Asin) || (intrinsicID == CORINFO_INTRINSIC_Asinh) ||
3976 (intrinsicID == CORINFO_INTRINSIC_Acos) || (intrinsicID == CORINFO_INTRINSIC_Acosh) ||
3977 (intrinsicID == CORINFO_INTRINSIC_Atan) || (intrinsicID == CORINFO_INTRINSIC_Atan2) ||
3978 (intrinsicID == CORINFO_INTRINSIC_Atanh) || (intrinsicID == CORINFO_INTRINSIC_Log10) ||
3979 (intrinsicID == CORINFO_INTRINSIC_Pow) || (intrinsicID == CORINFO_INTRINSIC_Exp) ||
3980 (intrinsicID == CORINFO_INTRINSIC_Ceiling) || (intrinsicID == CORINFO_INTRINSIC_Floor));
3984 #if defined(LEGACY_BACKEND)
3985 if (IsTargetIntrinsic(intrinsicID))
3986 #elif !defined(_TARGET_X86_)
3987 // Intrinsics that are not implemented directly by target instructions will
3988 // be re-materialized as users calls in rationalizer. For prefixed tail calls,
3989 // don't do this optimization, because
3990 // a) For back compatibility reasons on desktop.Net 4.6 / 4.6.1
3991 // b) It will be non-trivial task or too late to re-materialize a surviving
3992 // tail prefixed GT_INTRINSIC as tail call in rationalizer.
3993 if (!IsIntrinsicImplementedByUserCall(intrinsicID) || !tailCall)
3995 // On x86 RyuJIT, importing intrinsics that are implemented as user calls can cause incorrect calculation
3996 // of the depth of the stack if these intrinsics are used as arguments to another call. This causes bad
3997 // code generation for certain EH constructs.
3998 if (!IsIntrinsicImplementedByUserCall(intrinsicID))
4001 switch (sig->numArgs)
4004 op1 = impPopStack().val;
4006 #if FEATURE_X87_DOUBLES
4008 // X87 stack doesn't differentiate between float/double
4009 // so it doesn't need a cast, but everybody else does
4010 // Just double check it is at least a FP type
4011 noway_assert(varTypeIsFloating(op1));
4013 #else // FEATURE_X87_DOUBLES
4014 assert(varTypeIsFloating(op1));
4016 if (op1->TypeGet() != callType)
4018 op1 = gtNewCastNode(callType, op1, false, callType);
4021 #endif // FEATURE_X87_DOUBLES
4023 op1 = new (this, GT_INTRINSIC) GenTreeIntrinsic(genActualType(callType), op1, intrinsicID, method);
4027 op2 = impPopStack().val;
4028 op1 = impPopStack().val;
4030 #if FEATURE_X87_DOUBLES
4032 // X87 stack doesn't differentiate between float/double
4033 // so it doesn't need a cast, but everybody else does
4034 // Just double check it is at least a FP type
4035 noway_assert(varTypeIsFloating(op2));
4036 noway_assert(varTypeIsFloating(op1));
4038 #else // FEATURE_X87_DOUBLES
4039 assert(varTypeIsFloating(op1));
4040 assert(varTypeIsFloating(op2));
4042 if (op2->TypeGet() != callType)
4044 op2 = gtNewCastNode(callType, op2, false, callType);
4046 if (op1->TypeGet() != callType)
4048 op1 = gtNewCastNode(callType, op1, false, callType);
4051 #endif // FEATURE_X87_DOUBLES
4053 op1 = new (this, GT_INTRINSIC) GenTreeIntrinsic(genActualType(callType), op1, op2, intrinsicID, method);
4057 NO_WAY("Unsupported number of args for Math Instrinsic");
4060 #ifndef LEGACY_BACKEND
4061 if (IsIntrinsicImplementedByUserCall(intrinsicID))
4063 op1->gtFlags |= GTF_CALL;
4071 //------------------------------------------------------------------------
4072 // lookupNamedIntrinsic: map method to jit named intrinsic value
4075 // method -- method handle for method
4078 // Id for the named intrinsic, or Illegal if none.
4081 // method should have CORINFO_FLG_JIT_INTRINSIC set in its attributes,
4082 // otherwise it is not a named jit intrinsic.
4085 NamedIntrinsic Compiler::lookupNamedIntrinsic(CORINFO_METHOD_HANDLE method)
4087 NamedIntrinsic result = NI_Illegal;
4089 const char* className = nullptr;
4090 const char* namespaceName = nullptr;
4091 const char* methodName = info.compCompHnd->getMethodNameFromMetadata(method, &className, &namespaceName);
4093 if ((namespaceName == nullptr) || (className == nullptr) || (methodName == nullptr))
4098 if (strcmp(namespaceName, "System") == 0)
4100 if ((strcmp(className, "Enum") == 0) && (strcmp(methodName, "HasFlag") == 0))
4102 result = NI_System_Enum_HasFlag;
4104 else if ((strcmp(className, "MathF") == 0) && (strcmp(methodName, "Round") == 0))
4106 result = NI_MathF_Round;
4108 else if ((strcmp(className, "Math") == 0) && (strcmp(methodName, "Round") == 0))
4110 result = NI_Math_Round;
4113 else if (strcmp(namespaceName, "System.Collections.Generic") == 0)
4115 if ((strcmp(className, "EqualityComparer`1") == 0) && (strcmp(methodName, "get_Default") == 0))
4117 result = NI_System_Collections_Generic_EqualityComparer_get_Default;
4121 #ifdef FEATURE_HW_INTRINSICS
4122 #if defined(_TARGET_XARCH_)
4123 if ((namespaceName != nullptr) && strcmp(namespaceName, "System.Runtime.Intrinsics.X86") == 0)
4125 InstructionSet isa = lookupHWIntrinsicISA(className);
4126 result = lookupHWIntrinsic(methodName, isa);
4128 #elif defined(_TARGET_ARM64_)
4129 if ((namespaceName != nullptr) && strcmp(namespaceName, "System.Runtime.Intrinsics.Arm.Arm64") == 0)
4131 result = lookupHWIntrinsic(className, methodName);
4133 #else // !defined(_TARGET_XARCH_) && !defined(_TARGET_ARM64_)
4134 #error Unsupported platform
4135 #endif // !defined(_TARGET_XARCH_) && !defined(_TARGET_ARM64_)
4136 #endif // FEATURE_HW_INTRINSICS
4140 /*****************************************************************************/
4142 GenTree* Compiler::impArrayAccessIntrinsic(
4143 CORINFO_CLASS_HANDLE clsHnd, CORINFO_SIG_INFO* sig, int memberRef, bool readonlyCall, CorInfoIntrinsics intrinsicID)
4145 /* If we are generating SMALL_CODE, we don't want to use intrinsics for
4146 the following, as it generates fatter code.
4149 if (compCodeOpt() == SMALL_CODE)
4154 /* These intrinsics generate fatter (but faster) code and are only
4155 done if we don't need SMALL_CODE */
4157 unsigned rank = (intrinsicID == CORINFO_INTRINSIC_Array_Set) ? (sig->numArgs - 1) : sig->numArgs;
4159 // The rank 1 case is special because it has to handle two array formats
4160 // we will simply not do that case
4161 if (rank > GT_ARR_MAX_RANK || rank <= 1)
4166 CORINFO_CLASS_HANDLE arrElemClsHnd = nullptr;
4167 var_types elemType = JITtype2varType(info.compCompHnd->getChildType(clsHnd, &arrElemClsHnd));
4169 // For the ref case, we will only be able to inline if the types match
4170 // (verifier checks for this, we don't care for the nonverified case and the
4171 // type is final (so we don't need to do the cast)
4172 if ((intrinsicID != CORINFO_INTRINSIC_Array_Get) && !readonlyCall && varTypeIsGC(elemType))
4174 // Get the call site signature
4175 CORINFO_SIG_INFO LocalSig;
4176 eeGetCallSiteSig(memberRef, info.compScopeHnd, impTokenLookupContextHandle, &LocalSig);
4177 assert(LocalSig.hasThis());
4179 CORINFO_CLASS_HANDLE actualElemClsHnd;
4181 if (intrinsicID == CORINFO_INTRINSIC_Array_Set)
4183 // Fetch the last argument, the one that indicates the type we are setting.
4184 CORINFO_ARG_LIST_HANDLE argType = LocalSig.args;
4185 for (unsigned r = 0; r < rank; r++)
4187 argType = info.compCompHnd->getArgNext(argType);
4190 typeInfo argInfo = verParseArgSigToTypeInfo(&LocalSig, argType);
4191 actualElemClsHnd = argInfo.GetClassHandle();
4195 assert(intrinsicID == CORINFO_INTRINSIC_Array_Address);
4197 // Fetch the return type
4198 typeInfo retInfo = verMakeTypeInfo(LocalSig.retType, LocalSig.retTypeClass);
4199 assert(retInfo.IsByRef());
4200 actualElemClsHnd = retInfo.GetClassHandle();
4203 // if it's not final, we can't do the optimization
4204 if (!(info.compCompHnd->getClassAttribs(actualElemClsHnd) & CORINFO_FLG_FINAL))
4210 unsigned arrayElemSize;
4211 if (elemType == TYP_STRUCT)
4213 assert(arrElemClsHnd);
4215 arrayElemSize = info.compCompHnd->getClassSize(arrElemClsHnd);
4219 arrayElemSize = genTypeSize(elemType);
4222 if ((unsigned char)arrayElemSize != arrayElemSize)
4224 // arrayElemSize would be truncated as an unsigned char.
4225 // This means the array element is too large. Don't do the optimization.
4229 GenTree* val = nullptr;
4231 if (intrinsicID == CORINFO_INTRINSIC_Array_Set)
4233 // Assignment of a struct is more work, and there are more gets than sets.
4234 if (elemType == TYP_STRUCT)
4239 val = impPopStack().val;
4240 assert(genActualType(elemType) == genActualType(val->gtType) ||
4241 (elemType == TYP_FLOAT && val->gtType == TYP_DOUBLE) ||
4242 (elemType == TYP_INT && val->gtType == TYP_BYREF) ||
4243 (elemType == TYP_DOUBLE && val->gtType == TYP_FLOAT));
4246 noway_assert((unsigned char)GT_ARR_MAX_RANK == GT_ARR_MAX_RANK);
4248 GenTree* inds[GT_ARR_MAX_RANK];
4249 for (unsigned k = rank; k > 0; k--)
4251 inds[k - 1] = impPopStack().val;
4254 GenTree* arr = impPopStack().val;
4255 assert(arr->gtType == TYP_REF);
4258 new (this, GT_ARR_ELEM) GenTreeArrElem(TYP_BYREF, arr, static_cast<unsigned char>(rank),
4259 static_cast<unsigned char>(arrayElemSize), elemType, &inds[0]);
4261 if (intrinsicID != CORINFO_INTRINSIC_Array_Address)
4263 arrElem = gtNewOperNode(GT_IND, elemType, arrElem);
4266 if (intrinsicID == CORINFO_INTRINSIC_Array_Set)
4268 assert(val != nullptr);
4269 return gtNewAssignNode(arrElem, val);
4277 BOOL Compiler::verMergeEntryStates(BasicBlock* block, bool* changed)
4281 // do some basic checks first
4282 if (block->bbStackDepthOnEntry() != verCurrentState.esStackDepth)
4287 if (verCurrentState.esStackDepth > 0)
4289 // merge stack types
4290 StackEntry* parentStack = block->bbStackOnEntry();
4291 StackEntry* childStack = verCurrentState.esStack;
4293 for (i = 0; i < verCurrentState.esStackDepth; i++, parentStack++, childStack++)
4295 if (tiMergeToCommonParent(&parentStack->seTypeInfo, &childStack->seTypeInfo, changed) == FALSE)
4302 // merge initialization status of this ptr
4304 if (verTrackObjCtorInitState)
4306 // If we're tracking the CtorInitState, then it must not be unknown in the current state.
4307 assert(verCurrentState.thisInitialized != TIS_Bottom);
4309 // If the successor block's thisInit state is unknown, copy it from the current state.
4310 if (block->bbThisOnEntry() == TIS_Bottom)
4313 verSetThisInit(block, verCurrentState.thisInitialized);
4315 else if (verCurrentState.thisInitialized != block->bbThisOnEntry())
4317 if (block->bbThisOnEntry() != TIS_Top)
4320 verSetThisInit(block, TIS_Top);
4322 if (block->bbFlags & BBF_FAILED_VERIFICATION)
4324 // The block is bad. Control can flow through the block to any handler that catches the
4325 // verification exception, but the importer ignores bad blocks and therefore won't model
4326 // this flow in the normal way. To complete the merge into the bad block, the new state
4327 // needs to be manually pushed to the handlers that may be reached after the verification
4328 // exception occurs.
4330 // Usually, the new state was already propagated to the relevant handlers while processing
4331 // the predecessors of the bad block. The exception is when the bad block is at the start
4332 // of a try region, meaning it is protected by additional handlers that do not protect its
4335 if (block->hasTryIndex() && ((block->bbFlags & BBF_TRY_BEG) != 0))
4337 // Push TIS_Top to the handlers that protect the bad block. Note that this can cause
4338 // recursive calls back into this code path (if successors of the current bad block are
4339 // also bad blocks).
4341 ThisInitState origTIS = verCurrentState.thisInitialized;
4342 verCurrentState.thisInitialized = TIS_Top;
4343 impVerifyEHBlock(block, true);
4344 verCurrentState.thisInitialized = origTIS;
4352 assert(verCurrentState.thisInitialized == TIS_Bottom && block->bbThisOnEntry() == TIS_Bottom);
4358 /*****************************************************************************
4359 * 'logMsg' is true if a log message needs to be logged. false if the caller has
4360 * already logged it (presumably in a more detailed fashion than done here)
4361 * 'bVerificationException' is true for a verification exception, false for a
4362 * "call unauthorized by host" exception.
4365 void Compiler::verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg))
4367 block->bbJumpKind = BBJ_THROW;
4368 block->bbFlags |= BBF_FAILED_VERIFICATION;
4370 impCurStmtOffsSet(block->bbCodeOffs);
4373 // we need this since BeginTreeList asserts otherwise
4374 impTreeList = impTreeLast = nullptr;
4375 block->bbFlags &= ~BBF_IMPORTED;
4379 JITLOG((LL_ERROR, "Verification failure: while compiling %s near IL offset %x..%xh \n", info.compFullName,
4380 block->bbCodeOffs, block->bbCodeOffsEnd));
4383 printf("\n\nVerification failure: %s near IL %xh \n", info.compFullName, block->bbCodeOffs);
4387 if (JitConfig.DebugBreakOnVerificationFailure())
4395 // if the stack is non-empty evaluate all the side-effects
4396 if (verCurrentState.esStackDepth > 0)
4398 impEvalSideEffects();
4400 assert(verCurrentState.esStackDepth == 0);
4403 gtNewHelperCallNode(CORINFO_HELP_VERIFICATION, TYP_VOID, gtNewArgList(gtNewIconNode(block->bbCodeOffs)));
4404 // verCurrentState.esStackDepth = 0;
4405 impAppendTree(op1, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
4407 // The inliner is not able to handle methods that require throw block, so
4408 // make sure this methods never gets inlined.
4409 info.compCompHnd->setMethodAttribs(info.compMethodHnd, CORINFO_FLG_BAD_INLINEE);
4412 /*****************************************************************************
4415 void Compiler::verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg))
4418 // In AMD64, for historical reasons involving design limitations of JIT64, the VM has a
4419 // slightly different mechanism in which it calls the JIT to perform IL verification:
4420 // in the case of transparent methods the VM calls for a predicate IsVerifiable()
4421 // that consists of calling the JIT with the IMPORT_ONLY flag and with the IL verify flag on.
4422 // If the JIT determines the method is not verifiable, it should raise the exception to the VM and let
4423 // it bubble up until reported by the runtime. Currently in RyuJIT, this method doesn't bubble
4424 // up the exception, instead it embeds a throw inside the offending basic block and lets this
4425 // to fail upon runtime of the jitted method.
4427 // For AMD64 we don't want this behavior when the JIT has been called only for verification (i.e.
4428 // with the IMPORT_ONLY and IL Verification flag set) because this won't actually generate code,
4429 // just try to find out whether to fail this method before even actually jitting it. So, in case
4430 // we detect these two conditions, instead of generating a throw statement inside the offending
4431 // basic block, we immediately fail to JIT and notify the VM to make the IsVerifiable() predicate
4432 // to return false and make RyuJIT behave the same way JIT64 does.
4434 // The rationale behind this workaround is to avoid modifying the VM and maintain compatibility between JIT64 and
4435 // RyuJIT for the time being until we completely replace JIT64.
4436 // TODO-ARM64-Cleanup: We probably want to actually modify the VM in the future to avoid the unnecesary two passes.
4438 // In AMD64 we must make sure we're behaving the same way as JIT64, meaning we should only raise the verification
4439 // exception if we are only importing and verifying. The method verNeedsVerification() can also modify the
4440 // tiVerificationNeeded flag in the case it determines it can 'skip verification' during importation and defer it
4441 // to a runtime check. That's why we must assert one or the other (since the flag tiVerificationNeeded can
4442 // be turned off during importation).
4443 CLANG_FORMAT_COMMENT_ANCHOR;
4445 #ifdef _TARGET_64BIT_
4448 bool canSkipVerificationResult =
4449 info.compCompHnd->canSkipMethodVerification(info.compMethodHnd) != CORINFO_VERIFICATION_CANNOT_SKIP;
4450 assert(tiVerificationNeeded || canSkipVerificationResult);
4453 // Add the non verifiable flag to the compiler
4454 if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IMPORT_ONLY))
4456 tiIsVerifiableCode = FALSE;
4458 #endif //_TARGET_64BIT_
4459 verResetCurrentState(block, &verCurrentState);
4460 verConvertBBToThrowVerificationException(block DEBUGARG(logMsg));
4463 impNoteLastILoffs(); // Remember at which BC offset the tree was finished
4467 /******************************************************************************/
4468 typeInfo Compiler::verMakeTypeInfo(CorInfoType ciType, CORINFO_CLASS_HANDLE clsHnd)
4470 assert(ciType < CORINFO_TYPE_COUNT);
4475 case CORINFO_TYPE_STRING:
4476 case CORINFO_TYPE_CLASS:
4477 tiResult = verMakeTypeInfo(clsHnd);
4478 if (!tiResult.IsType(TI_REF))
4479 { // type must be consistent with element type
4484 #ifdef _TARGET_64BIT_
4485 case CORINFO_TYPE_NATIVEINT:
4486 case CORINFO_TYPE_NATIVEUINT:
4489 // If we have more precise information, use it
4490 return verMakeTypeInfo(clsHnd);
4494 return typeInfo::nativeInt();
4497 #endif // _TARGET_64BIT_
4499 case CORINFO_TYPE_VALUECLASS:
4500 case CORINFO_TYPE_REFANY:
4501 tiResult = verMakeTypeInfo(clsHnd);
4502 // type must be constant with element type;
4503 if (!tiResult.IsValueClass())
4508 case CORINFO_TYPE_VAR:
4509 return verMakeTypeInfo(clsHnd);
4511 case CORINFO_TYPE_PTR: // for now, pointers are treated as an error
4512 case CORINFO_TYPE_VOID:
4516 case CORINFO_TYPE_BYREF:
4518 CORINFO_CLASS_HANDLE childClassHandle;
4519 CorInfoType childType = info.compCompHnd->getChildType(clsHnd, &childClassHandle);
4520 return ByRef(verMakeTypeInfo(childType, childClassHandle));
4526 { // If we have more precise information, use it
4527 return typeInfo(TI_STRUCT, clsHnd);
4531 return typeInfo(JITtype2tiType(ciType));
4537 /******************************************************************************/
4539 typeInfo Compiler::verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd, bool bashStructToRef /* = false */)
4541 if (clsHnd == nullptr)
4546 // Byrefs should only occur in method and local signatures, which are accessed
4547 // using ICorClassInfo and ICorClassInfo.getChildType.
4548 // So findClass() and getClassAttribs() should not be called for byrefs
4550 if (JITtype2varType(info.compCompHnd->asCorInfoType(clsHnd)) == TYP_BYREF)
4552 assert(!"Did findClass() return a Byref?");
4556 unsigned attribs = info.compCompHnd->getClassAttribs(clsHnd);
4558 if (attribs & CORINFO_FLG_VALUECLASS)
4560 CorInfoType t = info.compCompHnd->getTypeForPrimitiveValueClass(clsHnd);
4562 // Meta-data validation should ensure that CORINF_TYPE_BYREF should
4563 // not occur here, so we may want to change this to an assert instead.
4564 if (t == CORINFO_TYPE_VOID || t == CORINFO_TYPE_BYREF || t == CORINFO_TYPE_PTR)
4569 #ifdef _TARGET_64BIT_
4570 if (t == CORINFO_TYPE_NATIVEINT || t == CORINFO_TYPE_NATIVEUINT)
4572 return typeInfo::nativeInt();
4574 #endif // _TARGET_64BIT_
4576 if (t != CORINFO_TYPE_UNDEF)
4578 return (typeInfo(JITtype2tiType(t)));
4580 else if (bashStructToRef)
4582 return (typeInfo(TI_REF, clsHnd));
4586 return (typeInfo(TI_STRUCT, clsHnd));
4589 else if (attribs & CORINFO_FLG_GENERIC_TYPE_VARIABLE)
4591 // See comment in _typeInfo.h for why we do it this way.
4592 return (typeInfo(TI_REF, clsHnd, true));
4596 return (typeInfo(TI_REF, clsHnd));
4600 /******************************************************************************/
4601 BOOL Compiler::verIsSDArray(typeInfo ti)
4603 if (ti.IsNullObjRef())
4604 { // nulls are SD arrays
4608 if (!ti.IsType(TI_REF))
4613 if (!info.compCompHnd->isSDArray(ti.GetClassHandleForObjRef()))
4620 /******************************************************************************/
4621 /* Given 'arrayObjectType' which is an array type, fetch the element type. */
4622 /* Returns an error type if anything goes wrong */
4624 typeInfo Compiler::verGetArrayElemType(typeInfo arrayObjectType)
4626 assert(!arrayObjectType.IsNullObjRef()); // you need to check for null explictly since that is a success case
4628 if (!verIsSDArray(arrayObjectType))
4633 CORINFO_CLASS_HANDLE childClassHandle = nullptr;
4634 CorInfoType ciType = info.compCompHnd->getChildType(arrayObjectType.GetClassHandleForObjRef(), &childClassHandle);
4636 return verMakeTypeInfo(ciType, childClassHandle);
4639 /*****************************************************************************
4641 typeInfo Compiler::verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args)
4643 CORINFO_CLASS_HANDLE classHandle;
4644 CorInfoType ciType = strip(info.compCompHnd->getArgType(sig, args, &classHandle));
4646 var_types type = JITtype2varType(ciType);
4647 if (varTypeIsGC(type))
4649 // For efficiency, getArgType only returns something in classHandle for
4650 // value types. For other types that have addition type info, you
4651 // have to call back explicitly
4652 classHandle = info.compCompHnd->getArgClass(sig, args);
4655 NO_WAY("Could not figure out Class specified in argument or local signature");
4659 return verMakeTypeInfo(ciType, classHandle);
4662 /*****************************************************************************/
4664 // This does the expensive check to figure out whether the method
4665 // needs to be verified. It is called only when we fail verification,
4666 // just before throwing the verification exception.
4668 BOOL Compiler::verNeedsVerification()
4670 // If we have previously determined that verification is NOT needed
4671 // (for example in Compiler::compCompile), that means verification is really not needed.
4672 // Return the same decision we made before.
4673 // (Note: This literally means that tiVerificationNeeded can never go from 0 to 1.)
4675 if (!tiVerificationNeeded)
4677 return tiVerificationNeeded;
4680 assert(tiVerificationNeeded);
4682 // Ok, we haven't concluded that verification is NOT needed. Consult the EE now to
4683 // obtain the answer.
4684 CorInfoCanSkipVerificationResult canSkipVerificationResult =
4685 info.compCompHnd->canSkipMethodVerification(info.compMethodHnd);
4687 // canSkipVerification will return one of the following three values:
4688 // CORINFO_VERIFICATION_CANNOT_SKIP = 0, // Cannot skip verification during jit time.
4689 // CORINFO_VERIFICATION_CAN_SKIP = 1, // Can skip verification during jit time.
4690 // CORINFO_VERIFICATION_RUNTIME_CHECK = 2, // Skip verification during jit time,
4691 // but need to insert a callout to the VM to ask during runtime
4692 // whether to skip verification or not.
4694 // Set tiRuntimeCalloutNeeded if canSkipVerification() instructs us to insert a callout for runtime check
4695 if (canSkipVerificationResult == CORINFO_VERIFICATION_RUNTIME_CHECK)
4697 tiRuntimeCalloutNeeded = true;
4700 if (canSkipVerificationResult == CORINFO_VERIFICATION_DONT_JIT)
4702 // Dev10 706080 - Testers don't like the assert, so just silence it
4703 // by not using the macros that invoke debugAssert.
4707 // When tiVerificationNeeded is true, JIT will do the verification during JIT time.
4708 // The following line means we will NOT do jit time verification if canSkipVerification
4709 // returns CORINFO_VERIFICATION_CAN_SKIP or CORINFO_VERIFICATION_RUNTIME_CHECK.
4710 tiVerificationNeeded = (canSkipVerificationResult == CORINFO_VERIFICATION_CANNOT_SKIP);
4711 return tiVerificationNeeded;
4714 BOOL Compiler::verIsByRefLike(const typeInfo& ti)
4720 if (!ti.IsType(TI_STRUCT))
4724 return info.compCompHnd->getClassAttribs(ti.GetClassHandleForValueClass()) & CORINFO_FLG_CONTAINS_STACK_PTR;
4727 BOOL Compiler::verIsSafeToReturnByRef(const typeInfo& ti)
4729 if (ti.IsPermanentHomeByRef())
4739 BOOL Compiler::verIsBoxable(const typeInfo& ti)
4741 return (ti.IsPrimitiveType() || ti.IsObjRef() // includes boxed generic type variables
4742 || ti.IsUnboxedGenericTypeVar() ||
4743 (ti.IsType(TI_STRUCT) &&
4744 // exclude byreflike structs
4745 !(info.compCompHnd->getClassAttribs(ti.GetClassHandleForValueClass()) & CORINFO_FLG_CONTAINS_STACK_PTR)));
4748 // Is it a boxed value type?
4749 bool Compiler::verIsBoxedValueType(typeInfo ti)
4751 if (ti.GetType() == TI_REF)
4753 CORINFO_CLASS_HANDLE clsHnd = ti.GetClassHandleForObjRef();
4754 return !!eeIsValueClass(clsHnd);
4762 /*****************************************************************************
4764 * Check if a TailCall is legal.
4767 bool Compiler::verCheckTailCallConstraint(
4769 CORINFO_RESOLVED_TOKEN* pResolvedToken,
4770 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a type parameter?
4771 bool speculative // If true, won't throw if verificatoin fails. Instead it will
4772 // return false to the caller.
4773 // If false, it will throw.
4777 CORINFO_SIG_INFO sig;
4778 unsigned int popCount = 0; // we can't pop the stack since impImportCall needs it, so
4779 // this counter is used to keep track of how many items have been
4782 CORINFO_METHOD_HANDLE methodHnd = nullptr;
4783 CORINFO_CLASS_HANDLE methodClassHnd = nullptr;
4784 unsigned methodClassFlgs = 0;
4786 assert(impOpcodeIsCallOpcode(opcode));
4788 if (compIsForInlining())
4793 // for calli, VerifyOrReturn that this is not a virtual method
4794 if (opcode == CEE_CALLI)
4796 /* Get the call sig */
4797 eeGetSig(pResolvedToken->token, info.compScopeHnd, impTokenLookupContextHandle, &sig);
4799 // We don't know the target method, so we have to infer the flags, or
4800 // assume the worst-case.
4801 mflags = (sig.callConv & CORINFO_CALLCONV_HASTHIS) ? 0 : CORINFO_FLG_STATIC;
4805 methodHnd = pResolvedToken->hMethod;
4807 mflags = info.compCompHnd->getMethodAttribs(methodHnd);
4809 // When verifying generic code we pair the method handle with its
4810 // owning class to get the exact method signature.
4811 methodClassHnd = pResolvedToken->hClass;
4812 assert(methodClassHnd);
4814 eeGetMethodSig(methodHnd, &sig, methodClassHnd);
4816 // opcode specific check
4817 methodClassFlgs = info.compCompHnd->getClassAttribs(methodClassHnd);
4820 // We must have got the methodClassHnd if opcode is not CEE_CALLI
4821 assert((methodHnd != nullptr && methodClassHnd != nullptr) || opcode == CEE_CALLI);
4823 if ((sig.callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG)
4825 eeGetCallSiteSig(pResolvedToken->token, info.compScopeHnd, impTokenLookupContextHandle, &sig);
4828 // check compatibility of the arguments
4829 unsigned int argCount;
4830 argCount = sig.numArgs;
4831 CORINFO_ARG_LIST_HANDLE args;
4835 typeInfo tiDeclared = verParseArgSigToTypeInfo(&sig, args).NormaliseForStack();
4837 // check that the argument is not a byref for tailcalls
4838 VerifyOrReturnSpeculative(!verIsByRefLike(tiDeclared), "tailcall on byrefs", speculative);
4840 // For unsafe code, we might have parameters containing pointer to the stack location.
4841 // Disallow the tailcall for this kind.
4842 CORINFO_CLASS_HANDLE classHandle;
4843 CorInfoType ciType = strip(info.compCompHnd->getArgType(&sig, args, &classHandle));
4844 VerifyOrReturnSpeculative(ciType != CORINFO_TYPE_PTR, "tailcall on CORINFO_TYPE_PTR", speculative);
4846 args = info.compCompHnd->getArgNext(args);
4850 popCount += sig.numArgs;
4852 // check for 'this' which is on non-static methods, not called via NEWOBJ
4853 if (!(mflags & CORINFO_FLG_STATIC))
4855 // Always update the popCount.
4856 // This is crucial for the stack calculation to be correct.
4857 typeInfo tiThis = impStackTop(popCount).seTypeInfo;
4860 if (opcode == CEE_CALLI)
4862 // For CALLI, we don't know the methodClassHnd. Therefore, let's check the "this" object
4864 if (tiThis.IsValueClass())
4868 VerifyOrReturnSpeculative(!verIsByRefLike(tiThis), "byref in tailcall", speculative);
4872 // Check type compatibility of the this argument
4873 typeInfo tiDeclaredThis = verMakeTypeInfo(methodClassHnd);
4874 if (tiDeclaredThis.IsValueClass())
4876 tiDeclaredThis.MakeByRef();
4879 VerifyOrReturnSpeculative(!verIsByRefLike(tiDeclaredThis), "byref in tailcall", speculative);
4883 // Tail calls on constrained calls should be illegal too:
4884 // when instantiated at a value type, a constrained call may pass the address of a stack allocated value
4885 VerifyOrReturnSpeculative(!pConstrainedResolvedToken, "byref in constrained tailcall", speculative);
4887 // Get the exact view of the signature for an array method
4888 if (sig.retType != CORINFO_TYPE_VOID)
4890 if (methodClassFlgs & CORINFO_FLG_ARRAY)
4892 assert(opcode != CEE_CALLI);
4893 eeGetCallSiteSig(pResolvedToken->token, info.compScopeHnd, impTokenLookupContextHandle, &sig);
4897 typeInfo tiCalleeRetType = verMakeTypeInfo(sig.retType, sig.retTypeClass);
4898 typeInfo tiCallerRetType =
4899 verMakeTypeInfo(info.compMethodInfo->args.retType, info.compMethodInfo->args.retTypeClass);
4901 // void return type gets morphed into the error type, so we have to treat them specially here
4902 if (sig.retType == CORINFO_TYPE_VOID)
4904 VerifyOrReturnSpeculative(info.compMethodInfo->args.retType == CORINFO_TYPE_VOID, "tailcall return mismatch",
4909 VerifyOrReturnSpeculative(tiCompatibleWith(NormaliseForStack(tiCalleeRetType),
4910 NormaliseForStack(tiCallerRetType), true),
4911 "tailcall return mismatch", speculative);
4914 // for tailcall, stack must be empty
4915 VerifyOrReturnSpeculative(verCurrentState.esStackDepth == popCount, "stack non-empty on tailcall", speculative);
4917 return true; // Yes, tailcall is legal
4920 /*****************************************************************************
4922 * Checks the IL verification rules for the call
4925 void Compiler::verVerifyCall(OPCODE opcode,
4926 CORINFO_RESOLVED_TOKEN* pResolvedToken,
4927 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
4930 const BYTE* delegateCreateStart,
4931 const BYTE* codeAddr,
4932 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName))
4935 CORINFO_SIG_INFO* sig = nullptr;
4936 unsigned int popCount = 0; // we can't pop the stack since impImportCall needs it, so
4937 // this counter is used to keep track of how many items have been
4940 // for calli, VerifyOrReturn that this is not a virtual method
4941 if (opcode == CEE_CALLI)
4943 Verify(false, "Calli not verifiable");
4947 //<NICE> It would be nice to cache the rest of it, but eeFindMethod is the big ticket item.
4948 mflags = callInfo->verMethodFlags;
4950 sig = &callInfo->verSig;
4952 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG)
4954 eeGetCallSiteSig(pResolvedToken->token, pResolvedToken->tokenScope, pResolvedToken->tokenContext, sig);
4957 // opcode specific check
4958 unsigned methodClassFlgs = callInfo->classFlags;
4962 // cannot do callvirt on valuetypes
4963 VerifyOrReturn(!(methodClassFlgs & CORINFO_FLG_VALUECLASS), "callVirt on value class");
4964 VerifyOrReturn(sig->hasThis(), "CallVirt on static method");
4969 assert(!tailCall); // Importer should not allow this
4970 VerifyOrReturn((mflags & CORINFO_FLG_CONSTRUCTOR) && !(mflags & CORINFO_FLG_STATIC),
4971 "newobj must be on instance");
4973 if (methodClassFlgs & CORINFO_FLG_DELEGATE)
4975 VerifyOrReturn(sig->numArgs == 2, "wrong number args to delegate ctor");
4976 typeInfo tiDeclaredObj = verParseArgSigToTypeInfo(sig, sig->args).NormaliseForStack();
4977 typeInfo tiDeclaredFtn =
4978 verParseArgSigToTypeInfo(sig, info.compCompHnd->getArgNext(sig->args)).NormaliseForStack();
4979 VerifyOrReturn(tiDeclaredFtn.IsNativeIntType(), "ftn arg needs to be a native int type");
4981 assert(popCount == 0);
4982 typeInfo tiActualObj = impStackTop(1).seTypeInfo;
4983 typeInfo tiActualFtn = impStackTop(0).seTypeInfo;
4985 VerifyOrReturn(tiActualFtn.IsMethod(), "delegate needs method as first arg");
4986 VerifyOrReturn(tiCompatibleWith(tiActualObj, tiDeclaredObj, true), "delegate object type mismatch");
4987 VerifyOrReturn(tiActualObj.IsNullObjRef() || tiActualObj.IsType(TI_REF),
4988 "delegate object type mismatch");
4990 CORINFO_CLASS_HANDLE objTypeHandle =
4991 tiActualObj.IsNullObjRef() ? nullptr : tiActualObj.GetClassHandleForObjRef();
4993 // the method signature must be compatible with the delegate's invoke method
4995 // check that for virtual functions, the type of the object used to get the
4996 // ftn ptr is the same as the type of the object passed to the delegate ctor.
4997 // since this is a bit of work to determine in general, we pattern match stylized
5000 // the delegate creation code check, which used to be done later, is now done here
5001 // so we can read delegateMethodRef directly from
5002 // from the preceding LDFTN or CEE_LDVIRTFN instruction sequence;
5003 // we then use it in our call to isCompatibleDelegate().
5005 mdMemberRef delegateMethodRef = mdMemberRefNil;
5006 VerifyOrReturn(verCheckDelegateCreation(delegateCreateStart, codeAddr, delegateMethodRef),
5007 "must create delegates with certain IL");
5009 CORINFO_RESOLVED_TOKEN delegateResolvedToken;
5010 delegateResolvedToken.tokenContext = impTokenLookupContextHandle;
5011 delegateResolvedToken.tokenScope = info.compScopeHnd;
5012 delegateResolvedToken.token = delegateMethodRef;
5013 delegateResolvedToken.tokenType = CORINFO_TOKENKIND_Method;
5014 info.compCompHnd->resolveToken(&delegateResolvedToken);
5016 CORINFO_CALL_INFO delegateCallInfo;
5017 eeGetCallInfo(&delegateResolvedToken, nullptr /* constraint typeRef */,
5018 addVerifyFlag(CORINFO_CALLINFO_SECURITYCHECKS), &delegateCallInfo);
5020 BOOL isOpenDelegate = FALSE;
5021 VerifyOrReturn(info.compCompHnd->isCompatibleDelegate(objTypeHandle, delegateResolvedToken.hClass,
5022 tiActualFtn.GetMethod(), pResolvedToken->hClass,
5024 "function incompatible with delegate");
5026 // check the constraints on the target method
5027 VerifyOrReturn(info.compCompHnd->satisfiesClassConstraints(delegateResolvedToken.hClass),
5028 "delegate target has unsatisfied class constraints");
5029 VerifyOrReturn(info.compCompHnd->satisfiesMethodConstraints(delegateResolvedToken.hClass,
5030 tiActualFtn.GetMethod()),
5031 "delegate target has unsatisfied method constraints");
5033 // See ECMA spec section 1.8.1.5.2 (Delegating via instance dispatch)
5034 // for additional verification rules for delegates
5035 CORINFO_METHOD_HANDLE actualMethodHandle = tiActualFtn.GetMethod();
5036 DWORD actualMethodAttribs = info.compCompHnd->getMethodAttribs(actualMethodHandle);
5037 if (impIsLDFTN_TOKEN(delegateCreateStart, codeAddr))
5040 if ((actualMethodAttribs & CORINFO_FLG_VIRTUAL) && ((actualMethodAttribs & CORINFO_FLG_FINAL) == 0)
5042 && StrictCheckForNonVirtualCallToVirtualMethod()
5046 if (info.compCompHnd->shouldEnforceCallvirtRestriction(info.compScopeHnd))
5048 VerifyOrReturn(tiActualObj.IsThisPtr() && lvaIsOriginalThisReadOnly() ||
5049 verIsBoxedValueType(tiActualObj),
5050 "The 'this' parameter to the call must be either the calling method's "
5051 "'this' parameter or "
5052 "a boxed value type.");
5057 if (actualMethodAttribs & CORINFO_FLG_PROTECTED)
5059 BOOL targetIsStatic = actualMethodAttribs & CORINFO_FLG_STATIC;
5061 Verify(targetIsStatic || !isOpenDelegate,
5062 "Unverifiable creation of an open instance delegate for a protected member.");
5064 CORINFO_CLASS_HANDLE instanceClassHnd = (tiActualObj.IsNullObjRef() || targetIsStatic)
5066 : tiActualObj.GetClassHandleForObjRef();
5068 // In the case of protected methods, it is a requirement that the 'this'
5069 // pointer be a subclass of the current context. Perform this check.
5070 Verify(info.compCompHnd->canAccessFamily(info.compMethodHnd, instanceClassHnd),
5071 "Accessing protected method through wrong type.");
5076 // fall thru to default checks
5078 VerifyOrReturn(!(mflags & CORINFO_FLG_ABSTRACT), "method abstract");
5080 VerifyOrReturn(!((mflags & CORINFO_FLG_CONSTRUCTOR) && (methodClassFlgs & CORINFO_FLG_DELEGATE)),
5081 "can only newobj a delegate constructor");
5083 // check compatibility of the arguments
5084 unsigned int argCount;
5085 argCount = sig->numArgs;
5086 CORINFO_ARG_LIST_HANDLE args;
5090 typeInfo tiActual = impStackTop(popCount + argCount).seTypeInfo;
5092 typeInfo tiDeclared = verParseArgSigToTypeInfo(sig, args).NormaliseForStack();
5093 VerifyOrReturn(tiCompatibleWith(tiActual, tiDeclared, true), "type mismatch");
5095 args = info.compCompHnd->getArgNext(args);
5101 popCount += sig->numArgs;
5103 // check for 'this' which are is non-static methods, not called via NEWOBJ
5104 CORINFO_CLASS_HANDLE instanceClassHnd = info.compClassHnd;
5105 if (!(mflags & CORINFO_FLG_STATIC) && (opcode != CEE_NEWOBJ))
5107 typeInfo tiThis = impStackTop(popCount).seTypeInfo;
5110 // If it is null, we assume we can access it (since it will AV shortly)
5111 // If it is anything but a reference class, there is no hierarchy, so
5112 // again, we don't need the precise instance class to compute 'protected' access
5113 if (tiThis.IsType(TI_REF))
5115 instanceClassHnd = tiThis.GetClassHandleForObjRef();
5118 // Check type compatibility of the this argument
5119 typeInfo tiDeclaredThis = verMakeTypeInfo(pResolvedToken->hClass);
5120 if (tiDeclaredThis.IsValueClass())
5122 tiDeclaredThis.MakeByRef();
5125 // If this is a call to the base class .ctor, set thisPtr Init for
5127 if (mflags & CORINFO_FLG_CONSTRUCTOR)
5129 if (verTrackObjCtorInitState && tiThis.IsThisPtr() &&
5130 verIsCallToInitThisPtr(info.compClassHnd, pResolvedToken->hClass))
5132 assert(verCurrentState.thisInitialized !=
5133 TIS_Bottom); // This should never be the case just from the logic of the verifier.
5134 VerifyOrReturn(verCurrentState.thisInitialized == TIS_Uninit,
5135 "Call to base class constructor when 'this' is possibly initialized");
5136 // Otherwise, 'this' is now initialized.
5137 verCurrentState.thisInitialized = TIS_Init;
5138 tiThis.SetInitialisedObjRef();
5142 // We allow direct calls to value type constructors
5143 // NB: we have to check that the contents of tiThis is a value type, otherwise we could use a
5144 // constrained callvirt to illegally re-enter a .ctor on a value of reference type.
5145 VerifyOrReturn(tiThis.IsByRef() && DereferenceByRef(tiThis).IsValueClass(),
5146 "Bad call to a constructor");
5150 if (pConstrainedResolvedToken != nullptr)
5152 VerifyOrReturn(tiThis.IsByRef(), "non-byref this type in constrained call");
5154 typeInfo tiConstraint = verMakeTypeInfo(pConstrainedResolvedToken->hClass);
5156 // We just dereference this and test for equality
5157 tiThis.DereferenceByRef();
5158 VerifyOrReturn(typeInfo::AreEquivalent(tiThis, tiConstraint),
5159 "this type mismatch with constrained type operand");
5161 // Now pretend the this type is the boxed constrained type, for the sake of subsequent checks
5162 tiThis = typeInfo(TI_REF, pConstrainedResolvedToken->hClass);
5165 // To support direct calls on readonly byrefs, just pretend tiDeclaredThis is readonly too
5166 if (tiDeclaredThis.IsByRef() && tiThis.IsReadonlyByRef())
5168 tiDeclaredThis.SetIsReadonlyByRef();
5171 VerifyOrReturn(tiCompatibleWith(tiThis, tiDeclaredThis, true), "this type mismatch");
5173 if (tiThis.IsByRef())
5175 // Find the actual type where the method exists (as opposed to what is declared
5176 // in the metadata). This is to prevent passing a byref as the "this" argument
5177 // while calling methods like System.ValueType.GetHashCode() which expect boxed objects.
5179 CORINFO_CLASS_HANDLE actualClassHnd = info.compCompHnd->getMethodClass(pResolvedToken->hMethod);
5180 VerifyOrReturn(eeIsValueClass(actualClassHnd),
5181 "Call to base type of valuetype (which is never a valuetype)");
5184 // Rules for non-virtual call to a non-final virtual method:
5187 // The "this" pointer is considered to be "possibly written" if
5188 // 1. Its address have been taken (LDARGA 0) anywhere in the method.
5190 // 2. It has been stored to (STARG.0) anywhere in the method.
5192 // A non-virtual call to a non-final virtual method is only allowed if
5193 // 1. The this pointer passed to the callee is an instance of a boxed value type.
5195 // 2. The this pointer passed to the callee is the current method's this pointer.
5196 // (and) The current method's this pointer is not "possibly written".
5198 // Thus the rule is that if you assign to this ANYWHERE you can't make "base" calls to
5199 // virtual methods. (Luckily this does affect .ctors, since they are not virtual).
5200 // This is stronger that is strictly needed, but implementing a laxer rule is significantly
5201 // hard and more error prone.
5203 if (opcode == CEE_CALL && (mflags & CORINFO_FLG_VIRTUAL) && ((mflags & CORINFO_FLG_FINAL) == 0)
5205 && StrictCheckForNonVirtualCallToVirtualMethod()
5209 if (info.compCompHnd->shouldEnforceCallvirtRestriction(info.compScopeHnd))
5212 tiThis.IsThisPtr() && lvaIsOriginalThisReadOnly() || verIsBoxedValueType(tiThis),
5213 "The 'this' parameter to the call must be either the calling method's 'this' parameter or "
5214 "a boxed value type.");
5219 // check any constraints on the callee's class and type parameters
5220 VerifyOrReturn(info.compCompHnd->satisfiesClassConstraints(pResolvedToken->hClass),
5221 "method has unsatisfied class constraints");
5222 VerifyOrReturn(info.compCompHnd->satisfiesMethodConstraints(pResolvedToken->hClass, pResolvedToken->hMethod),
5223 "method has unsatisfied method constraints");
5225 if (mflags & CORINFO_FLG_PROTECTED)
5227 VerifyOrReturn(info.compCompHnd->canAccessFamily(info.compMethodHnd, instanceClassHnd),
5228 "Can't access protected method");
5231 // Get the exact view of the signature for an array method
5232 if (sig->retType != CORINFO_TYPE_VOID)
5234 eeGetMethodSig(pResolvedToken->hMethod, sig, pResolvedToken->hClass);
5237 // "readonly." prefixed calls only allowed for the Address operation on arrays.
5238 // The methods supported by array types are under the control of the EE
5239 // so we can trust that only the Address operation returns a byref.
5242 typeInfo tiCalleeRetType = verMakeTypeInfo(sig->retType, sig->retTypeClass);
5243 VerifyOrReturn((methodClassFlgs & CORINFO_FLG_ARRAY) && tiCalleeRetType.IsByRef(),
5244 "unexpected use of readonly prefix");
5247 // Verify the tailcall
5250 verCheckTailCallConstraint(opcode, pResolvedToken, pConstrainedResolvedToken, false);
5254 /*****************************************************************************
5255 * Checks that a delegate creation is done using the following pattern:
5257 * ldvirtftn targetMemberRef
5259 * ldftn targetMemberRef
5261 * 'delegateCreateStart' points at the last dup or ldftn in this basic block (null if
5262 * not in this basic block)
5264 * targetMemberRef is read from the code sequence.
5265 * targetMemberRef is validated iff verificationNeeded.
5268 BOOL Compiler::verCheckDelegateCreation(const BYTE* delegateCreateStart,
5269 const BYTE* codeAddr,
5270 mdMemberRef& targetMemberRef)
5272 if (impIsLDFTN_TOKEN(delegateCreateStart, codeAddr))
5274 targetMemberRef = getU4LittleEndian(&delegateCreateStart[2]);
5277 else if (impIsDUP_LDVIRTFTN_TOKEN(delegateCreateStart, codeAddr))
5279 targetMemberRef = getU4LittleEndian(&delegateCreateStart[3]);
5286 typeInfo Compiler::verVerifySTIND(const typeInfo& tiTo, const typeInfo& value, const typeInfo& instrType)
5288 Verify(!tiTo.IsReadonlyByRef(), "write to readonly byref");
5289 typeInfo ptrVal = verVerifyLDIND(tiTo, instrType);
5290 typeInfo normPtrVal = typeInfo(ptrVal).NormaliseForStack();
5291 if (!tiCompatibleWith(value, normPtrVal, true))
5293 Verify(tiCompatibleWith(value, normPtrVal, true), "type mismatch");
5294 compUnsafeCastUsed = true;
5299 typeInfo Compiler::verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType)
5301 assert(!instrType.IsStruct());
5306 ptrVal = DereferenceByRef(ptr);
5307 if (instrType.IsObjRef() && !ptrVal.IsObjRef())
5309 Verify(false, "bad pointer");
5310 compUnsafeCastUsed = true;
5312 else if (!instrType.IsObjRef() && !typeInfo::AreEquivalent(instrType, ptrVal))
5314 Verify(false, "pointer not consistent with instr");
5315 compUnsafeCastUsed = true;
5320 Verify(false, "pointer not byref");
5321 compUnsafeCastUsed = true;
5327 // Verify that the field is used properly. 'tiThis' is NULL for statics,
5328 // 'fieldFlags' is the fields attributes, and mutator is TRUE if it is a
5329 // ld*flda or a st*fld.
5330 // 'enclosingClass' is given if we are accessing a field in some specific type.
5332 void Compiler::verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
5333 const CORINFO_FIELD_INFO& fieldInfo,
5334 const typeInfo* tiThis,
5336 BOOL allowPlainStructAsThis)
5338 CORINFO_CLASS_HANDLE enclosingClass = pResolvedToken->hClass;
5339 unsigned fieldFlags = fieldInfo.fieldFlags;
5340 CORINFO_CLASS_HANDLE instanceClass =
5341 info.compClassHnd; // for statics, we imagine the instance is the current class.
5343 bool isStaticField = ((fieldFlags & CORINFO_FLG_FIELD_STATIC) != 0);
5346 Verify(!(fieldFlags & CORINFO_FLG_FIELD_UNMANAGED), "mutating an RVA bases static");
5347 if ((fieldFlags & CORINFO_FLG_FIELD_FINAL))
5349 Verify((info.compFlags & CORINFO_FLG_CONSTRUCTOR) && enclosingClass == info.compClassHnd &&
5350 info.compIsStatic == isStaticField,
5351 "bad use of initonly field (set or address taken)");
5355 if (tiThis == nullptr)
5357 Verify(isStaticField, "used static opcode with non-static field");
5361 typeInfo tThis = *tiThis;
5363 if (allowPlainStructAsThis && tThis.IsValueClass())
5368 // If it is null, we assume we can access it (since it will AV shortly)
5369 // If it is anything but a refernce class, there is no hierarchy, so
5370 // again, we don't need the precise instance class to compute 'protected' access
5371 if (tiThis->IsType(TI_REF))
5373 instanceClass = tiThis->GetClassHandleForObjRef();
5376 // Note that even if the field is static, we require that the this pointer
5377 // satisfy the same constraints as a non-static field This happens to
5378 // be simpler and seems reasonable
5379 typeInfo tiDeclaredThis = verMakeTypeInfo(enclosingClass);
5380 if (tiDeclaredThis.IsValueClass())
5382 tiDeclaredThis.MakeByRef();
5384 // we allow read-only tThis, on any field access (even stores!), because if the
5385 // class implementor wants to prohibit stores he should make the field private.
5386 // we do this by setting the read-only bit on the type we compare tThis to.
5387 tiDeclaredThis.SetIsReadonlyByRef();
5389 else if (verTrackObjCtorInitState && tThis.IsThisPtr())
5391 // Any field access is legal on "uninitialized" this pointers.
5392 // The easiest way to implement this is to simply set the
5393 // initialized bit for the duration of the type check on the
5394 // field access only. It does not change the state of the "this"
5395 // for the function as a whole. Note that the "tThis" is a copy
5396 // of the original "this" type (*tiThis) passed in.
5397 tThis.SetInitialisedObjRef();
5400 Verify(tiCompatibleWith(tThis, tiDeclaredThis, true), "this type mismatch");
5403 // Presently the JIT does not check that we don't store or take the address of init-only fields
5404 // since we cannot guarantee their immutability and it is not a security issue.
5406 // check any constraints on the fields's class --- accessing the field might cause a class constructor to run.
5407 VerifyOrReturn(info.compCompHnd->satisfiesClassConstraints(enclosingClass),
5408 "field has unsatisfied class constraints");
5409 if (fieldFlags & CORINFO_FLG_FIELD_PROTECTED)
5411 Verify(info.compCompHnd->canAccessFamily(info.compMethodHnd, instanceClass),
5412 "Accessing protected method through wrong type.");
5416 void Compiler::verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode)
5418 if (tiOp1.IsNumberType())
5420 #ifdef _TARGET_64BIT_
5421 Verify(tiCompatibleWith(tiOp1, tiOp2, true), "Cond type mismatch");
5422 #else // _TARGET_64BIT
5423 // [10/17/2013] Consider changing this: to put on my verification lawyer hat,
5424 // this is non-conforming to the ECMA Spec: types don't have to be equivalent,
5425 // but compatible, since we can coalesce native int with int32 (see section III.1.5).
5426 Verify(typeInfo::AreEquivalent(tiOp1, tiOp2), "Cond type mismatch");
5427 #endif // !_TARGET_64BIT_
5429 else if (tiOp1.IsObjRef())
5441 Verify(FALSE, "Cond not allowed on object types");
5443 Verify(tiOp2.IsObjRef(), "Cond type mismatch");
5445 else if (tiOp1.IsByRef())
5447 Verify(tiOp2.IsByRef(), "Cond type mismatch");
5451 Verify(tiOp1.IsMethod() && tiOp2.IsMethod(), "Cond type mismatch");
5455 void Compiler::verVerifyThisPtrInitialised()
5457 if (verTrackObjCtorInitState)
5459 Verify(verCurrentState.thisInitialized == TIS_Init, "this ptr is not initialized");
5463 BOOL Compiler::verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target)
5465 // Either target == context, in this case calling an alternate .ctor
5466 // Or target is the immediate parent of context
5468 return ((target == context) || (target == info.compCompHnd->getParentType(context)));
5471 GenTree* Compiler::impImportLdvirtftn(GenTree* thisPtr,
5472 CORINFO_RESOLVED_TOKEN* pResolvedToken,
5473 CORINFO_CALL_INFO* pCallInfo)
5475 if ((pCallInfo->methodFlags & CORINFO_FLG_EnC) && !(pCallInfo->classFlags & CORINFO_FLG_INTERFACE))
5477 NO_WAY("Virtual call to a function added via EnC is not supported");
5480 // CoreRT generic virtual method
5481 if ((pCallInfo->sig.sigInst.methInstCount != 0) && IsTargetAbi(CORINFO_CORERT_ABI))
5483 GenTree* runtimeMethodHandle = nullptr;
5484 if (pCallInfo->exactContextNeedsRuntimeLookup)
5486 runtimeMethodHandle =
5487 impRuntimeLookupToTree(pResolvedToken, &pCallInfo->codePointerLookup, pCallInfo->hMethod);
5491 runtimeMethodHandle = gtNewIconEmbMethHndNode(pResolvedToken->hMethod);
5493 return gtNewHelperCallNode(CORINFO_HELP_GVMLOOKUP_FOR_SLOT, TYP_I_IMPL,
5494 gtNewArgList(thisPtr, runtimeMethodHandle));
5497 #ifdef FEATURE_READYTORUN_COMPILER
5498 if (opts.IsReadyToRun())
5500 if (!pCallInfo->exactContextNeedsRuntimeLookup)
5503 gtNewHelperCallNode(CORINFO_HELP_READYTORUN_VIRTUAL_FUNC_PTR, TYP_I_IMPL, gtNewArgList(thisPtr));
5505 call->setEntryPoint(pCallInfo->codePointerLookup.constLookup);
5510 // We need a runtime lookup. CoreRT has a ReadyToRun helper for that too.
5511 if (IsTargetAbi(CORINFO_CORERT_ABI))
5513 GenTree* ctxTree = getRuntimeContextTree(pCallInfo->codePointerLookup.lookupKind.runtimeLookupKind);
5515 return impReadyToRunHelperToTree(pResolvedToken, CORINFO_HELP_READYTORUN_GENERIC_HANDLE, TYP_I_IMPL,
5516 gtNewArgList(ctxTree), &pCallInfo->codePointerLookup.lookupKind);
5521 // Get the exact descriptor for the static callsite
5522 GenTree* exactTypeDesc = impParentClassTokenToHandle(pResolvedToken);
5523 if (exactTypeDesc == nullptr)
5524 { // compDonotInline()
5528 GenTree* exactMethodDesc = impTokenToHandle(pResolvedToken);
5529 if (exactMethodDesc == nullptr)
5530 { // compDonotInline()
5534 GenTreeArgList* helpArgs = gtNewArgList(exactMethodDesc);
5536 helpArgs = gtNewListNode(exactTypeDesc, helpArgs);
5538 helpArgs = gtNewListNode(thisPtr, helpArgs);
5540 // Call helper function. This gets the target address of the final destination callsite.
5542 return gtNewHelperCallNode(CORINFO_HELP_VIRTUAL_FUNC_PTR, TYP_I_IMPL, helpArgs);
5545 //------------------------------------------------------------------------
5546 // impImportAndPushBox: build and import a value-type box
5549 // pResolvedToken - resolved token from the box operation
5555 // The value to be boxed is popped from the stack, and a tree for
5556 // the boxed value is pushed. This method may create upstream
5557 // statements, spill side effecting trees, and create new temps.
5559 // If importing an inlinee, we may also discover the inline must
5560 // fail. If so there is no new value pushed on the stack. Callers
5561 // should use CompDoNotInline after calling this method to see if
5562 // ongoing importation should be aborted.
5565 // Boxing of ref classes results in the same value as the value on
5566 // the top of the stack, so is handled inline in impImportBlockCode
5567 // for the CEE_BOX case. Only value or primitive type boxes make it
5570 // Boxing for nullable types is done via a helper call; boxing
5571 // of other value types is expanded inline or handled via helper
5572 // call, depending on the jit's codegen mode.
5574 // When the jit is operating in size and time constrained modes,
5575 // using a helper call here can save jit time and code size. But it
5576 // also may inhibit cleanup optimizations that could have also had a
5577 // even greater benefit effect on code size and jit time. An optimal
5578 // strategy may need to peek ahead and see if it is easy to tell how
5579 // the box is being used. For now, we defer.
5581 void Compiler::impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken)
5583 // Spill any special side effects
5584 impSpillSpecialSideEff();
5586 // Get get the expression to box from the stack.
5587 GenTree* op1 = nullptr;
5588 GenTree* op2 = nullptr;
5589 StackEntry se = impPopStack();
5590 CORINFO_CLASS_HANDLE operCls = se.seTypeInfo.GetClassHandle();
5591 GenTree* exprToBox = se.val;
5593 // Look at what helper we should use.
5594 CorInfoHelpFunc boxHelper = info.compCompHnd->getBoxHelper(pResolvedToken->hClass);
5596 // Determine what expansion to prefer.
5598 // In size/time/debuggable constrained modes, the helper call
5599 // expansion for box is generally smaller and is preferred, unless
5600 // the value to box is a struct that comes from a call. In that
5601 // case the call can construct its return value directly into the
5602 // box payload, saving possibly some up-front zeroing.
5604 // Currently primitive type boxes always get inline expanded. We may
5605 // want to do the same for small structs if they don't come from
5606 // calls and don't have GC pointers, since explicitly copying such
5607 // structs is cheap.
5608 JITDUMP("\nCompiler::impImportAndPushBox -- handling BOX(value class) via");
5609 bool canExpandInline = (boxHelper == CORINFO_HELP_BOX);
5610 bool optForSize = !exprToBox->IsCall() && (operCls != nullptr) && (opts.compDbgCode || opts.MinOpts());
5611 bool expandInline = canExpandInline && !optForSize;
5615 JITDUMP(" inline allocate/copy sequence\n");
5617 // we are doing 'normal' boxing. This means that we can inline the box operation
5618 // Box(expr) gets morphed into
5619 // temp = new(clsHnd)
5620 // cpobj(temp+4, expr, clsHnd)
5622 // The code paths differ slightly below for structs and primitives because
5623 // "cpobj" differs in these cases. In one case you get
5624 // impAssignStructPtr(temp+4, expr, clsHnd)
5625 // and the other you get
5628 if (opts.MinOpts() || opts.compDbgCode)
5630 // For minopts/debug code, try and minimize the total number
5631 // of box temps by reusing an existing temp when possible.
5632 if (impBoxTempInUse || impBoxTemp == BAD_VAR_NUM)
5634 impBoxTemp = lvaGrabTemp(true DEBUGARG("Reusable Box Helper"));
5639 // When optimizing, use a new temp for each box operation
5640 // since we then know the exact class of the box temp.
5641 impBoxTemp = lvaGrabTemp(true DEBUGARG("Single-def Box Helper"));
5642 lvaTable[impBoxTemp].lvType = TYP_REF;
5643 const bool isExact = true;
5644 lvaSetClass(impBoxTemp, pResolvedToken->hClass, isExact);
5647 // needs to stay in use until this box expression is appended
5648 // some other node. We approximate this by keeping it alive until
5649 // the opcode stack becomes empty
5650 impBoxTempInUse = true;
5652 #ifdef FEATURE_READYTORUN_COMPILER
5653 bool usingReadyToRunHelper = false;
5655 if (opts.IsReadyToRun())
5657 op1 = impReadyToRunHelperToTree(pResolvedToken, CORINFO_HELP_READYTORUN_NEW, TYP_REF);
5658 usingReadyToRunHelper = (op1 != nullptr);
5661 if (!usingReadyToRunHelper)
5664 // TODO: ReadyToRun: When generic dictionary lookups are necessary, replace the lookup call
5665 // and the newfast call with a single call to a dynamic R2R cell that will:
5666 // 1) Load the context
5667 // 2) Perform the generic dictionary lookup and caching, and generate the appropriate stub
5668 // 3) Allocate and return the new object for boxing
5669 // Reason: performance (today, we'll always use the slow helper for the R2R generics case)
5671 // Ensure that the value class is restored
5672 op2 = impTokenToHandle(pResolvedToken, nullptr, TRUE /* mustRestoreHandle */);
5675 // We must be backing out of an inline.
5676 assert(compDonotInline());
5680 op1 = gtNewAllocObjNode(info.compCompHnd->getNewHelper(pResolvedToken, info.compMethodHnd),
5681 pResolvedToken->hClass, TYP_REF, op2);
5684 /* Remember that this basic block contains 'new' of an object, and so does this method */
5685 compCurBB->bbFlags |= BBF_HAS_NEWOBJ;
5686 optMethodFlags |= OMF_HAS_NEWOBJ;
5688 GenTree* asg = gtNewTempAssign(impBoxTemp, op1);
5690 GenTree* asgStmt = impAppendTree(asg, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
5692 op1 = gtNewLclvNode(impBoxTemp, TYP_REF);
5693 op2 = gtNewIconNode(TARGET_POINTER_SIZE, TYP_I_IMPL);
5694 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1, op2);
5696 if (varTypeIsStruct(exprToBox))
5698 assert(info.compCompHnd->getClassSize(pResolvedToken->hClass) == info.compCompHnd->getClassSize(operCls));
5699 op1 = impAssignStructPtr(op1, exprToBox, operCls, (unsigned)CHECK_SPILL_ALL);
5703 var_types lclTyp = exprToBox->TypeGet();
5704 if (lclTyp == TYP_BYREF)
5706 lclTyp = TYP_I_IMPL;
5708 CorInfoType jitType = info.compCompHnd->asCorInfoType(pResolvedToken->hClass);
5709 if (impIsPrimitive(jitType))
5711 lclTyp = JITtype2varType(jitType);
5713 assert(genActualType(exprToBox->TypeGet()) == genActualType(lclTyp) ||
5714 varTypeIsFloating(lclTyp) == varTypeIsFloating(exprToBox->TypeGet()));
5715 var_types srcTyp = exprToBox->TypeGet();
5716 var_types dstTyp = lclTyp;
5718 if (srcTyp != dstTyp)
5720 assert((varTypeIsFloating(srcTyp) && varTypeIsFloating(dstTyp)) ||
5721 (varTypeIsIntegral(srcTyp) && varTypeIsIntegral(dstTyp)));
5722 exprToBox = gtNewCastNode(dstTyp, exprToBox, false, dstTyp);
5724 op1 = gtNewAssignNode(gtNewOperNode(GT_IND, lclTyp, op1), exprToBox);
5727 // Spill eval stack to flush out any pending side effects.
5728 impSpillSideEffects(true, (unsigned)CHECK_SPILL_ALL DEBUGARG("impImportAndPushBox"));
5730 // Set up this copy as a second assignment.
5731 GenTree* copyStmt = impAppendTree(op1, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
5733 op1 = gtNewLclvNode(impBoxTemp, TYP_REF);
5735 // Record that this is a "box" node and keep track of the matching parts.
5736 op1 = new (this, GT_BOX) GenTreeBox(TYP_REF, op1, asgStmt, copyStmt);
5738 // If it is a value class, mark the "box" node. We can use this information
5739 // to optimise several cases:
5740 // "box(x) == null" --> false
5741 // "(box(x)).CallAnInterfaceMethod(...)" --> "(&x).CallAValueTypeMethod"
5742 // "(box(x)).CallAnObjectMethod(...)" --> "(&x).CallAValueTypeMethod"
5744 op1->gtFlags |= GTF_BOX_VALUE;
5745 assert(op1->IsBoxedValue());
5746 assert(asg->gtOper == GT_ASG);
5750 // Don't optimize, just call the helper and be done with it.
5751 JITDUMP(" helper call because: %s\n", canExpandInline ? "optimizing for size" : "nullable");
5752 assert(operCls != nullptr);
5754 // Ensure that the value class is restored
5755 op2 = impTokenToHandle(pResolvedToken, nullptr, TRUE /* mustRestoreHandle */);
5758 // We must be backing out of an inline.
5759 assert(compDonotInline());
5763 GenTreeArgList* args = gtNewArgList(op2, impGetStructAddr(exprToBox, operCls, (unsigned)CHECK_SPILL_ALL, true));
5764 op1 = gtNewHelperCallNode(boxHelper, TYP_REF, args);
5767 /* Push the result back on the stack, */
5768 /* even if clsHnd is a value class we want the TI_REF */
5769 typeInfo tiRetVal = typeInfo(TI_REF, info.compCompHnd->getTypeForBox(pResolvedToken->hClass));
5770 impPushOnStack(op1, tiRetVal);
5773 //------------------------------------------------------------------------
5774 // impImportNewObjArray: Build and import `new` of multi-dimmensional array
5777 // pResolvedToken - The CORINFO_RESOLVED_TOKEN that has been initialized
5778 // by a call to CEEInfo::resolveToken().
5779 // pCallInfo - The CORINFO_CALL_INFO that has been initialized
5780 // by a call to CEEInfo::getCallInfo().
5783 // The multi-dimensional array constructor arguments (array dimensions) are
5784 // pushed on the IL stack on entry to this method.
5787 // Multi-dimensional array constructors are imported as calls to a JIT
5788 // helper, not as regular calls.
5790 void Compiler::impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo)
5792 GenTree* classHandle = impParentClassTokenToHandle(pResolvedToken);
5793 if (classHandle == nullptr)
5794 { // compDonotInline()
5798 assert(pCallInfo->sig.numArgs);
5801 GenTreeArgList* args;
5804 // There are two different JIT helpers that can be used to allocate
5805 // multi-dimensional arrays:
5807 // - CORINFO_HELP_NEW_MDARR - takes the array dimensions as varargs.
5808 // This variant is deprecated. It should be eventually removed.
5810 // - CORINFO_HELP_NEW_MDARR_NONVARARG - takes the array dimensions as
5811 // pointer to block of int32s. This variant is more portable.
5813 // The non-varargs helper is enabled for CoreRT only for now. Enabling this
5814 // unconditionally would require ReadyToRun version bump.
5816 CLANG_FORMAT_COMMENT_ANCHOR;
5818 if (!opts.IsReadyToRun() || IsTargetAbi(CORINFO_CORERT_ABI))
5821 // Reuse the temp used to pass the array dimensions to avoid bloating
5822 // the stack frame in case there are multiple calls to multi-dim array
5823 // constructors within a single method.
5824 if (lvaNewObjArrayArgs == BAD_VAR_NUM)
5826 lvaNewObjArrayArgs = lvaGrabTemp(false DEBUGARG("NewObjArrayArgs"));
5827 lvaTable[lvaNewObjArrayArgs].lvType = TYP_BLK;
5828 lvaTable[lvaNewObjArrayArgs].lvExactSize = 0;
5831 // Increase size of lvaNewObjArrayArgs to be the largest size needed to hold 'numArgs' integers
5832 // for our call to CORINFO_HELP_NEW_MDARR_NONVARARG.
5833 lvaTable[lvaNewObjArrayArgs].lvExactSize =
5834 max(lvaTable[lvaNewObjArrayArgs].lvExactSize, pCallInfo->sig.numArgs * sizeof(INT32));
5836 // The side-effects may include allocation of more multi-dimensional arrays. Spill all side-effects
5837 // to ensure that the shared lvaNewObjArrayArgs local variable is only ever used to pass arguments
5838 // to one allocation at a time.
5839 impSpillSideEffects(true, (unsigned)CHECK_SPILL_ALL DEBUGARG("impImportNewObjArray"));
5842 // The arguments of the CORINFO_HELP_NEW_MDARR_NONVARARG helper are:
5843 // - Array class handle
5844 // - Number of dimension arguments
5845 // - Pointer to block of int32 dimensions - address of lvaNewObjArrayArgs temp.
5848 node = gtNewLclvNode(lvaNewObjArrayArgs, TYP_BLK);
5849 node = gtNewOperNode(GT_ADDR, TYP_I_IMPL, node);
5851 // Pop dimension arguments from the stack one at a time and store it
5852 // into lvaNewObjArrayArgs temp.
5853 for (int i = pCallInfo->sig.numArgs - 1; i >= 0; i--)
5855 GenTree* arg = impImplicitIorI4Cast(impPopStack().val, TYP_INT);
5857 GenTree* dest = gtNewLclvNode(lvaNewObjArrayArgs, TYP_BLK);
5858 dest = gtNewOperNode(GT_ADDR, TYP_I_IMPL, dest);
5859 dest = gtNewOperNode(GT_ADD, TYP_I_IMPL, dest,
5860 new (this, GT_CNS_INT) GenTreeIntCon(TYP_I_IMPL, sizeof(INT32) * i));
5861 dest = gtNewOperNode(GT_IND, TYP_INT, dest);
5863 node = gtNewOperNode(GT_COMMA, node->TypeGet(), gtNewAssignNode(dest, arg), node);
5866 args = gtNewArgList(node);
5868 // pass number of arguments to the helper
5869 args = gtNewListNode(gtNewIconNode(pCallInfo->sig.numArgs), args);
5871 args = gtNewListNode(classHandle, args);
5873 node = gtNewHelperCallNode(CORINFO_HELP_NEW_MDARR_NONVARARG, TYP_REF, args);
5878 // The varargs helper needs the type and method handles as last
5879 // and last-1 param (this is a cdecl call, so args will be
5880 // pushed in reverse order on the CPU stack)
5883 args = gtNewArgList(classHandle);
5885 // pass number of arguments to the helper
5886 args = gtNewListNode(gtNewIconNode(pCallInfo->sig.numArgs), args);
5888 unsigned argFlags = 0;
5889 args = impPopList(pCallInfo->sig.numArgs, &pCallInfo->sig, args);
5891 node = gtNewHelperCallNode(CORINFO_HELP_NEW_MDARR, TYP_REF, args);
5893 // varargs, so we pop the arguments
5894 node->gtFlags |= GTF_CALL_POP_ARGS;
5897 // At the present time we don't track Caller pop arguments
5898 // that have GC references in them
5899 for (GenTreeArgList* temp = args; temp; temp = temp->Rest())
5901 assert(temp->Current()->gtType != TYP_REF);
5906 node->gtFlags |= args->gtFlags & GTF_GLOB_EFFECT;
5907 node->gtCall.compileTimeHelperArgumentHandle = (CORINFO_GENERIC_HANDLE)pResolvedToken->hClass;
5909 // Remember that this basic block contains 'new' of a md array
5910 compCurBB->bbFlags |= BBF_HAS_NEWARRAY;
5912 impPushOnStack(node, typeInfo(TI_REF, pResolvedToken->hClass));
5915 GenTree* Compiler::impTransformThis(GenTree* thisPtr,
5916 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
5917 CORINFO_THIS_TRANSFORM transform)
5921 case CORINFO_DEREF_THIS:
5923 GenTree* obj = thisPtr;
5925 // This does a LDIND on the obj, which should be a byref. pointing to a ref
5926 impBashVarAddrsToI(obj);
5927 assert(genActualType(obj->gtType) == TYP_I_IMPL || obj->gtType == TYP_BYREF);
5928 CorInfoType constraintTyp = info.compCompHnd->asCorInfoType(pConstrainedResolvedToken->hClass);
5930 obj = gtNewOperNode(GT_IND, JITtype2varType(constraintTyp), obj);
5931 // ldind could point anywhere, example a boxed class static int
5932 obj->gtFlags |= (GTF_EXCEPT | GTF_GLOB_REF | GTF_IND_TGTANYWHERE);
5937 case CORINFO_BOX_THIS:
5939 // Constraint calls where there might be no
5940 // unboxed entry point require us to implement the call via helper.
5941 // These only occur when a possible target of the call
5942 // may have inherited an implementation of an interface
5943 // method from System.Object or System.ValueType. The EE does not provide us with
5944 // "unboxed" versions of these methods.
5946 GenTree* obj = thisPtr;
5948 assert(obj->TypeGet() == TYP_BYREF || obj->TypeGet() == TYP_I_IMPL);
5949 obj = gtNewObjNode(pConstrainedResolvedToken->hClass, obj);
5950 obj->gtFlags |= GTF_EXCEPT;
5952 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(pConstrainedResolvedToken->hClass);
5953 var_types objType = JITtype2varType(jitTyp);
5954 if (impIsPrimitive(jitTyp))
5956 if (obj->OperIsBlk())
5958 obj->ChangeOperUnchecked(GT_IND);
5960 // Obj could point anywhere, example a boxed class static int
5961 obj->gtFlags |= GTF_IND_TGTANYWHERE;
5962 obj->gtOp.gtOp2 = nullptr; // must be zero for tree walkers
5965 obj->gtType = JITtype2varType(jitTyp);
5966 assert(varTypeIsArithmetic(obj->gtType));
5969 // This pushes on the dereferenced byref
5970 // This is then used immediately to box.
5971 impPushOnStack(obj, verMakeTypeInfo(pConstrainedResolvedToken->hClass).NormaliseForStack());
5973 // This pops off the byref-to-a-value-type remaining on the stack and
5974 // replaces it with a boxed object.
5975 // This is then used as the object to the virtual call immediately below.
5976 impImportAndPushBox(pConstrainedResolvedToken);
5977 if (compDonotInline())
5982 obj = impPopStack().val;
5985 case CORINFO_NO_THIS_TRANSFORM:
5991 //------------------------------------------------------------------------
5992 // impCanPInvokeInline: check whether PInvoke inlining should enabled in current method.
5995 // true if PInvoke inlining should be enabled in current method, false otherwise
5998 // Checks a number of ambient conditions where we could pinvoke but choose not to
6000 bool Compiler::impCanPInvokeInline()
6002 return getInlinePInvokeEnabled() && (!opts.compDbgCode) && (compCodeOpt() != SMALL_CODE) &&
6003 (!opts.compNoPInvokeInlineCB) // profiler is preventing inline pinvoke
6007 //------------------------------------------------------------------------
6008 // impCanPInvokeInlineCallSite: basic legality checks using information
6009 // from a call to see if the call qualifies as an inline pinvoke.
6012 // block - block contaning the call, or for inlinees, block
6013 // containing the call being inlined
6016 // true if this call can legally qualify as an inline pinvoke, false otherwise
6019 // For runtimes that support exception handling interop there are
6020 // restrictions on using inline pinvoke in handler regions.
6022 // * We have to disable pinvoke inlining inside of filters because
6023 // in case the main execution (i.e. in the try block) is inside
6024 // unmanaged code, we cannot reuse the inlined stub (we still need
6025 // the original state until we are in the catch handler)
6027 // * We disable pinvoke inlining inside handlers since the GSCookie
6028 // is in the inlined Frame (see
6029 // CORINFO_EE_INFO::InlinedCallFrameInfo::offsetOfGSCookie), but
6030 // this would not protect framelets/return-address of handlers.
6032 // These restrictions are currently also in place for CoreCLR but
6033 // can be relaxed when coreclr/#8459 is addressed.
6035 bool Compiler::impCanPInvokeInlineCallSite(BasicBlock* block)
6037 if (block->hasHndIndex())
6042 // The remaining limitations do not apply to CoreRT
6043 if (IsTargetAbi(CORINFO_CORERT_ABI))
6048 #ifdef _TARGET_AMD64_
6049 // On x64, we disable pinvoke inlining inside of try regions.
6050 // Here is the comment from JIT64 explaining why:
6052 // [VSWhidbey: 611015] - because the jitted code links in the
6053 // Frame (instead of the stub) we rely on the Frame not being
6054 // 'active' until inside the stub. This normally happens by the
6055 // stub setting the return address pointer in the Frame object
6056 // inside the stub. On a normal return, the return address
6057 // pointer is zeroed out so the Frame can be safely re-used, but
6058 // if an exception occurs, nobody zeros out the return address
6059 // pointer. Thus if we re-used the Frame object, it would go
6060 // 'active' as soon as we link it into the Frame chain.
6062 // Technically we only need to disable PInvoke inlining if we're
6063 // in a handler or if we're in a try body with a catch or
6064 // filter/except where other non-handler code in this method
6065 // might run and try to re-use the dirty Frame object.
6067 // A desktop test case where this seems to matter is
6068 // jit\jit64\ebvts\mcpp\sources2\ijw\__clrcall\vector_ctor_dtor.02\deldtor_clr.exe
6069 if (block->hasTryIndex())
6073 #endif // _TARGET_AMD64_
6078 //------------------------------------------------------------------------
6079 // impCheckForPInvokeCall examine call to see if it is a pinvoke and if so
6080 // if it can be expressed as an inline pinvoke.
6083 // call - tree for the call
6084 // methHnd - handle for the method being called (may be null)
6085 // sig - signature of the method being called
6086 // mflags - method flags for the method being called
6087 // block - block contaning the call, or for inlinees, block
6088 // containing the call being inlined
6091 // Sets GTF_CALL_M_PINVOKE on the call for pinvokes.
6093 // Also sets GTF_CALL_UNMANAGED on call for inline pinvokes if the
6094 // call passes a combination of legality and profitabilty checks.
6096 // If GTF_CALL_UNMANAGED is set, increments info.compCallUnmanaged
6098 void Compiler::impCheckForPInvokeCall(
6099 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block)
6101 CorInfoUnmanagedCallConv unmanagedCallConv;
6103 // If VM flagged it as Pinvoke, flag the call node accordingly
6104 if ((mflags & CORINFO_FLG_PINVOKE) != 0)
6106 call->gtCallMoreFlags |= GTF_CALL_M_PINVOKE;
6111 if ((mflags & CORINFO_FLG_PINVOKE) == 0 || (mflags & CORINFO_FLG_NOSECURITYWRAP) == 0)
6116 unmanagedCallConv = info.compCompHnd->getUnmanagedCallConv(methHnd);
6120 CorInfoCallConv callConv = CorInfoCallConv(sig->callConv & CORINFO_CALLCONV_MASK);
6121 if (callConv == CORINFO_CALLCONV_NATIVEVARARG)
6123 // Used by the IL Stubs.
6124 callConv = CORINFO_CALLCONV_C;
6126 static_assert_no_msg((unsigned)CORINFO_CALLCONV_C == (unsigned)CORINFO_UNMANAGED_CALLCONV_C);
6127 static_assert_no_msg((unsigned)CORINFO_CALLCONV_STDCALL == (unsigned)CORINFO_UNMANAGED_CALLCONV_STDCALL);
6128 static_assert_no_msg((unsigned)CORINFO_CALLCONV_THISCALL == (unsigned)CORINFO_UNMANAGED_CALLCONV_THISCALL);
6129 unmanagedCallConv = CorInfoUnmanagedCallConv(callConv);
6131 assert(!call->gtCallCookie);
6134 if (unmanagedCallConv != CORINFO_UNMANAGED_CALLCONV_C && unmanagedCallConv != CORINFO_UNMANAGED_CALLCONV_STDCALL &&
6135 unmanagedCallConv != CORINFO_UNMANAGED_CALLCONV_THISCALL)
6139 optNativeCallCount++;
6141 if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB) && methHnd == nullptr)
6143 // PInvoke CALLI in IL stubs must be inlined
6148 if (!impCanPInvokeInlineCallSite(block))
6153 // PInvoke CALL in IL stubs must be inlined on CoreRT. Skip the ambient conditions checks and
6154 // profitability checks
6155 if (!(opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB) && IsTargetAbi(CORINFO_CORERT_ABI)))
6157 if (!impCanPInvokeInline())
6162 // Size-speed tradeoff: don't use inline pinvoke at rarely
6163 // executed call sites. The non-inline version is more
6165 if (block->isRunRarely())
6171 // The expensive check should be last
6172 if (info.compCompHnd->pInvokeMarshalingRequired(methHnd, sig))
6178 JITLOG((LL_INFO1000000, "\nInline a CALLI PINVOKE call from method %s", info.compFullName));
6180 call->gtFlags |= GTF_CALL_UNMANAGED;
6181 info.compCallUnmanaged++;
6183 // AMD64 convention is same for native and managed
6184 if (unmanagedCallConv == CORINFO_UNMANAGED_CALLCONV_C)
6186 call->gtFlags |= GTF_CALL_POP_ARGS;
6189 if (unmanagedCallConv == CORINFO_UNMANAGED_CALLCONV_THISCALL)
6191 call->gtCallMoreFlags |= GTF_CALL_M_UNMGD_THISCALL;
6195 GenTreeCall* Compiler::impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset)
6197 var_types callRetTyp = JITtype2varType(sig->retType);
6199 /* The function pointer is on top of the stack - It may be a
6200 * complex expression. As it is evaluated after the args,
6201 * it may cause registered args to be spilled. Simply spill it.
6204 // Ignore this trivial case.
6205 if (impStackTop().val->gtOper != GT_LCL_VAR)
6207 impSpillStackEntry(verCurrentState.esStackDepth - 1,
6208 BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impImportIndirectCall"));
6211 /* Get the function pointer */
6213 GenTree* fptr = impPopStack().val;
6215 // The function pointer is typically a sized to match the target pointer size
6216 // However, stubgen IL optimization can change LDC.I8 to LDC.I4
6217 // See ILCodeStream::LowerOpcode
6218 assert(genActualType(fptr->gtType) == TYP_I_IMPL || genActualType(fptr->gtType) == TYP_INT);
6221 // This temporary must never be converted to a double in stress mode,
6222 // because that can introduce a call to the cast helper after the
6223 // arguments have already been evaluated.
6225 if (fptr->OperGet() == GT_LCL_VAR)
6227 lvaTable[fptr->gtLclVarCommon.gtLclNum].lvKeepType = 1;
6231 /* Create the call node */
6233 GenTreeCall* call = gtNewIndCallNode(fptr, callRetTyp, nullptr, ilOffset);
6235 call->gtFlags |= GTF_EXCEPT | (fptr->gtFlags & GTF_GLOB_EFFECT);
6240 /*****************************************************************************/
6242 void Compiler::impPopArgsForUnmanagedCall(GenTree* call, CORINFO_SIG_INFO* sig)
6244 assert(call->gtFlags & GTF_CALL_UNMANAGED);
6246 /* Since we push the arguments in reverse order (i.e. right -> left)
6247 * spill any side effects from the stack
6249 * OBS: If there is only one side effect we do not need to spill it
6250 * thus we have to spill all side-effects except last one
6253 unsigned lastLevelWithSideEffects = UINT_MAX;
6255 unsigned argsToReverse = sig->numArgs;
6257 // For "thiscall", the first argument goes in a register. Since its
6258 // order does not need to be changed, we do not need to spill it
6260 if (call->gtCall.gtCallMoreFlags & GTF_CALL_M_UNMGD_THISCALL)
6262 assert(argsToReverse);
6266 #ifndef _TARGET_X86_
6267 // Don't reverse args on ARM or x64 - first four args always placed in regs in order
6271 for (unsigned level = verCurrentState.esStackDepth - argsToReverse; level < verCurrentState.esStackDepth; level++)
6273 if (verCurrentState.esStack[level].val->gtFlags & GTF_ORDER_SIDEEFF)
6275 assert(lastLevelWithSideEffects == UINT_MAX);
6277 impSpillStackEntry(level,
6278 BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impPopArgsForUnmanagedCall - other side effect"));
6280 else if (verCurrentState.esStack[level].val->gtFlags & GTF_SIDE_EFFECT)
6282 if (lastLevelWithSideEffects != UINT_MAX)
6284 /* We had a previous side effect - must spill it */
6285 impSpillStackEntry(lastLevelWithSideEffects,
6286 BAD_VAR_NUM DEBUGARG(false) DEBUGARG("impPopArgsForUnmanagedCall - side effect"));
6288 /* Record the level for the current side effect in case we will spill it */
6289 lastLevelWithSideEffects = level;
6293 /* This is the first side effect encountered - record its level */
6295 lastLevelWithSideEffects = level;
6300 /* The argument list is now "clean" - no out-of-order side effects
6301 * Pop the argument list in reverse order */
6303 GenTree* args = call->gtCall.gtCallArgs = impPopRevList(sig->numArgs, sig, sig->numArgs - argsToReverse);
6305 if (call->gtCall.gtCallMoreFlags & GTF_CALL_M_UNMGD_THISCALL)
6307 GenTree* thisPtr = args->Current();
6308 impBashVarAddrsToI(thisPtr);
6309 assert(thisPtr->TypeGet() == TYP_I_IMPL || thisPtr->TypeGet() == TYP_BYREF);
6314 call->gtFlags |= args->gtFlags & GTF_GLOB_EFFECT;
6318 //------------------------------------------------------------------------
6319 // impInitClass: Build a node to initialize the class before accessing the
6320 // field if necessary
6323 // pResolvedToken - The CORINFO_RESOLVED_TOKEN that has been initialized
6324 // by a call to CEEInfo::resolveToken().
6326 // Return Value: If needed, a pointer to the node that will perform the class
6327 // initializtion. Otherwise, nullptr.
6330 GenTree* Compiler::impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken)
6332 CorInfoInitClassResult initClassResult =
6333 info.compCompHnd->initClass(pResolvedToken->hField, info.compMethodHnd, impTokenLookupContextHandle);
6335 if ((initClassResult & CORINFO_INITCLASS_USE_HELPER) == 0)
6341 GenTree* node = impParentClassTokenToHandle(pResolvedToken, &runtimeLookup);
6343 if (node == nullptr)
6345 assert(compDonotInline());
6351 node = gtNewHelperCallNode(CORINFO_HELP_INITCLASS, TYP_VOID, gtNewArgList(node));
6355 // Call the shared non gc static helper, as its the fastest
6356 node = fgGetSharedCCtor(pResolvedToken->hClass);
6362 GenTree* Compiler::impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp)
6364 GenTree* op1 = nullptr;
6373 ival = *((bool*)fldAddr);
6377 ival = *((signed char*)fldAddr);
6381 ival = *((unsigned char*)fldAddr);
6385 ival = *((short*)fldAddr);
6389 ival = *((unsigned short*)fldAddr);
6394 ival = *((int*)fldAddr);
6396 op1 = gtNewIconNode(ival);
6401 lval = *((__int64*)fldAddr);
6402 op1 = gtNewLconNode(lval);
6406 dval = *((float*)fldAddr);
6407 op1 = gtNewDconNode(dval);
6408 #if !FEATURE_X87_DOUBLES
6409 // X87 stack doesn't differentiate between float/double
6410 // so R4 is treated as R8, but everybody else does
6411 op1->gtType = TYP_FLOAT;
6412 #endif // FEATURE_X87_DOUBLES
6416 dval = *((double*)fldAddr);
6417 op1 = gtNewDconNode(dval);
6421 assert(!"Unexpected lclTyp");
6428 GenTree* Compiler::impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6429 CORINFO_ACCESS_FLAGS access,
6430 CORINFO_FIELD_INFO* pFieldInfo,
6435 switch (pFieldInfo->fieldAccessor)
6437 case CORINFO_FIELD_STATIC_GENERICS_STATIC_HELPER:
6439 assert(!compIsForInlining());
6441 // We first call a special helper to get the statics base pointer
6442 op1 = impParentClassTokenToHandle(pResolvedToken);
6444 // compIsForInlining() is false so we should not neve get NULL here
6445 assert(op1 != nullptr);
6447 var_types type = TYP_BYREF;
6449 switch (pFieldInfo->helper)
6451 case CORINFO_HELP_GETGENERICS_NONGCTHREADSTATIC_BASE:
6454 case CORINFO_HELP_GETGENERICS_GCSTATIC_BASE:
6455 case CORINFO_HELP_GETGENERICS_NONGCSTATIC_BASE:
6456 case CORINFO_HELP_GETGENERICS_GCTHREADSTATIC_BASE:
6459 assert(!"unknown generic statics helper");
6463 op1 = gtNewHelperCallNode(pFieldInfo->helper, type, gtNewArgList(op1));
6465 FieldSeqNode* fs = GetFieldSeqStore()->CreateSingleton(pResolvedToken->hField);
6466 op1 = gtNewOperNode(GT_ADD, type, op1,
6467 new (this, GT_CNS_INT) GenTreeIntCon(TYP_I_IMPL, pFieldInfo->offset, fs));
6471 case CORINFO_FIELD_STATIC_SHARED_STATIC_HELPER:
6473 #ifdef FEATURE_READYTORUN_COMPILER
6474 if (opts.IsReadyToRun())
6476 unsigned callFlags = 0;
6478 if (info.compCompHnd->getClassAttribs(pResolvedToken->hClass) & CORINFO_FLG_BEFOREFIELDINIT)
6480 callFlags |= GTF_CALL_HOISTABLE;
6483 op1 = gtNewHelperCallNode(CORINFO_HELP_READYTORUN_STATIC_BASE, TYP_BYREF);
6484 op1->gtFlags |= callFlags;
6486 op1->gtCall.setEntryPoint(pFieldInfo->fieldLookup);
6491 op1 = fgGetStaticsCCtorHelper(pResolvedToken->hClass, pFieldInfo->helper);
6495 FieldSeqNode* fs = GetFieldSeqStore()->CreateSingleton(pResolvedToken->hField);
6496 op1 = gtNewOperNode(GT_ADD, op1->TypeGet(), op1,
6497 new (this, GT_CNS_INT) GenTreeIntCon(TYP_INT, pFieldInfo->offset, fs));
6502 case CORINFO_FIELD_STATIC_READYTORUN_HELPER:
6504 #ifdef FEATURE_READYTORUN_COMPILER
6505 noway_assert(opts.IsReadyToRun());
6506 CORINFO_LOOKUP_KIND kind = info.compCompHnd->getLocationOfThisType(info.compMethodHnd);
6507 assert(kind.needsRuntimeLookup);
6509 GenTree* ctxTree = getRuntimeContextTree(kind.runtimeLookupKind);
6510 GenTreeArgList* args = gtNewArgList(ctxTree);
6512 unsigned callFlags = 0;
6514 if (info.compCompHnd->getClassAttribs(pResolvedToken->hClass) & CORINFO_FLG_BEFOREFIELDINIT)
6516 callFlags |= GTF_CALL_HOISTABLE;
6518 var_types type = TYP_BYREF;
6519 op1 = gtNewHelperCallNode(CORINFO_HELP_READYTORUN_GENERIC_STATIC_BASE, type, args);
6520 op1->gtFlags |= callFlags;
6522 op1->gtCall.setEntryPoint(pFieldInfo->fieldLookup);
6523 FieldSeqNode* fs = GetFieldSeqStore()->CreateSingleton(pResolvedToken->hField);
6524 op1 = gtNewOperNode(GT_ADD, type, op1,
6525 new (this, GT_CNS_INT) GenTreeIntCon(TYP_I_IMPL, pFieldInfo->offset, fs));
6528 #endif // FEATURE_READYTORUN_COMPILER
6534 if (!(access & CORINFO_ACCESS_ADDRESS))
6536 // In future, it may be better to just create the right tree here instead of folding it later.
6537 op1 = gtNewFieldRef(lclTyp, pResolvedToken->hField);
6539 if (pFieldInfo->fieldFlags & CORINFO_FLG_FIELD_INITCLASS)
6541 op1->gtFlags |= GTF_FLD_INITCLASS;
6544 if (pFieldInfo->fieldFlags & CORINFO_FLG_FIELD_STATIC_IN_HEAP)
6546 op1->gtType = TYP_REF; // points at boxed object
6547 FieldSeqNode* firstElemFldSeq =
6548 GetFieldSeqStore()->CreateSingleton(FieldSeqStore::FirstElemPseudoField);
6549 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1,
6550 new (this, GT_CNS_INT)
6551 GenTreeIntCon(TYP_I_IMPL, TARGET_POINTER_SIZE, firstElemFldSeq));
6553 if (varTypeIsStruct(lclTyp))
6555 // Constructor adds GTF_GLOB_REF. Note that this is *not* GTF_EXCEPT.
6556 op1 = gtNewObjNode(pFieldInfo->structType, op1);
6560 op1 = gtNewOperNode(GT_IND, lclTyp, op1);
6561 op1->gtFlags |= GTF_GLOB_REF | GTF_IND_NONFAULTING;
6569 void** pFldAddr = nullptr;
6570 void* fldAddr = info.compCompHnd->getFieldAddress(pResolvedToken->hField, (void**)&pFldAddr);
6572 FieldSeqNode* fldSeq = GetFieldSeqStore()->CreateSingleton(pResolvedToken->hField);
6574 /* Create the data member node */
6575 op1 = gtNewIconHandleNode(pFldAddr == nullptr ? (size_t)fldAddr : (size_t)pFldAddr, GTF_ICON_STATIC_HDL,
6578 if (pFieldInfo->fieldFlags & CORINFO_FLG_FIELD_INITCLASS)
6580 op1->gtFlags |= GTF_ICON_INITCLASS;
6583 if (pFldAddr != nullptr)
6585 // There are two cases here, either the static is RVA based,
6586 // in which case the type of the FIELD node is not a GC type
6587 // and the handle to the RVA is a TYP_I_IMPL. Or the FIELD node is
6588 // a GC type and the handle to it is a TYP_BYREF in the GC heap
6589 // because handles to statics now go into the large object heap
6591 var_types handleTyp = (var_types)(varTypeIsGC(lclTyp) ? TYP_BYREF : TYP_I_IMPL);
6592 op1 = gtNewOperNode(GT_IND, handleTyp, op1);
6593 op1->gtFlags |= GTF_IND_INVARIANT | GTF_IND_NONFAULTING;
6600 if (pFieldInfo->fieldFlags & CORINFO_FLG_FIELD_STATIC_IN_HEAP)
6602 op1 = gtNewOperNode(GT_IND, TYP_REF, op1);
6604 FieldSeqNode* fldSeq = GetFieldSeqStore()->CreateSingleton(FieldSeqStore::FirstElemPseudoField);
6606 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1,
6607 new (this, GT_CNS_INT) GenTreeIntCon(TYP_I_IMPL, TARGET_POINTER_SIZE, fldSeq));
6610 if (!(access & CORINFO_ACCESS_ADDRESS))
6612 op1 = gtNewOperNode(GT_IND, lclTyp, op1);
6613 op1->gtFlags |= GTF_GLOB_REF;
6619 // In general try to call this before most of the verification work. Most people expect the access
6620 // exceptions before the verification exceptions. If you do this after, that usually doesn't happen. Turns
6621 // out if you can't access something we also think that you're unverifiable for other reasons.
6622 void Compiler::impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall)
6624 if (result != CORINFO_ACCESS_ALLOWED)
6626 impHandleAccessAllowedInternal(result, helperCall);
6630 void Compiler::impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall)
6634 case CORINFO_ACCESS_ALLOWED:
6636 case CORINFO_ACCESS_ILLEGAL:
6637 // if we're verifying, then we need to reject the illegal access to ensure that we don't think the
6638 // method is verifiable. Otherwise, delay the exception to runtime.
6639 if (compIsForImportOnly())
6641 info.compCompHnd->ThrowExceptionForHelper(helperCall);
6645 impInsertHelperCall(helperCall);
6648 case CORINFO_ACCESS_RUNTIME_CHECK:
6649 impInsertHelperCall(helperCall);
6654 void Compiler::impInsertHelperCall(CORINFO_HELPER_DESC* helperInfo)
6656 // Construct the argument list
6657 GenTreeArgList* args = nullptr;
6658 assert(helperInfo->helperNum != CORINFO_HELP_UNDEF);
6659 for (unsigned i = helperInfo->numArgs; i > 0; --i)
6661 const CORINFO_HELPER_ARG& helperArg = helperInfo->args[i - 1];
6662 GenTree* currentArg = nullptr;
6663 switch (helperArg.argType)
6665 case CORINFO_HELPER_ARG_TYPE_Field:
6666 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(
6667 info.compCompHnd->getFieldClass(helperArg.fieldHandle));
6668 currentArg = gtNewIconEmbFldHndNode(helperArg.fieldHandle);
6670 case CORINFO_HELPER_ARG_TYPE_Method:
6671 info.compCompHnd->methodMustBeLoadedBeforeCodeIsRun(helperArg.methodHandle);
6672 currentArg = gtNewIconEmbMethHndNode(helperArg.methodHandle);
6674 case CORINFO_HELPER_ARG_TYPE_Class:
6675 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(helperArg.classHandle);
6676 currentArg = gtNewIconEmbClsHndNode(helperArg.classHandle);
6678 case CORINFO_HELPER_ARG_TYPE_Module:
6679 currentArg = gtNewIconEmbScpHndNode(helperArg.moduleHandle);
6681 case CORINFO_HELPER_ARG_TYPE_Const:
6682 currentArg = gtNewIconNode(helperArg.constant);
6685 NO_WAY("Illegal helper arg type");
6687 args = (currentArg == nullptr) ? gtNewArgList(currentArg) : gtNewListNode(currentArg, args);
6691 * Mark as CSE'able, and hoistable. Consider marking hoistable unless you're in the inlinee.
6692 * Also, consider sticking this in the first basic block.
6694 GenTree* callout = gtNewHelperCallNode(helperInfo->helperNum, TYP_VOID, args);
6695 impAppendTree(callout, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
6698 // Checks whether the return types of caller and callee are compatible
6699 // so that callee can be tail called. Note that here we don't check
6700 // compatibility in IL Verifier sense, but on the lines of return type
6701 // sizes are equal and get returned in the same return register.
6702 bool Compiler::impTailCallRetTypeCompatible(var_types callerRetType,
6703 CORINFO_CLASS_HANDLE callerRetTypeClass,
6704 var_types calleeRetType,
6705 CORINFO_CLASS_HANDLE calleeRetTypeClass)
6707 // Note that we can not relax this condition with genActualType() as the
6708 // calling convention dictates that the caller of a function with a small
6709 // typed return value is responsible for normalizing the return val.
6710 if (callerRetType == calleeRetType)
6715 // If the class handles are the same and not null, the return types are compatible.
6716 if ((callerRetTypeClass != nullptr) && (callerRetTypeClass == calleeRetTypeClass))
6721 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
6723 if (callerRetType == TYP_VOID)
6725 // This needs to be allowed to support the following IL pattern that Jit64 allows:
6730 // Note that the above IL pattern is not valid as per IL verification rules.
6731 // Therefore, only full trust code can take advantage of this pattern.
6735 // These checks return true if the return value type sizes are the same and
6736 // get returned in the same return register i.e. caller doesn't need to normalize
6737 // return value. Some of the tail calls permitted by below checks would have
6738 // been rejected by IL Verifier before we reached here. Therefore, only full
6739 // trust code can make those tail calls.
6740 unsigned callerRetTypeSize = 0;
6741 unsigned calleeRetTypeSize = 0;
6742 bool isCallerRetTypMBEnreg =
6743 VarTypeIsMultiByteAndCanEnreg(callerRetType, callerRetTypeClass, &callerRetTypeSize, true);
6744 bool isCalleeRetTypMBEnreg =
6745 VarTypeIsMultiByteAndCanEnreg(calleeRetType, calleeRetTypeClass, &calleeRetTypeSize, true);
6747 if (varTypeIsIntegral(callerRetType) || isCallerRetTypMBEnreg)
6749 return (varTypeIsIntegral(calleeRetType) || isCalleeRetTypMBEnreg) && (callerRetTypeSize == calleeRetTypeSize);
6751 #endif // _TARGET_AMD64_ || _TARGET_ARM64_
6759 PREFIX_TAILCALL_EXPLICIT = 0x00000001, // call has "tail" IL prefix
6760 PREFIX_TAILCALL_IMPLICIT =
6761 0x00000010, // call is treated as having "tail" prefix even though there is no "tail" IL prefix
6762 PREFIX_TAILCALL = (PREFIX_TAILCALL_EXPLICIT | PREFIX_TAILCALL_IMPLICIT),
6763 PREFIX_VOLATILE = 0x00000100,
6764 PREFIX_UNALIGNED = 0x00001000,
6765 PREFIX_CONSTRAINED = 0x00010000,
6766 PREFIX_READONLY = 0x00100000
6769 /********************************************************************************
6771 * Returns true if the current opcode and and the opcodes following it correspond
6772 * to a supported tail call IL pattern.
6775 bool Compiler::impIsTailCallILPattern(bool tailPrefixed,
6777 const BYTE* codeAddrOfNextOpcode,
6778 const BYTE* codeEnd,
6780 bool* isCallPopAndRet /* = nullptr */)
6782 // Bail out if the current opcode is not a call.
6783 if (!impOpcodeIsCallOpcode(curOpcode))
6788 #if !FEATURE_TAILCALL_OPT_SHARED_RETURN
6789 // If shared ret tail opt is not enabled, we will enable
6790 // it for recursive methods.
6794 // we can actually handle if the ret is in a fallthrough block, as long as that is the only part of the
6795 // sequence. Make sure we don't go past the end of the IL however.
6796 codeEnd = min(codeEnd + 1, info.compCode + info.compILCodeSize);
6799 // Bail out if there is no next opcode after call
6800 if (codeAddrOfNextOpcode >= codeEnd)
6805 // Scan the opcodes to look for the following IL patterns if either
6806 // i) the call is not tail prefixed (i.e. implicit tail call) or
6807 // ii) if tail prefixed, IL verification is not needed for the method.
6809 // Only in the above two cases we can allow the below tail call patterns
6810 // violating ECMA spec.
6826 #if !defined(FEATURE_CORECLR) && defined(_TARGET_AMD64_)
6829 nextOpcode = (OPCODE)getU1LittleEndian(codeAddrOfNextOpcode);
6830 codeAddrOfNextOpcode += sizeof(__int8);
6831 } while ((codeAddrOfNextOpcode < codeEnd) && // Haven't reached end of method
6832 (!tailPrefixed || !tiVerificationNeeded) && // Not ".tail" prefixed or method requires no IL verification
6833 ((nextOpcode == CEE_NOP) || ((nextOpcode == CEE_POP) && (++cntPop == 1)))); // Next opcode = nop or exactly
6834 // one pop seen so far.
6836 nextOpcode = (OPCODE)getU1LittleEndian(codeAddrOfNextOpcode);
6837 #endif // !FEATURE_CORECLR && _TARGET_AMD64_
6839 if (isCallPopAndRet)
6841 // Allow call+pop+ret to be tail call optimized if caller ret type is void
6842 *isCallPopAndRet = (nextOpcode == CEE_RET) && (cntPop == 1);
6845 #if !defined(FEATURE_CORECLR) && defined(_TARGET_AMD64_)
6847 // Tail call IL pattern could be either of the following
6848 // 1) call/callvirt/calli + ret
6849 // 2) call/callvirt/calli + pop + ret in a method returning void.
6850 return (nextOpcode == CEE_RET) && ((cntPop == 0) || ((cntPop == 1) && (info.compRetType == TYP_VOID)));
6852 return (nextOpcode == CEE_RET) && (cntPop == 0);
6853 #endif // !FEATURE_CORECLR && _TARGET_AMD64_
6856 /*****************************************************************************
6858 * Determine whether the call could be converted to an implicit tail call
6861 bool Compiler::impIsImplicitTailCallCandidate(
6862 OPCODE opcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive)
6865 #if FEATURE_TAILCALL_OPT
6866 if (!opts.compTailCallOpt)
6871 if (opts.compDbgCode || opts.MinOpts())
6876 // must not be tail prefixed
6877 if (prefixFlags & PREFIX_TAILCALL_EXPLICIT)
6882 #if !FEATURE_TAILCALL_OPT_SHARED_RETURN
6883 // the block containing call is marked as BBJ_RETURN
6884 // We allow shared ret tail call optimization on recursive calls even under
6885 // !FEATURE_TAILCALL_OPT_SHARED_RETURN.
6886 if (!isRecursive && (compCurBB->bbJumpKind != BBJ_RETURN))
6888 #endif // !FEATURE_TAILCALL_OPT_SHARED_RETURN
6890 // must be call+ret or call+pop+ret
6891 if (!impIsTailCallILPattern(false, opcode, codeAddrOfNextOpcode, codeEnd, isRecursive))
6899 #endif // FEATURE_TAILCALL_OPT
6902 //------------------------------------------------------------------------
6903 // impImportCall: import a call-inspiring opcode
6906 // opcode - opcode that inspires the call
6907 // pResolvedToken - resolved token for the call target
6908 // pConstrainedResolvedToken - resolved constraint token (or nullptr)
6909 // newObjThis - tree for this pointer or uninitalized newobj temp (or nullptr)
6910 // prefixFlags - IL prefix flags for the call
6911 // callInfo - EE supplied info for the call
6912 // rawILOffset - IL offset of the opcode
6915 // Type of the call's return value.
6916 // If we're importing an inlinee and have realized the inline must fail, the call return type should be TYP_UNDEF.
6917 // However we can't assert for this here yet because there are cases we miss. See issue #13272.
6921 // opcode can be CEE_CALL, CEE_CALLI, CEE_CALLVIRT, or CEE_NEWOBJ.
6923 // For CEE_NEWOBJ, newobjThis should be the temp grabbed for the allocated
6924 // uninitalized object.
6927 #pragma warning(push)
6928 #pragma warning(disable : 21000) // Suppress PREFast warning about overly large function
6931 var_types Compiler::impImportCall(OPCODE opcode,
6932 CORINFO_RESOLVED_TOKEN* pResolvedToken,
6933 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
6934 GenTree* newobjThis,
6936 CORINFO_CALL_INFO* callInfo,
6937 IL_OFFSET rawILOffset)
6939 assert(opcode == CEE_CALL || opcode == CEE_CALLVIRT || opcode == CEE_NEWOBJ || opcode == CEE_CALLI);
6941 IL_OFFSETX ilOffset = impCurILOffset(rawILOffset, true);
6942 var_types callRetTyp = TYP_COUNT;
6943 CORINFO_SIG_INFO* sig = nullptr;
6944 CORINFO_METHOD_HANDLE methHnd = nullptr;
6945 CORINFO_CLASS_HANDLE clsHnd = nullptr;
6946 unsigned clsFlags = 0;
6947 unsigned mflags = 0;
6948 unsigned argFlags = 0;
6949 GenTree* call = nullptr;
6950 GenTreeArgList* args = nullptr;
6951 CORINFO_THIS_TRANSFORM constraintCallThisTransform = CORINFO_NO_THIS_TRANSFORM;
6952 CORINFO_CONTEXT_HANDLE exactContextHnd = nullptr;
6953 bool exactContextNeedsRuntimeLookup = false;
6954 bool canTailCall = true;
6955 const char* szCanTailCallFailReason = nullptr;
6956 int tailCall = prefixFlags & PREFIX_TAILCALL;
6957 bool readonlyCall = (prefixFlags & PREFIX_READONLY) != 0;
6959 CORINFO_RESOLVED_TOKEN* ldftnToken = nullptr;
6961 // Synchronized methods need to call CORINFO_HELP_MON_EXIT at the end. We could
6962 // do that before tailcalls, but that is probably not the intended
6963 // semantic. So just disallow tailcalls from synchronized methods.
6964 // Also, popping arguments in a varargs function is more work and NYI
6965 // If we have a security object, we have to keep our frame around for callers
6966 // to see any imperative security.
6967 if (info.compFlags & CORINFO_FLG_SYNCH)
6969 canTailCall = false;
6970 szCanTailCallFailReason = "Caller is synchronized";
6972 #if !FEATURE_FIXED_OUT_ARGS
6973 else if (info.compIsVarArgs)
6975 canTailCall = false;
6976 szCanTailCallFailReason = "Caller is varargs";
6978 #endif // FEATURE_FIXED_OUT_ARGS
6979 else if (opts.compNeedSecurityCheck)
6981 canTailCall = false;
6982 szCanTailCallFailReason = "Caller requires a security check.";
6985 // We only need to cast the return value of pinvoke inlined calls that return small types
6987 // TODO-AMD64-Cleanup: Remove this when we stop interoperating with JIT64, or if we decide to stop
6988 // widening everything! CoreCLR does not support JIT64 interoperation so no need to widen there.
6989 // The existing x64 JIT doesn't bother widening all types to int, so we have to assume for
6990 // the time being that the callee might be compiled by the other JIT and thus the return
6991 // value will need to be widened by us (or not widened at all...)
6993 // ReadyToRun code sticks with default calling convention that does not widen small return types.
6995 bool checkForSmallType = opts.IsJit64Compat() || opts.IsReadyToRun();
6996 bool bIntrinsicImported = false;
6998 CORINFO_SIG_INFO calliSig;
6999 GenTreeArgList* extraArg = nullptr;
7001 /*-------------------------------------------------------------------------
7002 * First create the call node
7005 if (opcode == CEE_CALLI)
7007 /* Get the call site sig */
7008 eeGetSig(pResolvedToken->token, info.compScopeHnd, impTokenLookupContextHandle, &calliSig);
7010 callRetTyp = JITtype2varType(calliSig.retType);
7012 call = impImportIndirectCall(&calliSig, ilOffset);
7014 // We don't know the target method, so we have to infer the flags, or
7015 // assume the worst-case.
7016 mflags = (calliSig.callConv & CORINFO_CALLCONV_HASTHIS) ? 0 : CORINFO_FLG_STATIC;
7021 unsigned structSize =
7022 (callRetTyp == TYP_STRUCT) ? info.compCompHnd->getClassSize(calliSig.retTypeSigClass) : 0;
7023 printf("\nIn Compiler::impImportCall: opcode is %s, kind=%d, callRetType is %s, structSize is %d\n",
7024 opcodeNames[opcode], callInfo->kind, varTypeName(callRetTyp), structSize);
7027 // This should be checked in impImportBlockCode.
7028 assert(!compIsForInlining() || !(impInlineInfo->inlineCandidateInfo->dwRestrictions & INLINE_RESPECT_BOUNDARY));
7033 // We cannot lazily obtain the signature of a CALLI call because it has no method
7034 // handle that we can use, so we need to save its full call signature here.
7035 assert(call->gtCall.callSig == nullptr);
7036 call->gtCall.callSig = new (this, CMK_CorSig) CORINFO_SIG_INFO;
7037 *call->gtCall.callSig = calliSig;
7040 if (IsTargetAbi(CORINFO_CORERT_ABI))
7042 bool managedCall = (((calliSig.callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_STDCALL) &&
7043 ((calliSig.callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_C) &&
7044 ((calliSig.callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_THISCALL) &&
7045 ((calliSig.callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_FASTCALL));
7048 addFatPointerCandidate(call->AsCall());
7052 else // (opcode != CEE_CALLI)
7054 CorInfoIntrinsics intrinsicID = CORINFO_INTRINSIC_Count;
7056 // Passing CORINFO_CALLINFO_ALLOWINSTPARAM indicates that this JIT is prepared to
7057 // supply the instantiation parameters necessary to make direct calls to underlying
7058 // shared generic code, rather than calling through instantiating stubs. If the
7059 // returned signature has CORINFO_CALLCONV_PARAMTYPE then this indicates that the JIT
7060 // must indeed pass an instantiation parameter.
7062 methHnd = callInfo->hMethod;
7064 sig = &(callInfo->sig);
7065 callRetTyp = JITtype2varType(sig->retType);
7067 mflags = callInfo->methodFlags;
7072 unsigned structSize = (callRetTyp == TYP_STRUCT) ? info.compCompHnd->getClassSize(sig->retTypeSigClass) : 0;
7073 printf("\nIn Compiler::impImportCall: opcode is %s, kind=%d, callRetType is %s, structSize is %d\n",
7074 opcodeNames[opcode], callInfo->kind, varTypeName(callRetTyp), structSize);
7077 if (compIsForInlining())
7079 /* Does this call site have security boundary restrictions? */
7081 if (impInlineInfo->inlineCandidateInfo->dwRestrictions & INLINE_RESPECT_BOUNDARY)
7083 compInlineResult->NoteFatal(InlineObservation::CALLSITE_CROSS_BOUNDARY_SECURITY);
7087 /* Does the inlinee need a security check token on the frame */
7089 if (mflags & CORINFO_FLG_SECURITYCHECK)
7091 compInlineResult->NoteFatal(InlineObservation::CALLEE_NEEDS_SECURITY_CHECK);
7095 /* Does the inlinee use StackCrawlMark */
7097 if (mflags & CORINFO_FLG_DONT_INLINE_CALLER)
7099 compInlineResult->NoteFatal(InlineObservation::CALLEE_STACK_CRAWL_MARK);
7103 /* For now ignore delegate invoke */
7105 if (mflags & CORINFO_FLG_DELEGATE_INVOKE)
7107 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_DELEGATE_INVOKE);
7111 /* For now ignore varargs */
7112 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_NATIVEVARARG)
7114 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_NATIVE_VARARGS);
7118 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG)
7120 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_MANAGED_VARARGS);
7124 if ((mflags & CORINFO_FLG_VIRTUAL) && (sig->sigInst.methInstCount != 0) && (opcode == CEE_CALLVIRT))
7126 compInlineResult->NoteFatal(InlineObservation::CALLEE_IS_GENERIC_VIRTUAL);
7131 clsHnd = pResolvedToken->hClass;
7133 clsFlags = callInfo->classFlags;
7136 // If this is a call to JitTestLabel.Mark, do "early inlining", and record the test attribute.
7138 // This recognition should really be done by knowing the methHnd of the relevant Mark method(s).
7139 // These should be in mscorlib.h, and available through a JIT/EE interface call.
7140 const char* modName;
7141 const char* className;
7142 const char* methodName;
7143 if ((className = eeGetClassName(clsHnd)) != nullptr &&
7144 strcmp(className, "System.Runtime.CompilerServices.JitTestLabel") == 0 &&
7145 (methodName = eeGetMethodName(methHnd, &modName)) != nullptr && strcmp(methodName, "Mark") == 0)
7147 return impImportJitTestLabelMark(sig->numArgs);
7151 // <NICE> Factor this into getCallInfo </NICE>
7152 bool isSpecialIntrinsic = false;
7153 if ((mflags & (CORINFO_FLG_INTRINSIC | CORINFO_FLG_JIT_INTRINSIC)) != 0)
7155 const bool isTail = canTailCall && (tailCall != 0);
7157 call = impIntrinsic(newobjThis, clsHnd, methHnd, sig, mflags, pResolvedToken->token, readonlyCall, isTail,
7158 pConstrainedResolvedToken, callInfo->thisTransform, &intrinsicID, &isSpecialIntrinsic);
7160 if (compDonotInline())
7165 if (call != nullptr)
7167 assert(!(mflags & CORINFO_FLG_VIRTUAL) || (mflags & CORINFO_FLG_FINAL) ||
7168 (clsFlags & CORINFO_FLG_FINAL));
7170 #ifdef FEATURE_READYTORUN_COMPILER
7171 if (call->OperGet() == GT_INTRINSIC)
7173 if (opts.IsReadyToRun())
7175 noway_assert(callInfo->kind == CORINFO_CALL);
7176 call->gtIntrinsic.gtEntryPoint = callInfo->codePointerLookup.constLookup;
7180 call->gtIntrinsic.gtEntryPoint.addr = nullptr;
7181 call->gtIntrinsic.gtEntryPoint.accessType = IAT_VALUE;
7186 bIntrinsicImported = true;
7194 call = impSIMDIntrinsic(opcode, newobjThis, clsHnd, methHnd, sig, pResolvedToken->token);
7195 if (call != nullptr)
7197 bIntrinsicImported = true;
7201 #endif // FEATURE_SIMD
7203 if ((mflags & CORINFO_FLG_VIRTUAL) && (mflags & CORINFO_FLG_EnC) && (opcode == CEE_CALLVIRT))
7205 NO_WAY("Virtual call to a function added via EnC is not supported");
7208 if ((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_DEFAULT &&
7209 (sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_VARARG &&
7210 (sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_NATIVEVARARG)
7212 BADCODE("Bad calling convention");
7215 //-------------------------------------------------------------------------
7216 // Construct the call node
7218 // Work out what sort of call we're making.
7219 // Dispense with virtual calls implemented via LDVIRTFTN immediately.
7221 constraintCallThisTransform = callInfo->thisTransform;
7222 exactContextHnd = callInfo->contextHandle;
7223 exactContextNeedsRuntimeLookup = callInfo->exactContextNeedsRuntimeLookup == TRUE;
7225 // Recursive call is treated as a loop to the begining of the method.
7226 if (gtIsRecursiveCall(methHnd))
7231 JITDUMP("\nFound recursive call in the method. Mark BB%02u to BB%02u as having a backward branch.\n",
7232 fgFirstBB->bbNum, compCurBB->bbNum);
7235 fgMarkBackwardJump(fgFirstBB, compCurBB);
7238 switch (callInfo->kind)
7241 case CORINFO_VIRTUALCALL_STUB:
7243 assert(!(mflags & CORINFO_FLG_STATIC)); // can't call a static method
7244 assert(!(clsFlags & CORINFO_FLG_VALUECLASS));
7245 if (callInfo->stubLookup.lookupKind.needsRuntimeLookup)
7248 if (compIsForInlining())
7250 // Don't import runtime lookups when inlining
7251 // Inlining has to be aborted in such a case
7252 /* XXX Fri 3/20/2009
7253 * By the way, this would never succeed. If the handle lookup is into the generic
7254 * dictionary for a candidate, you'll generate different dictionary offsets and the
7255 * inlined code will crash.
7257 * To anyone code reviewing this, when could this ever succeed in the future? It'll
7258 * always have a handle lookup. These lookups are safe intra-module, but we're just
7261 compInlineResult->NoteFatal(InlineObservation::CALLSITE_HAS_COMPLEX_HANDLE);
7265 GenTree* stubAddr = impRuntimeLookupToTree(pResolvedToken, &callInfo->stubLookup, methHnd);
7266 assert(!compDonotInline());
7268 // This is the rough code to set up an indirect stub call
7269 assert(stubAddr != nullptr);
7271 // The stubAddr may be a
7272 // complex expression. As it is evaluated after the args,
7273 // it may cause registered args to be spilled. Simply spill it.
7275 unsigned lclNum = lvaGrabTemp(true DEBUGARG("VirtualCall with runtime lookup"));
7276 impAssignTempGen(lclNum, stubAddr, (unsigned)CHECK_SPILL_ALL);
7277 stubAddr = gtNewLclvNode(lclNum, TYP_I_IMPL);
7279 // Create the actual call node
7281 assert((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_VARARG &&
7282 (sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_NATIVEVARARG);
7284 call = gtNewIndCallNode(stubAddr, callRetTyp, nullptr);
7286 call->gtFlags |= GTF_EXCEPT | (stubAddr->gtFlags & GTF_GLOB_EFFECT);
7287 call->gtFlags |= GTF_CALL_VIRT_STUB;
7290 // No tailcalls allowed for these yet...
7291 canTailCall = false;
7292 szCanTailCallFailReason = "VirtualCall with runtime lookup";
7297 // ok, the stub is available at compile type.
7299 call = gtNewCallNode(CT_USER_FUNC, callInfo->hMethod, callRetTyp, nullptr, ilOffset);
7300 call->gtCall.gtStubCallStubAddr = callInfo->stubLookup.constLookup.addr;
7301 call->gtFlags |= GTF_CALL_VIRT_STUB;
7302 assert(callInfo->stubLookup.constLookup.accessType != IAT_PPVALUE);
7303 if (callInfo->stubLookup.constLookup.accessType == IAT_PVALUE)
7305 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_VIRTSTUB_REL_INDIRECT;
7309 #ifdef FEATURE_READYTORUN_COMPILER
7310 if (opts.IsReadyToRun())
7312 // Null check is sometimes needed for ready to run to handle
7313 // non-virtual <-> virtual changes between versions
7314 if (callInfo->nullInstanceCheck)
7316 call->gtFlags |= GTF_CALL_NULLCHECK;
7324 case CORINFO_VIRTUALCALL_VTABLE:
7326 assert(!(mflags & CORINFO_FLG_STATIC)); // can't call a static method
7327 assert(!(clsFlags & CORINFO_FLG_VALUECLASS));
7328 call = gtNewCallNode(CT_USER_FUNC, callInfo->hMethod, callRetTyp, nullptr, ilOffset);
7329 call->gtFlags |= GTF_CALL_VIRT_VTABLE;
7333 case CORINFO_VIRTUALCALL_LDVIRTFTN:
7335 if (compIsForInlining())
7337 compInlineResult->NoteFatal(InlineObservation::CALLSITE_HAS_CALL_VIA_LDVIRTFTN);
7341 assert(!(mflags & CORINFO_FLG_STATIC)); // can't call a static method
7342 assert(!(clsFlags & CORINFO_FLG_VALUECLASS));
7343 // OK, We've been told to call via LDVIRTFTN, so just
7344 // take the call now....
7346 args = impPopList(sig->numArgs, sig);
7348 GenTree* thisPtr = impPopStack().val;
7349 thisPtr = impTransformThis(thisPtr, pConstrainedResolvedToken, callInfo->thisTransform);
7350 assert(thisPtr != nullptr);
7352 // Clone the (possibly transformed) "this" pointer
7353 GenTree* thisPtrCopy;
7354 thisPtr = impCloneExpr(thisPtr, &thisPtrCopy, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
7355 nullptr DEBUGARG("LDVIRTFTN this pointer"));
7357 GenTree* fptr = impImportLdvirtftn(thisPtr, pResolvedToken, callInfo);
7358 assert(fptr != nullptr);
7360 thisPtr = nullptr; // can't reuse it
7362 // Now make an indirect call through the function pointer
7364 unsigned lclNum = lvaGrabTemp(true DEBUGARG("VirtualCall through function pointer"));
7365 impAssignTempGen(lclNum, fptr, (unsigned)CHECK_SPILL_ALL);
7366 fptr = gtNewLclvNode(lclNum, TYP_I_IMPL);
7368 // Create the actual call node
7370 call = gtNewIndCallNode(fptr, callRetTyp, args, ilOffset);
7371 call->gtCall.gtCallObjp = thisPtrCopy;
7372 call->gtFlags |= GTF_EXCEPT | (fptr->gtFlags & GTF_GLOB_EFFECT);
7374 if ((sig->sigInst.methInstCount != 0) && IsTargetAbi(CORINFO_CORERT_ABI))
7376 // CoreRT generic virtual method: need to handle potential fat function pointers
7377 addFatPointerCandidate(call->AsCall());
7379 #ifdef FEATURE_READYTORUN_COMPILER
7380 if (opts.IsReadyToRun())
7382 // Null check is needed for ready to run to handle
7383 // non-virtual <-> virtual changes between versions
7384 call->gtFlags |= GTF_CALL_NULLCHECK;
7388 // Sine we are jumping over some code, check that its OK to skip that code
7389 assert((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_VARARG &&
7390 (sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_NATIVEVARARG);
7396 // This is for a non-virtual, non-interface etc. call
7397 call = gtNewCallNode(CT_USER_FUNC, callInfo->hMethod, callRetTyp, nullptr, ilOffset);
7399 // We remove the nullcheck for the GetType call instrinsic.
7400 // TODO-CQ: JIT64 does not introduce the null check for many more helper calls
7402 if (callInfo->nullInstanceCheck &&
7403 !((mflags & CORINFO_FLG_INTRINSIC) != 0 && (intrinsicID == CORINFO_INTRINSIC_Object_GetType)))
7405 call->gtFlags |= GTF_CALL_NULLCHECK;
7408 #ifdef FEATURE_READYTORUN_COMPILER
7409 if (opts.IsReadyToRun())
7411 call->gtCall.setEntryPoint(callInfo->codePointerLookup.constLookup);
7417 case CORINFO_CALL_CODE_POINTER:
7419 // The EE has asked us to call by computing a code pointer and then doing an
7420 // indirect call. This is because a runtime lookup is required to get the code entry point.
7422 // These calls always follow a uniform calling convention, i.e. no extra hidden params
7423 assert((sig->callConv & CORINFO_CALLCONV_PARAMTYPE) == 0);
7425 assert((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_VARARG);
7426 assert((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_NATIVEVARARG);
7429 impLookupToTree(pResolvedToken, &callInfo->codePointerLookup, GTF_ICON_FTN_ADDR, callInfo->hMethod);
7431 if (compDonotInline())
7436 // Now make an indirect call through the function pointer
7438 unsigned lclNum = lvaGrabTemp(true DEBUGARG("Indirect call through function pointer"));
7439 impAssignTempGen(lclNum, fptr, (unsigned)CHECK_SPILL_ALL);
7440 fptr = gtNewLclvNode(lclNum, TYP_I_IMPL);
7442 call = gtNewIndCallNode(fptr, callRetTyp, nullptr, ilOffset);
7443 call->gtFlags |= GTF_EXCEPT | (fptr->gtFlags & GTF_GLOB_EFFECT);
7444 if (callInfo->nullInstanceCheck)
7446 call->gtFlags |= GTF_CALL_NULLCHECK;
7453 assert(!"unknown call kind");
7457 //-------------------------------------------------------------------------
7460 PREFIX_ASSUME(call != nullptr);
7462 if (mflags & CORINFO_FLG_NOGCCHECK)
7464 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_NOGCCHECK;
7467 // Mark call if it's one of the ones we will maybe treat as an intrinsic
7468 if (isSpecialIntrinsic)
7470 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_SPECIAL_INTRINSIC;
7474 assert(clsHnd || (opcode == CEE_CALLI)); // We're never verifying for CALLI, so this is not set.
7476 /* Some sanity checks */
7478 // CALL_VIRT and NEWOBJ must have a THIS pointer
7479 assert((opcode != CEE_CALLVIRT && opcode != CEE_NEWOBJ) || (sig->callConv & CORINFO_CALLCONV_HASTHIS));
7480 // static bit and hasThis are negations of one another
7481 assert(((mflags & CORINFO_FLG_STATIC) != 0) == ((sig->callConv & CORINFO_CALLCONV_HASTHIS) == 0));
7482 assert(call != nullptr);
7484 /*-------------------------------------------------------------------------
7485 * Check special-cases etc
7488 /* Special case - Check if it is a call to Delegate.Invoke(). */
7490 if (mflags & CORINFO_FLG_DELEGATE_INVOKE)
7492 assert(!compIsForInlining());
7493 assert(!(mflags & CORINFO_FLG_STATIC)); // can't call a static method
7494 assert(mflags & CORINFO_FLG_FINAL);
7496 /* Set the delegate flag */
7497 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_DELEGATE_INV;
7499 if (callInfo->secureDelegateInvoke)
7501 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_SECURE_DELEGATE_INV;
7504 if (opcode == CEE_CALLVIRT)
7506 assert(mflags & CORINFO_FLG_FINAL);
7508 /* It should have the GTF_CALL_NULLCHECK flag set. Reset it */
7509 assert(call->gtFlags & GTF_CALL_NULLCHECK);
7510 call->gtFlags &= ~GTF_CALL_NULLCHECK;
7514 CORINFO_CLASS_HANDLE actualMethodRetTypeSigClass;
7515 actualMethodRetTypeSigClass = sig->retTypeSigClass;
7516 if (varTypeIsStruct(callRetTyp))
7518 callRetTyp = impNormStructType(actualMethodRetTypeSigClass);
7519 call->gtType = callRetTyp;
7523 /* Check for varargs */
7524 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG ||
7525 (sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_NATIVEVARARG)
7527 BADCODE("Varargs not supported.");
7529 #endif // !FEATURE_VARARG
7532 if (call->gtCall.callSig == nullptr)
7534 call->gtCall.callSig = new (this, CMK_CorSig) CORINFO_SIG_INFO;
7535 *call->gtCall.callSig = *sig;
7537 #endif // UNIX_X86_ABI
7539 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG ||
7540 (sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_NATIVEVARARG)
7542 assert(!compIsForInlining());
7544 /* Set the right flags */
7546 call->gtFlags |= GTF_CALL_POP_ARGS;
7547 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_VARARGS;
7549 /* Can't allow tailcall for varargs as it is caller-pop. The caller
7550 will be expecting to pop a certain number of arguments, but if we
7551 tailcall to a function with a different number of arguments, we
7552 are hosed. There are ways around this (caller remembers esp value,
7553 varargs is not caller-pop, etc), but not worth it. */
7554 CLANG_FORMAT_COMMENT_ANCHOR;
7559 canTailCall = false;
7560 szCanTailCallFailReason = "Callee is varargs";
7564 /* Get the total number of arguments - this is already correct
7565 * for CALLI - for methods we have to get it from the call site */
7567 if (opcode != CEE_CALLI)
7570 unsigned numArgsDef = sig->numArgs;
7572 eeGetCallSiteSig(pResolvedToken->token, info.compScopeHnd, impTokenLookupContextHandle, sig);
7575 // We cannot lazily obtain the signature of a vararg call because using its method
7576 // handle will give us only the declared argument list, not the full argument list.
7577 assert(call->gtCall.callSig == nullptr);
7578 call->gtCall.callSig = new (this, CMK_CorSig) CORINFO_SIG_INFO;
7579 *call->gtCall.callSig = *sig;
7582 // For vararg calls we must be sure to load the return type of the
7583 // method actually being called, as well as the return types of the
7584 // specified in the vararg signature. With type equivalency, these types
7585 // may not be the same.
7586 if (sig->retTypeSigClass != actualMethodRetTypeSigClass)
7588 if (actualMethodRetTypeSigClass != nullptr && sig->retType != CORINFO_TYPE_CLASS &&
7589 sig->retType != CORINFO_TYPE_BYREF && sig->retType != CORINFO_TYPE_PTR &&
7590 sig->retType != CORINFO_TYPE_VAR)
7592 // Make sure that all valuetypes (including enums) that we push are loaded.
7593 // This is to guarantee that if a GC is triggerred from the prestub of this methods,
7594 // all valuetypes in the method signature are already loaded.
7595 // We need to be able to find the size of the valuetypes, but we cannot
7596 // do a class-load from within GC.
7597 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(actualMethodRetTypeSigClass);
7601 assert(numArgsDef <= sig->numArgs);
7604 /* We will have "cookie" as the last argument but we cannot push
7605 * it on the operand stack because we may overflow, so we append it
7606 * to the arg list next after we pop them */
7609 if (mflags & CORINFO_FLG_SECURITYCHECK)
7611 assert(!compIsForInlining());
7613 // Need security prolog/epilog callouts when there is
7614 // imperative security in the method. This is to give security a
7615 // chance to do any setup in the prolog and cleanup in the epilog if needed.
7617 if (compIsForInlining())
7619 // Cannot handle this if the method being imported is an inlinee by itself.
7620 // Because inlinee method does not have its own frame.
7622 compInlineResult->NoteFatal(InlineObservation::CALLEE_NEEDS_SECURITY_CHECK);
7627 tiSecurityCalloutNeeded = true;
7629 // If the current method calls a method which needs a security check,
7630 // (i.e. the method being compiled has imperative security)
7631 // we need to reserve a slot for the security object in
7632 // the current method's stack frame
7633 opts.compNeedSecurityCheck = true;
7637 //--------------------------- Inline NDirect ------------------------------
7639 // For inline cases we technically should look at both the current
7640 // block and the call site block (or just the latter if we've
7641 // fused the EH trees). However the block-related checks pertain to
7642 // EH and we currently won't inline a method with EH. So for
7643 // inlinees, just checking the call site block is sufficient.
7645 // New lexical block here to avoid compilation errors because of GOTOs.
7646 BasicBlock* block = compIsForInlining() ? impInlineInfo->iciBlock : compCurBB;
7647 impCheckForPInvokeCall(call->AsCall(), methHnd, sig, mflags, block);
7650 if (call->gtFlags & GTF_CALL_UNMANAGED)
7652 // We set up the unmanaged call by linking the frame, disabling GC, etc
7653 // This needs to be cleaned up on return
7656 canTailCall = false;
7657 szCanTailCallFailReason = "Callee is native";
7660 checkForSmallType = true;
7662 impPopArgsForUnmanagedCall(call, sig);
7666 else if ((opcode == CEE_CALLI) && (((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_STDCALL) ||
7667 ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_C) ||
7668 ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_THISCALL) ||
7669 ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_FASTCALL)))
7671 if (!info.compCompHnd->canGetCookieForPInvokeCalliSig(sig))
7673 // Normally this only happens with inlining.
7674 // However, a generic method (or type) being NGENd into another module
7675 // can run into this issue as well. There's not an easy fall-back for NGEN
7676 // so instead we fallback to JIT.
7677 if (compIsForInlining())
7679 compInlineResult->NoteFatal(InlineObservation::CALLSITE_CANT_EMBED_PINVOKE_COOKIE);
7683 IMPL_LIMITATION("Can't get PInvoke cookie (cross module generics)");
7689 GenTree* cookie = eeGetPInvokeCookie(sig);
7691 // This cookie is required to be either a simple GT_CNS_INT or
7692 // an indirection of a GT_CNS_INT
7694 GenTree* cookieConst = cookie;
7695 if (cookie->gtOper == GT_IND)
7697 cookieConst = cookie->gtOp.gtOp1;
7699 assert(cookieConst->gtOper == GT_CNS_INT);
7701 // Setting GTF_DONT_CSE on the GT_CNS_INT as well as on the GT_IND (if it exists) will ensure that
7702 // we won't allow this tree to participate in any CSE logic
7704 cookie->gtFlags |= GTF_DONT_CSE;
7705 cookieConst->gtFlags |= GTF_DONT_CSE;
7707 call->gtCall.gtCallCookie = cookie;
7711 canTailCall = false;
7712 szCanTailCallFailReason = "PInvoke calli";
7716 /*-------------------------------------------------------------------------
7717 * Create the argument list
7720 //-------------------------------------------------------------------------
7721 // Special case - for varargs we have an implicit last argument
7723 if ((sig->callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG)
7725 assert(!compIsForInlining());
7727 void *varCookie, *pVarCookie;
7728 if (!info.compCompHnd->canGetVarArgsHandle(sig))
7730 compInlineResult->NoteFatal(InlineObservation::CALLSITE_CANT_EMBED_VARARGS_COOKIE);
7734 varCookie = info.compCompHnd->getVarArgsHandle(sig, &pVarCookie);
7735 assert((!varCookie) != (!pVarCookie));
7736 GenTree* cookie = gtNewIconEmbHndNode(varCookie, pVarCookie, GTF_ICON_VARG_HDL, sig);
7738 assert(extraArg == nullptr);
7739 extraArg = gtNewArgList(cookie);
7742 //-------------------------------------------------------------------------
7743 // Extra arg for shared generic code and array methods
7745 // Extra argument containing instantiation information is passed in the
7746 // following circumstances:
7747 // (a) To the "Address" method on array classes; the extra parameter is
7748 // the array's type handle (a TypeDesc)
7749 // (b) To shared-code instance methods in generic structs; the extra parameter
7750 // is the struct's type handle (a vtable ptr)
7751 // (c) To shared-code per-instantiation non-generic static methods in generic
7752 // classes and structs; the extra parameter is the type handle
7753 // (d) To shared-code generic methods; the extra parameter is an
7754 // exact-instantiation MethodDesc
7756 // We also set the exact type context associated with the call so we can
7757 // inline the call correctly later on.
7759 if (sig->callConv & CORINFO_CALLCONV_PARAMTYPE)
7761 assert(call->gtCall.gtCallType == CT_USER_FUNC);
7762 if (clsHnd == nullptr)
7764 NO_WAY("CALLI on parameterized type");
7767 assert(opcode != CEE_CALLI);
7772 // Instantiated generic method
7773 if (((SIZE_T)exactContextHnd & CORINFO_CONTEXTFLAGS_MASK) == CORINFO_CONTEXTFLAGS_METHOD)
7775 CORINFO_METHOD_HANDLE exactMethodHandle =
7776 (CORINFO_METHOD_HANDLE)((SIZE_T)exactContextHnd & ~CORINFO_CONTEXTFLAGS_MASK);
7778 if (!exactContextNeedsRuntimeLookup)
7780 #ifdef FEATURE_READYTORUN_COMPILER
7781 if (opts.IsReadyToRun())
7784 impReadyToRunLookupToTree(&callInfo->instParamLookup, GTF_ICON_METHOD_HDL, exactMethodHandle);
7785 if (instParam == nullptr)
7787 assert(compDonotInline());
7794 instParam = gtNewIconEmbMethHndNode(exactMethodHandle);
7795 info.compCompHnd->methodMustBeLoadedBeforeCodeIsRun(exactMethodHandle);
7800 instParam = impTokenToHandle(pResolvedToken, &runtimeLookup, TRUE /*mustRestoreHandle*/);
7801 if (instParam == nullptr)
7803 assert(compDonotInline());
7809 // otherwise must be an instance method in a generic struct,
7810 // a static method in a generic type, or a runtime-generated array method
7813 assert(((SIZE_T)exactContextHnd & CORINFO_CONTEXTFLAGS_MASK) == CORINFO_CONTEXTFLAGS_CLASS);
7814 CORINFO_CLASS_HANDLE exactClassHandle =
7815 (CORINFO_CLASS_HANDLE)((SIZE_T)exactContextHnd & ~CORINFO_CONTEXTFLAGS_MASK);
7817 if (compIsForInlining() && (clsFlags & CORINFO_FLG_ARRAY) != 0)
7819 compInlineResult->NoteFatal(InlineObservation::CALLEE_IS_ARRAY_METHOD);
7823 if ((clsFlags & CORINFO_FLG_ARRAY) && readonlyCall)
7825 // We indicate "readonly" to the Address operation by using a null
7827 instParam = gtNewIconNode(0, TYP_REF);
7829 else if (!exactContextNeedsRuntimeLookup)
7831 #ifdef FEATURE_READYTORUN_COMPILER
7832 if (opts.IsReadyToRun())
7835 impReadyToRunLookupToTree(&callInfo->instParamLookup, GTF_ICON_CLASS_HDL, exactClassHandle);
7836 if (instParam == nullptr)
7838 assert(compDonotInline());
7845 instParam = gtNewIconEmbClsHndNode(exactClassHandle);
7846 info.compCompHnd->classMustBeLoadedBeforeCodeIsRun(exactClassHandle);
7851 // If the EE was able to resolve a constrained call, the instantiating parameter to use is the type
7852 // by which the call was constrained with. We embed pConstrainedResolvedToken as the extra argument
7853 // because pResolvedToken is an interface method and interface types make a poor generic context.
7854 if (pConstrainedResolvedToken)
7856 instParam = impTokenToHandle(pConstrainedResolvedToken, &runtimeLookup, TRUE /*mustRestoreHandle*/,
7857 FALSE /* importParent */);
7861 instParam = impParentClassTokenToHandle(pResolvedToken, &runtimeLookup, TRUE /*mustRestoreHandle*/);
7864 if (instParam == nullptr)
7866 assert(compDonotInline());
7872 assert(extraArg == nullptr);
7873 extraArg = gtNewArgList(instParam);
7876 // Inlining may need the exact type context (exactContextHnd) if we're inlining shared generic code, in particular
7877 // to inline 'polytypic' operations such as static field accesses, type tests and method calls which
7878 // rely on the exact context. The exactContextHnd is passed back to the JitInterface at appropriate points.
7879 // exactContextHnd is not currently required when inlining shared generic code into shared
7880 // generic code, since the inliner aborts whenever shared code polytypic operations are encountered
7881 // (e.g. anything marked needsRuntimeLookup)
7882 if (exactContextNeedsRuntimeLookup)
7884 exactContextHnd = nullptr;
7887 if ((opcode == CEE_NEWOBJ) && ((clsFlags & CORINFO_FLG_DELEGATE) != 0))
7889 // Only verifiable cases are supported.
7890 // dup; ldvirtftn; newobj; or ldftn; newobj.
7891 // IL test could contain unverifiable sequence, in this case optimization should not be done.
7892 if (impStackHeight() > 0)
7894 typeInfo delegateTypeInfo = impStackTop().seTypeInfo;
7895 if (delegateTypeInfo.IsToken())
7897 ldftnToken = delegateTypeInfo.GetToken();
7902 //-------------------------------------------------------------------------
7903 // The main group of arguments
7905 args = call->gtCall.gtCallArgs = impPopList(sig->numArgs, sig, extraArg);
7909 call->gtFlags |= args->gtFlags & GTF_GLOB_EFFECT;
7912 //-------------------------------------------------------------------------
7913 // The "this" pointer
7915 if (!(mflags & CORINFO_FLG_STATIC) && !((opcode == CEE_NEWOBJ) && (newobjThis == nullptr)))
7919 if (opcode == CEE_NEWOBJ)
7925 obj = impPopStack().val;
7926 obj = impTransformThis(obj, pConstrainedResolvedToken, constraintCallThisTransform);
7927 if (compDonotInline())
7933 // Store the "this" value in the call
7934 call->gtFlags |= obj->gtFlags & GTF_GLOB_EFFECT;
7935 call->gtCall.gtCallObjp = obj;
7937 // Is this a virtual or interface call?
7938 if (call->gtCall.IsVirtual())
7940 // only true object pointers can be virtual
7941 assert(obj->gtType == TYP_REF);
7943 // See if we can devirtualize.
7944 impDevirtualizeCall(call->AsCall(), &callInfo->hMethod, &callInfo->methodFlags, &callInfo->contextHandle,
7950 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_NONVIRT_SAME_THIS;
7954 //-------------------------------------------------------------------------
7955 // The "this" pointer for "newobj"
7957 if (opcode == CEE_NEWOBJ)
7959 if (clsFlags & CORINFO_FLG_VAROBJSIZE)
7961 assert(!(clsFlags & CORINFO_FLG_ARRAY)); // arrays handled separately
7962 // This is a 'new' of a variable sized object, wher
7963 // the constructor is to return the object. In this case
7964 // the constructor claims to return VOID but we know it
7965 // actually returns the new object
7966 assert(callRetTyp == TYP_VOID);
7967 callRetTyp = TYP_REF;
7968 call->gtType = TYP_REF;
7969 impSpillSpecialSideEff();
7971 impPushOnStack(call, typeInfo(TI_REF, clsHnd));
7975 if (clsFlags & CORINFO_FLG_DELEGATE)
7977 // New inliner morph it in impImportCall.
7978 // This will allow us to inline the call to the delegate constructor.
7979 call = fgOptimizeDelegateConstructor(call->AsCall(), &exactContextHnd, ldftnToken);
7982 if (!bIntrinsicImported)
7985 #if defined(DEBUG) || defined(INLINE_DATA)
7987 // Keep track of the raw IL offset of the call
7988 call->gtCall.gtRawILOffset = rawILOffset;
7990 #endif // defined(DEBUG) || defined(INLINE_DATA)
7992 // Is it an inline candidate?
7993 impMarkInlineCandidate(call, exactContextHnd, exactContextNeedsRuntimeLookup, callInfo);
7996 // append the call node.
7997 impAppendTree(call, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
7999 // Now push the value of the 'new onto the stack
8001 // This is a 'new' of a non-variable sized object.
8002 // Append the new node (op1) to the statement list,
8003 // and then push the local holding the value of this
8004 // new instruction on the stack.
8006 if (clsFlags & CORINFO_FLG_VALUECLASS)
8008 assert(newobjThis->gtOper == GT_ADDR && newobjThis->gtOp.gtOp1->gtOper == GT_LCL_VAR);
8010 unsigned tmp = newobjThis->gtOp.gtOp1->gtLclVarCommon.gtLclNum;
8011 impPushOnStack(gtNewLclvNode(tmp, lvaGetRealType(tmp)), verMakeTypeInfo(clsHnd).NormaliseForStack());
8015 if (newobjThis->gtOper == GT_COMMA)
8017 // In coreclr the callout can be inserted even if verification is disabled
8018 // so we cannot rely on tiVerificationNeeded alone
8020 // We must have inserted the callout. Get the real newobj.
8021 newobjThis = newobjThis->gtOp.gtOp2;
8024 assert(newobjThis->gtOper == GT_LCL_VAR);
8025 impPushOnStack(gtNewLclvNode(newobjThis->gtLclVarCommon.gtLclNum, TYP_REF), typeInfo(TI_REF, clsHnd));
8035 // This check cannot be performed for implicit tail calls for the reason
8036 // that impIsImplicitTailCallCandidate() is not checking whether return
8037 // types are compatible before marking a call node with PREFIX_TAILCALL_IMPLICIT.
8038 // As a result it is possible that in the following case, we find that
8039 // the type stack is non-empty if Callee() is considered for implicit
8041 // int Caller(..) { .... void Callee(); ret val; ... }
8043 // Note that we cannot check return type compatibility before ImpImportCall()
8044 // as we don't have required info or need to duplicate some of the logic of
8047 // For implicit tail calls, we perform this check after return types are
8048 // known to be compatible.
8049 if ((tailCall & PREFIX_TAILCALL_EXPLICIT) && (verCurrentState.esStackDepth != 0))
8051 BADCODE("Stack should be empty after tailcall");
8054 // Note that we can not relax this condition with genActualType() as
8055 // the calling convention dictates that the caller of a function with
8056 // a small-typed return value is responsible for normalizing the return val
8059 !impTailCallRetTypeCompatible(info.compRetType, info.compMethodInfo->args.retTypeClass, callRetTyp,
8060 callInfo->sig.retTypeClass))
8062 canTailCall = false;
8063 szCanTailCallFailReason = "Return types are not tail call compatible";
8066 // Stack empty check for implicit tail calls.
8067 if (canTailCall && (tailCall & PREFIX_TAILCALL_IMPLICIT) && (verCurrentState.esStackDepth != 0))
8069 #ifdef _TARGET_AMD64_
8070 // JIT64 Compatibility: Opportunistic tail call stack mismatch throws a VerificationException
8071 // in JIT64, not an InvalidProgramException.
8072 Verify(false, "Stack should be empty after tailcall");
8073 #else // _TARGET_64BIT_
8074 BADCODE("Stack should be empty after tailcall");
8075 #endif //!_TARGET_64BIT_
8078 // assert(compCurBB is not a catch, finally or filter block);
8079 // assert(compCurBB is not a try block protected by a finally block);
8081 // Check for permission to tailcall
8082 bool explicitTailCall = (tailCall & PREFIX_TAILCALL_EXPLICIT) != 0;
8084 assert(!explicitTailCall || compCurBB->bbJumpKind == BBJ_RETURN);
8088 // True virtual or indirect calls, shouldn't pass in a callee handle.
8089 CORINFO_METHOD_HANDLE exactCalleeHnd =
8090 ((call->gtCall.gtCallType != CT_USER_FUNC) || call->gtCall.IsVirtual()) ? nullptr : methHnd;
8091 GenTree* thisArg = call->gtCall.gtCallObjp;
8093 if (info.compCompHnd->canTailCall(info.compMethodHnd, methHnd, exactCalleeHnd, explicitTailCall))
8096 if (explicitTailCall)
8098 // In case of explicit tail calls, mark it so that it is not considered
8100 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_EXPLICIT_TAILCALL;
8104 printf("\nGTF_CALL_M_EXPLICIT_TAILCALL bit set for call ");
8112 #if FEATURE_TAILCALL_OPT
8113 // Must be an implicit tail call.
8114 assert((tailCall & PREFIX_TAILCALL_IMPLICIT) != 0);
8116 // It is possible that a call node is both an inline candidate and marked
8117 // for opportunistic tail calling. In-lining happens before morhphing of
8118 // trees. If in-lining of an in-line candidate gets aborted for whatever
8119 // reason, it will survive to the morphing stage at which point it will be
8120 // transformed into a tail call after performing additional checks.
8122 call->gtCall.gtCallMoreFlags |= GTF_CALL_M_IMPLICIT_TAILCALL;
8126 printf("\nGTF_CALL_M_IMPLICIT_TAILCALL bit set for call ");
8132 #else //! FEATURE_TAILCALL_OPT
8133 NYI("Implicit tail call prefix on a target which doesn't support opportunistic tail calls");
8135 #endif // FEATURE_TAILCALL_OPT
8138 // we can't report success just yet...
8142 canTailCall = false;
8143 // canTailCall reported its reasons already
8147 printf("\ninfo.compCompHnd->canTailCall returned false for call ");
8156 // If this assert fires it means that canTailCall was set to false without setting a reason!
8157 assert(szCanTailCallFailReason != nullptr);
8162 printf("\nRejecting %splicit tail call for call ", explicitTailCall ? "ex" : "im");
8164 printf(": %s\n", szCanTailCallFailReason);
8167 info.compCompHnd->reportTailCallDecision(info.compMethodHnd, methHnd, explicitTailCall, TAILCALL_FAIL,
8168 szCanTailCallFailReason);
8172 // Note: we assume that small return types are already normalized by the managed callee
8173 // or by the pinvoke stub for calls to unmanaged code.
8175 if (!bIntrinsicImported)
8178 // Things needed to be checked when bIntrinsicImported is false.
8181 assert(call->gtOper == GT_CALL);
8182 assert(sig != nullptr);
8184 // Tail calls require us to save the call site's sig info so we can obtain an argument
8185 // copying thunk from the EE later on.
8186 if (call->gtCall.callSig == nullptr)
8188 call->gtCall.callSig = new (this, CMK_CorSig) CORINFO_SIG_INFO;
8189 *call->gtCall.callSig = *sig;
8192 if (compIsForInlining() && opcode == CEE_CALLVIRT)
8194 GenTree* callObj = call->gtCall.gtCallObjp;
8195 assert(callObj != nullptr);
8197 if ((call->gtCall.IsVirtual() || (call->gtFlags & GTF_CALL_NULLCHECK)) &&
8198 impInlineIsGuaranteedThisDerefBeforeAnySideEffects(call->gtCall.gtCallArgs, callObj,
8199 impInlineInfo->inlArgInfo))
8201 impInlineInfo->thisDereferencedFirst = true;
8205 #if defined(DEBUG) || defined(INLINE_DATA)
8207 // Keep track of the raw IL offset of the call
8208 call->gtCall.gtRawILOffset = rawILOffset;
8210 #endif // defined(DEBUG) || defined(INLINE_DATA)
8212 // Is it an inline candidate?
8213 impMarkInlineCandidate(call, exactContextHnd, exactContextNeedsRuntimeLookup, callInfo);
8217 // Push or append the result of the call
8218 if (callRetTyp == TYP_VOID)
8220 if (opcode == CEE_NEWOBJ)
8222 // we actually did push something, so don't spill the thing we just pushed.
8223 assert(verCurrentState.esStackDepth > 0);
8224 impAppendTree(call, verCurrentState.esStackDepth - 1, impCurStmtOffs);
8228 impAppendTree(call, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
8233 impSpillSpecialSideEff();
8235 if (clsFlags & CORINFO_FLG_ARRAY)
8237 eeGetCallSiteSig(pResolvedToken->token, pResolvedToken->tokenScope, pResolvedToken->tokenContext, sig);
8240 // Find the return type used for verification by interpreting the method signature.
8241 // NB: we are clobbering the already established sig.
8242 if (tiVerificationNeeded)
8244 // Actually, we never get the sig for the original method.
8245 sig = &(callInfo->verSig);
8248 typeInfo tiRetVal = verMakeTypeInfo(sig->retType, sig->retTypeClass);
8249 tiRetVal.NormaliseForStack();
8251 // The CEE_READONLY prefix modifies the verification semantics of an Address
8252 // operation on an array type.
8253 if ((clsFlags & CORINFO_FLG_ARRAY) && readonlyCall && tiRetVal.IsByRef())
8255 tiRetVal.SetIsReadonlyByRef();
8258 if (tiVerificationNeeded)
8260 // We assume all calls return permanent home byrefs. If they
8261 // didn't they wouldn't be verifiable. This is also covering
8262 // the Address() helper for multidimensional arrays.
8263 if (tiRetVal.IsByRef())
8265 tiRetVal.SetIsPermanentHomeByRef();
8271 // Sometimes "call" is not a GT_CALL (if we imported an intrinsic that didn't turn into a call)
8273 bool fatPointerCandidate = call->AsCall()->IsFatPointerCandidate();
8274 if (varTypeIsStruct(callRetTyp))
8276 call = impFixupCallStructReturn(call->AsCall(), sig->retTypeClass);
8279 if ((call->gtFlags & GTF_CALL_INLINE_CANDIDATE) != 0)
8281 assert(opts.OptEnabled(CLFLG_INLINING));
8282 assert(!fatPointerCandidate); // We should not try to inline calli.
8284 // Make the call its own tree (spill the stack if needed).
8285 impAppendTree(call, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
8287 // TODO: Still using the widened type.
8288 call = gtNewInlineCandidateReturnExpr(call, genActualType(callRetTyp));
8292 if (fatPointerCandidate)
8294 // fatPointer candidates should be in statements of the form call() or var = call().
8295 // Such form allows to find statements with fat calls without walking through whole trees
8296 // and removes problems with cutting trees.
8297 assert(!bIntrinsicImported);
8298 assert(IsTargetAbi(CORINFO_CORERT_ABI));
8299 if (call->OperGet() != GT_LCL_VAR) // can be already converted by impFixupCallStructReturn.
8301 unsigned calliSlot = lvaGrabTemp(true DEBUGARG("calli"));
8302 LclVarDsc* varDsc = &lvaTable[calliSlot];
8303 varDsc->lvVerTypeInfo = tiRetVal;
8304 impAssignTempGen(calliSlot, call, tiRetVal.GetClassHandle(), (unsigned)CHECK_SPILL_NONE);
8305 // impAssignTempGen can change src arg list and return type for call that returns struct.
8306 var_types type = genActualType(lvaTable[calliSlot].TypeGet());
8307 call = gtNewLclvNode(calliSlot, type);
8311 // For non-candidates we must also spill, since we
8312 // might have locals live on the eval stack that this
8315 // Suppress this for certain well-known call targets
8316 // that we know won't modify locals, eg calls that are
8317 // recognized in gtCanOptimizeTypeEquality. Otherwise
8318 // we may break key fragile pattern matches later on.
8319 bool spillStack = true;
8322 GenTreeCall* callNode = call->AsCall();
8323 if ((callNode->gtCallType == CT_HELPER) && gtIsTypeHandleToRuntimeTypeHelper(callNode))
8327 else if ((callNode->gtCallMoreFlags & GTF_CALL_M_SPECIAL_INTRINSIC) != 0)
8335 impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("non-inline candidate call"));
8340 if (!bIntrinsicImported)
8342 //-------------------------------------------------------------------------
8344 /* If the call is of a small type and the callee is managed, the callee will normalize the result
8346 However, we need to normalize small type values returned by unmanaged
8347 functions (pinvoke). The pinvoke stub does the normalization, but we need to do it here
8348 if we use the shorter inlined pinvoke stub. */
8350 if (checkForSmallType && varTypeIsIntegral(callRetTyp) && genTypeSize(callRetTyp) < genTypeSize(TYP_INT))
8352 call = gtNewCastNode(genActualType(callRetTyp), call, false, callRetTyp);
8356 impPushOnStack(call, tiRetVal);
8359 // VSD functions get a new call target each time we getCallInfo, so clear the cache.
8360 // Also, the call info cache for CALLI instructions is largely incomplete, so clear it out.
8361 // if ( (opcode == CEE_CALLI) || (callInfoCache.fetchCallInfo().kind == CORINFO_VIRTUALCALL_STUB))
8362 // callInfoCache.uncacheCallInfo();
8367 #pragma warning(pop)
8370 bool Compiler::impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo)
8372 CorInfoType corType = methInfo->args.retType;
8374 if ((corType == CORINFO_TYPE_VALUECLASS) || (corType == CORINFO_TYPE_REFANY))
8376 // We have some kind of STRUCT being returned
8378 structPassingKind howToReturnStruct = SPK_Unknown;
8380 var_types returnType = getReturnTypeForStruct(methInfo->args.retTypeClass, &howToReturnStruct);
8382 if (howToReturnStruct == SPK_ByReference)
8393 var_types Compiler::impImportJitTestLabelMark(int numArgs)
8395 TestLabelAndNum tlAndN;
8399 StackEntry se = impPopStack();
8400 assert(se.seTypeInfo.GetType() == TI_INT);
8401 GenTree* val = se.val;
8402 assert(val->IsCnsIntOrI());
8403 tlAndN.m_tl = (TestLabel)val->AsIntConCommon()->IconValue();
8405 else if (numArgs == 3)
8407 StackEntry se = impPopStack();
8408 assert(se.seTypeInfo.GetType() == TI_INT);
8409 GenTree* val = se.val;
8410 assert(val->IsCnsIntOrI());
8411 tlAndN.m_num = val->AsIntConCommon()->IconValue();
8413 assert(se.seTypeInfo.GetType() == TI_INT);
8415 assert(val->IsCnsIntOrI());
8416 tlAndN.m_tl = (TestLabel)val->AsIntConCommon()->IconValue();
8423 StackEntry expSe = impPopStack();
8424 GenTree* node = expSe.val;
8426 // There are a small number of special cases, where we actually put the annotation on a subnode.
8427 if (tlAndN.m_tl == TL_LoopHoist && tlAndN.m_num >= 100)
8429 // A loop hoist annotation with value >= 100 means that the expression should be a static field access,
8430 // a GT_IND of a static field address, which should be the sum of a (hoistable) helper call and possibly some
8431 // offset within the the static field block whose address is returned by the helper call.
8432 // The annotation is saying that this address calculation, but not the entire access, should be hoisted.
8433 GenTree* helperCall = nullptr;
8434 assert(node->OperGet() == GT_IND);
8435 tlAndN.m_num -= 100;
8436 GetNodeTestData()->Set(node->gtOp.gtOp1, tlAndN);
8437 GetNodeTestData()->Remove(node);
8441 GetNodeTestData()->Set(node, tlAndN);
8444 impPushOnStack(node, expSe.seTypeInfo);
8445 return node->TypeGet();
8449 //-----------------------------------------------------------------------------------
8450 // impFixupCallStructReturn: For a call node that returns a struct type either
8451 // adjust the return type to an enregisterable type, or set the flag to indicate
8452 // struct return via retbuf arg.
8455 // call - GT_CALL GenTree node
8456 // retClsHnd - Class handle of return type of the call
8459 // Returns new GenTree node after fixing struct return of call node
8461 GenTree* Compiler::impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd)
8463 if (!varTypeIsStruct(call))
8468 call->gtRetClsHnd = retClsHnd;
8470 #if FEATURE_MULTIREG_RET
8471 // Initialize Return type descriptor of call node
8472 ReturnTypeDesc* retTypeDesc = call->GetReturnTypeDesc();
8473 retTypeDesc->InitializeStructReturnType(this, retClsHnd);
8474 #endif // FEATURE_MULTIREG_RET
8476 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
8478 // Not allowed for FEATURE_CORCLR which is the only SKU available for System V OSs.
8479 assert(!call->IsVarargs() && "varargs not allowed for System V OSs.");
8481 // The return type will remain as the incoming struct type unless normalized to a
8482 // single eightbyte return type below.
8483 call->gtReturnType = call->gtType;
8485 unsigned retRegCount = retTypeDesc->GetReturnRegCount();
8486 if (retRegCount != 0)
8488 if (retRegCount == 1)
8490 // struct returned in a single register
8491 call->gtReturnType = retTypeDesc->GetReturnRegType(0);
8495 // must be a struct returned in two registers
8496 assert(retRegCount == 2);
8498 if ((!call->CanTailCall()) && (!call->IsInlineCandidate()))
8500 // Force a call returning multi-reg struct to be always of the IR form
8503 // No need to assign a multi-reg struct to a local var if:
8504 // - It is a tail call or
8505 // - The call is marked for in-lining later
8506 return impAssignMultiRegTypeToVar(call, retClsHnd);
8512 // struct not returned in registers i.e returned via hiddden retbuf arg.
8513 call->gtCallMoreFlags |= GTF_CALL_M_RETBUFFARG;
8516 #else // not FEATURE_UNIX_AMD64_STRUCT_PASSING
8518 // Check for TYP_STRUCT type that wraps a primitive type
8519 // Such structs are returned using a single register
8520 // and we change the return type on those calls here.
8522 structPassingKind howToReturnStruct;
8523 var_types returnType = getReturnTypeForStruct(retClsHnd, &howToReturnStruct);
8525 if (howToReturnStruct == SPK_ByReference)
8527 assert(returnType == TYP_UNKNOWN);
8528 call->gtCallMoreFlags |= GTF_CALL_M_RETBUFFARG;
8532 assert(returnType != TYP_UNKNOWN);
8533 call->gtReturnType = returnType;
8535 // ToDo: Refactor this common code sequence into its own method as it is used 4+ times
8536 if ((returnType == TYP_LONG) && (compLongUsed == false))
8538 compLongUsed = true;
8540 else if (((returnType == TYP_FLOAT) || (returnType == TYP_DOUBLE)) && (compFloatingPointUsed == false))
8542 compFloatingPointUsed = true;
8545 #if FEATURE_MULTIREG_RET
8546 unsigned retRegCount = retTypeDesc->GetReturnRegCount();
8547 assert(retRegCount != 0);
8549 if (retRegCount >= 2)
8551 if ((!call->CanTailCall()) && (!call->IsInlineCandidate()))
8553 // Force a call returning multi-reg struct to be always of the IR form
8556 // No need to assign a multi-reg struct to a local var if:
8557 // - It is a tail call or
8558 // - The call is marked for in-lining later
8559 return impAssignMultiRegTypeToVar(call, retClsHnd);
8562 #endif // FEATURE_MULTIREG_RET
8565 #endif // not FEATURE_UNIX_AMD64_STRUCT_PASSING
8570 /*****************************************************************************
8571 For struct return values, re-type the operand in the case where the ABI
8572 does not use a struct return buffer
8573 Note that this method is only call for !_TARGET_X86_
8576 GenTree* Compiler::impFixupStructReturnType(GenTree* op, CORINFO_CLASS_HANDLE retClsHnd)
8578 assert(varTypeIsStruct(info.compRetType));
8579 assert(info.compRetBuffArg == BAD_VAR_NUM);
8581 #if defined(_TARGET_XARCH_)
8583 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
8584 // No VarArgs for CoreCLR on x64 Unix
8585 assert(!info.compIsVarArgs);
8587 // Is method returning a multi-reg struct?
8588 if (varTypeIsStruct(info.compRetNativeType) && IsMultiRegReturnedType(retClsHnd))
8590 // In case of multi-reg struct return, we force IR to be one of the following:
8591 // GT_RETURN(lclvar) or GT_RETURN(call). If op is anything other than a
8592 // lclvar or call, it is assigned to a temp to create: temp = op and GT_RETURN(tmp).
8594 if (op->gtOper == GT_LCL_VAR)
8596 // Make sure that this struct stays in memory and doesn't get promoted.
8597 unsigned lclNum = op->gtLclVarCommon.gtLclNum;
8598 lvaTable[lclNum].lvIsMultiRegRet = true;
8600 // TODO-1stClassStructs: Handle constant propagation and CSE-ing of multireg returns.
8601 op->gtFlags |= GTF_DONT_CSE;
8606 if (op->gtOper == GT_CALL)
8611 return impAssignMultiRegTypeToVar(op, retClsHnd);
8613 #else // !FEATURE_UNIX_AMD64_STRUCT_PASSING
8614 assert(info.compRetNativeType != TYP_STRUCT);
8615 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
8617 #elif FEATURE_MULTIREG_RET && defined(_TARGET_ARM_)
8619 if (varTypeIsStruct(info.compRetNativeType) && !info.compIsVarArgs && IsHfa(retClsHnd))
8621 if (op->gtOper == GT_LCL_VAR)
8623 // This LCL_VAR is an HFA return value, it stays as a TYP_STRUCT
8624 unsigned lclNum = op->gtLclVarCommon.gtLclNum;
8625 // Make sure this struct type stays as struct so that we can return it as an HFA
8626 lvaTable[lclNum].lvIsMultiRegRet = true;
8628 // TODO-1stClassStructs: Handle constant propagation and CSE-ing of multireg returns.
8629 op->gtFlags |= GTF_DONT_CSE;
8634 if (op->gtOper == GT_CALL)
8636 if (op->gtCall.IsVarargs())
8638 // We cannot tail call because control needs to return to fixup the calling
8639 // convention for result return.
8640 op->gtCall.gtCallMoreFlags &= ~GTF_CALL_M_TAILCALL;
8641 op->gtCall.gtCallMoreFlags &= ~GTF_CALL_M_EXPLICIT_TAILCALL;
8648 return impAssignMultiRegTypeToVar(op, retClsHnd);
8651 #elif FEATURE_MULTIREG_RET && defined(_TARGET_ARM64_)
8653 // Is method returning a multi-reg struct?
8654 if (IsMultiRegReturnedType(retClsHnd))
8656 if (op->gtOper == GT_LCL_VAR)
8658 // This LCL_VAR stays as a TYP_STRUCT
8659 unsigned lclNum = op->gtLclVarCommon.gtLclNum;
8661 // Make sure this struct type is not struct promoted
8662 lvaTable[lclNum].lvIsMultiRegRet = true;
8664 // TODO-1stClassStructs: Handle constant propagation and CSE-ing of multireg returns.
8665 op->gtFlags |= GTF_DONT_CSE;
8670 if (op->gtOper == GT_CALL)
8672 if (op->gtCall.IsVarargs())
8674 // We cannot tail call because control needs to return to fixup the calling
8675 // convention for result return.
8676 op->gtCall.gtCallMoreFlags &= ~GTF_CALL_M_TAILCALL;
8677 op->gtCall.gtCallMoreFlags &= ~GTF_CALL_M_EXPLICIT_TAILCALL;
8684 return impAssignMultiRegTypeToVar(op, retClsHnd);
8687 #endif // FEATURE_MULTIREG_RET && FEATURE_HFA
8690 // adjust the type away from struct to integral
8691 // and no normalizing
8692 if (op->gtOper == GT_LCL_VAR)
8694 op->ChangeOper(GT_LCL_FLD);
8696 else if (op->gtOper == GT_OBJ)
8698 GenTree* op1 = op->AsObj()->Addr();
8700 // We will fold away OBJ/ADDR
8701 // except for OBJ/ADDR/INDEX
8702 // as the array type influences the array element's offset
8703 // Later in this method we change op->gtType to info.compRetNativeType
8704 // This is not correct when op is a GT_INDEX as the starting offset
8705 // for the array elements 'elemOffs' is different for an array of
8706 // TYP_REF than an array of TYP_STRUCT (which simply wraps a TYP_REF)
8707 // Also refer to the GTF_INX_REFARR_LAYOUT flag
8709 if ((op1->gtOper == GT_ADDR) && (op1->gtOp.gtOp1->gtOper != GT_INDEX))
8711 // Change '*(&X)' to 'X' and see if we can do better
8712 op = op1->gtOp.gtOp1;
8713 goto REDO_RETURN_NODE;
8715 op->gtObj.gtClass = NO_CLASS_HANDLE;
8716 op->ChangeOperUnchecked(GT_IND);
8717 op->gtFlags |= GTF_IND_TGTANYWHERE;
8719 else if (op->gtOper == GT_CALL)
8721 if (op->AsCall()->TreatAsHasRetBufArg(this))
8723 // This must be one of those 'special' helpers that don't
8724 // really have a return buffer, but instead use it as a way
8725 // to keep the trees cleaner with fewer address-taken temps.
8727 // Well now we have to materialize the the return buffer as
8728 // an address-taken temp. Then we can return the temp.
8730 // NOTE: this code assumes that since the call directly
8731 // feeds the return, then the call must be returning the
8732 // same structure/class/type.
8734 unsigned tmpNum = lvaGrabTemp(true DEBUGARG("pseudo return buffer"));
8736 // No need to spill anything as we're about to return.
8737 impAssignTempGen(tmpNum, op, info.compMethodInfo->args.retTypeClass, (unsigned)CHECK_SPILL_NONE);
8739 // Don't create both a GT_ADDR & GT_OBJ just to undo all of that; instead,
8740 // jump directly to a GT_LCL_FLD.
8741 op = gtNewLclvNode(tmpNum, info.compRetNativeType);
8742 op->ChangeOper(GT_LCL_FLD);
8746 assert(info.compRetNativeType == op->gtCall.gtReturnType);
8748 // Don't change the gtType of the node just yet, it will get changed later.
8752 #if defined(FEATURE_HW_INTRINSICS) && defined(_TARGET_ARM64_)
8753 else if ((op->gtOper == GT_HWIntrinsic) && varTypeIsSIMD(op->gtType))
8755 // TODO-ARM64-FIXME Implement ARM64 ABI for Short Vectors properly
8756 // assert(op->gtType == info.compRetNativeType)
8757 if (op->gtType != info.compRetNativeType)
8759 // Insert a register move to keep target type of SIMD intrinsic intact
8760 op = gtNewScalarHWIntrinsicNode(info.compRetNativeType, op, NI_ARM64_NONE_MOV);
8764 else if (op->gtOper == GT_COMMA)
8766 op->gtOp.gtOp2 = impFixupStructReturnType(op->gtOp.gtOp2, retClsHnd);
8769 op->gtType = info.compRetNativeType;
8774 /*****************************************************************************
8775 CEE_LEAVE may be jumping out of a protected block, viz, a catch or a
8776 finally-protected try. We find the finally blocks protecting the current
8777 offset (in order) by walking over the complete exception table and
8778 finding enclosing clauses. This assumes that the table is sorted.
8779 This will create a series of BBJ_CALLFINALLY -> BBJ_CALLFINALLY ... -> BBJ_ALWAYS.
8781 If we are leaving a catch handler, we need to attach the
8782 CPX_ENDCATCHes to the correct BBJ_CALLFINALLY blocks.
8784 After this function, the BBJ_LEAVE block has been converted to a different type.
8787 #if !FEATURE_EH_FUNCLETS
8789 void Compiler::impImportLeave(BasicBlock* block)
8794 printf("\nBefore import CEE_LEAVE:\n");
8795 fgDispBasicBlocks();
8800 bool invalidatePreds = false; // If we create new blocks, invalidate the predecessor lists (if created)
8801 unsigned blkAddr = block->bbCodeOffs;
8802 BasicBlock* leaveTarget = block->bbJumpDest;
8803 unsigned jmpAddr = leaveTarget->bbCodeOffs;
8805 // LEAVE clears the stack, spill side effects, and set stack to 0
8807 impSpillSideEffects(true, (unsigned)CHECK_SPILL_ALL DEBUGARG("impImportLeave"));
8808 verCurrentState.esStackDepth = 0;
8810 assert(block->bbJumpKind == BBJ_LEAVE);
8811 assert(fgBBs == (BasicBlock**)0xCDCD || fgLookupBB(jmpAddr) != NULL); // should be a BB boundary
8813 BasicBlock* step = DUMMY_INIT(NULL);
8814 unsigned encFinallies = 0; // Number of enclosing finallies.
8815 GenTree* endCatches = NULL;
8816 GenTree* endLFin = NULL; // The statement tree to indicate the end of locally-invoked finally.
8821 for (XTnum = 0, HBtab = compHndBBtab; XTnum < compHndBBtabCount; XTnum++, HBtab++)
8823 // Grab the handler offsets
8825 IL_OFFSET tryBeg = HBtab->ebdTryBegOffs();
8826 IL_OFFSET tryEnd = HBtab->ebdTryEndOffs();
8827 IL_OFFSET hndBeg = HBtab->ebdHndBegOffs();
8828 IL_OFFSET hndEnd = HBtab->ebdHndEndOffs();
8830 /* Is this a catch-handler we are CEE_LEAVEing out of?
8831 * If so, we need to call CORINFO_HELP_ENDCATCH.
8834 if (jitIsBetween(blkAddr, hndBeg, hndEnd) && !jitIsBetween(jmpAddr, hndBeg, hndEnd))
8836 // Can't CEE_LEAVE out of a finally/fault handler
8837 if (HBtab->HasFinallyOrFaultHandler())
8838 BADCODE("leave out of fault/finally block");
8840 // Create the call to CORINFO_HELP_ENDCATCH
8841 GenTree* endCatch = gtNewHelperCallNode(CORINFO_HELP_ENDCATCH, TYP_VOID);
8843 // Make a list of all the currently pending endCatches
8845 endCatches = gtNewOperNode(GT_COMMA, TYP_VOID, endCatches, endCatch);
8847 endCatches = endCatch;
8852 printf("impImportLeave - BB%02u jumping out of catch handler EH#%u, adding call to "
8853 "CORINFO_HELP_ENDCATCH\n",
8854 block->bbNum, XTnum);
8858 else if (HBtab->HasFinallyHandler() && jitIsBetween(blkAddr, tryBeg, tryEnd) &&
8859 !jitIsBetween(jmpAddr, tryBeg, tryEnd))
8861 /* This is a finally-protected try we are jumping out of */
8863 /* If there are any pending endCatches, and we have already
8864 jumped out of a finally-protected try, then the endCatches
8865 have to be put in a block in an outer try for async
8866 exceptions to work correctly.
8867 Else, just use append to the original block */
8869 BasicBlock* callBlock;
8871 assert(!encFinallies == !endLFin); // if we have finallies, we better have an endLFin tree, and vice-versa
8873 if (encFinallies == 0)
8875 assert(step == DUMMY_INIT(NULL));
8877 callBlock->bbJumpKind = BBJ_CALLFINALLY; // convert the BBJ_LEAVE to BBJ_CALLFINALLY
8880 impAppendTree(endCatches, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
8885 printf("impImportLeave - jumping out of a finally-protected try, convert block to BBJ_CALLFINALLY "
8887 callBlock->dspToString());
8893 assert(step != DUMMY_INIT(NULL));
8895 /* Calling the finally block */
8896 callBlock = fgNewBBinRegion(BBJ_CALLFINALLY, XTnum + 1, 0, step);
8897 assert(step->bbJumpKind == BBJ_ALWAYS);
8898 step->bbJumpDest = callBlock; // the previous call to a finally returns to this call (to the next
8899 // finally in the chain)
8900 step->bbJumpDest->bbRefs++;
8902 /* The new block will inherit this block's weight */
8903 callBlock->setBBWeight(block->bbWeight);
8904 callBlock->bbFlags |= block->bbFlags & BBF_RUN_RARELY;
8909 printf("impImportLeave - jumping out of a finally-protected try, new BBJ_CALLFINALLY block %s\n",
8910 callBlock->dspToString());
8918 lastStmt = gtNewStmt(endCatches);
8919 endLFin->gtNext = lastStmt;
8920 lastStmt->gtPrev = endLFin;
8927 // note that this sets BBF_IMPORTED on the block
8928 impEndTreeList(callBlock, endLFin, lastStmt);
8931 step = fgNewBBafter(BBJ_ALWAYS, callBlock, true);
8932 /* The new block will inherit this block's weight */
8933 step->setBBWeight(block->bbWeight);
8934 step->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED | BBF_KEEP_BBJ_ALWAYS;
8939 printf("impImportLeave - jumping out of a finally-protected try, created step (BBJ_ALWAYS) block %s\n",
8940 step->dspToString());
8944 unsigned finallyNesting = compHndBBtab[XTnum].ebdHandlerNestingLevel;
8945 assert(finallyNesting <= compHndBBtabCount);
8947 callBlock->bbJumpDest = HBtab->ebdHndBeg; // This callBlock will call the "finally" handler.
8948 endLFin = new (this, GT_END_LFIN) GenTreeVal(GT_END_LFIN, TYP_VOID, finallyNesting);
8949 endLFin = gtNewStmt(endLFin);
8954 invalidatePreds = true;
8958 /* Append any remaining endCatches, if any */
8960 assert(!encFinallies == !endLFin);
8962 if (encFinallies == 0)
8964 assert(step == DUMMY_INIT(NULL));
8965 block->bbJumpKind = BBJ_ALWAYS; // convert the BBJ_LEAVE to a BBJ_ALWAYS
8968 impAppendTree(endCatches, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
8973 printf("impImportLeave - no enclosing finally-protected try blocks; convert CEE_LEAVE block to BBJ_ALWAYS "
8975 block->dspToString());
8981 // If leaveTarget is the start of another try block, we want to make sure that
8982 // we do not insert finalStep into that try block. Hence, we find the enclosing
8984 unsigned tryIndex = bbFindInnermostCommonTryRegion(step, leaveTarget);
8986 // Insert a new BB either in the try region indicated by tryIndex or
8987 // the handler region indicated by leaveTarget->bbHndIndex,
8988 // depending on which is the inner region.
8989 BasicBlock* finalStep = fgNewBBinRegion(BBJ_ALWAYS, tryIndex, leaveTarget->bbHndIndex, step);
8990 finalStep->bbFlags |= BBF_KEEP_BBJ_ALWAYS;
8991 step->bbJumpDest = finalStep;
8993 /* The new block will inherit this block's weight */
8994 finalStep->setBBWeight(block->bbWeight);
8995 finalStep->bbFlags |= block->bbFlags & BBF_RUN_RARELY;
9000 printf("impImportLeave - finalStep block required (encFinallies(%d) > 0), new block %s\n", encFinallies,
9001 finalStep->dspToString());
9009 lastStmt = gtNewStmt(endCatches);
9010 endLFin->gtNext = lastStmt;
9011 lastStmt->gtPrev = endLFin;
9018 impEndTreeList(finalStep, endLFin, lastStmt);
9020 finalStep->bbJumpDest = leaveTarget; // this is the ultimate destination of the LEAVE
9022 // Queue up the jump target for importing
9024 impImportBlockPending(leaveTarget);
9026 invalidatePreds = true;
9029 if (invalidatePreds && fgComputePredsDone)
9031 JITDUMP("\n**** impImportLeave - Removing preds after creating new blocks\n");
9036 fgVerifyHandlerTab();
9040 printf("\nAfter import CEE_LEAVE:\n");
9041 fgDispBasicBlocks();
9047 #else // FEATURE_EH_FUNCLETS
9049 void Compiler::impImportLeave(BasicBlock* block)
9054 printf("\nBefore import CEE_LEAVE in BB%02u (targetting BB%02u):\n", block->bbNum, block->bbJumpDest->bbNum);
9055 fgDispBasicBlocks();
9060 bool invalidatePreds = false; // If we create new blocks, invalidate the predecessor lists (if created)
9061 unsigned blkAddr = block->bbCodeOffs;
9062 BasicBlock* leaveTarget = block->bbJumpDest;
9063 unsigned jmpAddr = leaveTarget->bbCodeOffs;
9065 // LEAVE clears the stack, spill side effects, and set stack to 0
9067 impSpillSideEffects(true, (unsigned)CHECK_SPILL_ALL DEBUGARG("impImportLeave"));
9068 verCurrentState.esStackDepth = 0;
9070 assert(block->bbJumpKind == BBJ_LEAVE);
9071 assert(fgBBs == (BasicBlock**)0xCDCD || fgLookupBB(jmpAddr) != nullptr); // should be a BB boundary
9073 BasicBlock* step = nullptr;
9077 // No step type; step == NULL.
9080 // Is the step block the BBJ_ALWAYS block of a BBJ_CALLFINALLY/BBJ_ALWAYS pair?
9081 // That is, is step->bbJumpDest where a finally will return to?
9084 // The step block is a catch return.
9087 // The step block is in a "try", created as the target for a finally return or the target for a catch return.
9090 StepType stepType = ST_None;
9095 for (XTnum = 0, HBtab = compHndBBtab; XTnum < compHndBBtabCount; XTnum++, HBtab++)
9097 // Grab the handler offsets
9099 IL_OFFSET tryBeg = HBtab->ebdTryBegOffs();
9100 IL_OFFSET tryEnd = HBtab->ebdTryEndOffs();
9101 IL_OFFSET hndBeg = HBtab->ebdHndBegOffs();
9102 IL_OFFSET hndEnd = HBtab->ebdHndEndOffs();
9104 /* Is this a catch-handler we are CEE_LEAVEing out of?
9107 if (jitIsBetween(blkAddr, hndBeg, hndEnd) && !jitIsBetween(jmpAddr, hndBeg, hndEnd))
9109 // Can't CEE_LEAVE out of a finally/fault handler
9110 if (HBtab->HasFinallyOrFaultHandler())
9112 BADCODE("leave out of fault/finally block");
9115 /* We are jumping out of a catch */
9117 if (step == nullptr)
9120 step->bbJumpKind = BBJ_EHCATCHRET; // convert the BBJ_LEAVE to BBJ_EHCATCHRET
9121 stepType = ST_Catch;
9126 printf("impImportLeave - jumping out of a catch (EH#%u), convert block BB%02u to BBJ_EHCATCHRET "
9128 XTnum, step->bbNum);
9134 BasicBlock* exitBlock;
9136 /* Create a new catch exit block in the catch region for the existing step block to jump to in this
9138 exitBlock = fgNewBBinRegion(BBJ_EHCATCHRET, 0, XTnum + 1, step);
9140 assert(step->bbJumpKind == BBJ_ALWAYS || step->bbJumpKind == BBJ_EHCATCHRET);
9141 step->bbJumpDest = exitBlock; // the previous step (maybe a call to a nested finally, or a nested catch
9142 // exit) returns to this block
9143 step->bbJumpDest->bbRefs++;
9145 #if defined(_TARGET_ARM_)
9146 if (stepType == ST_FinallyReturn)
9148 assert(step->bbJumpKind == BBJ_ALWAYS);
9149 // Mark the target of a finally return
9150 step->bbJumpDest->bbFlags |= BBF_FINALLY_TARGET;
9152 #endif // defined(_TARGET_ARM_)
9154 /* The new block will inherit this block's weight */
9155 exitBlock->setBBWeight(block->bbWeight);
9156 exitBlock->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED;
9158 /* This exit block is the new step */
9160 stepType = ST_Catch;
9162 invalidatePreds = true;
9167 printf("impImportLeave - jumping out of a catch (EH#%u), new BBJ_EHCATCHRET block BB%02u\n", XTnum,
9173 else if (HBtab->HasFinallyHandler() && jitIsBetween(blkAddr, tryBeg, tryEnd) &&
9174 !jitIsBetween(jmpAddr, tryBeg, tryEnd))
9176 /* We are jumping out of a finally-protected try */
9178 BasicBlock* callBlock;
9180 if (step == nullptr)
9182 #if FEATURE_EH_CALLFINALLY_THUNKS
9184 // Put the call to the finally in the enclosing region.
9185 unsigned callFinallyTryIndex =
9186 (HBtab->ebdEnclosingTryIndex == EHblkDsc::NO_ENCLOSING_INDEX) ? 0 : HBtab->ebdEnclosingTryIndex + 1;
9187 unsigned callFinallyHndIndex =
9188 (HBtab->ebdEnclosingHndIndex == EHblkDsc::NO_ENCLOSING_INDEX) ? 0 : HBtab->ebdEnclosingHndIndex + 1;
9189 callBlock = fgNewBBinRegion(BBJ_CALLFINALLY, callFinallyTryIndex, callFinallyHndIndex, block);
9191 // Convert the BBJ_LEAVE to BBJ_ALWAYS, jumping to the new BBJ_CALLFINALLY. This is because
9192 // the new BBJ_CALLFINALLY is in a different EH region, thus it can't just replace the BBJ_LEAVE,
9193 // which might be in the middle of the "try". In most cases, the BBJ_ALWAYS will jump to the
9194 // next block, and flow optimizations will remove it.
9195 block->bbJumpKind = BBJ_ALWAYS;
9196 block->bbJumpDest = callBlock;
9197 block->bbJumpDest->bbRefs++;
9199 /* The new block will inherit this block's weight */
9200 callBlock->setBBWeight(block->bbWeight);
9201 callBlock->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED;
9206 printf("impImportLeave - jumping out of a finally-protected try (EH#%u), convert block BB%02u to "
9207 "BBJ_ALWAYS, add BBJ_CALLFINALLY block BB%02u\n",
9208 XTnum, block->bbNum, callBlock->bbNum);
9212 #else // !FEATURE_EH_CALLFINALLY_THUNKS
9215 callBlock->bbJumpKind = BBJ_CALLFINALLY; // convert the BBJ_LEAVE to BBJ_CALLFINALLY
9220 printf("impImportLeave - jumping out of a finally-protected try (EH#%u), convert block BB%02u to "
9221 "BBJ_CALLFINALLY block\n",
9222 XTnum, callBlock->bbNum);
9226 #endif // !FEATURE_EH_CALLFINALLY_THUNKS
9230 // Calling the finally block. We already have a step block that is either the call-to-finally from a
9231 // more nested try/finally (thus we are jumping out of multiple nested 'try' blocks, each protected by
9232 // a 'finally'), or the step block is the return from a catch.
9234 // Due to ThreadAbortException, we can't have the catch return target the call-to-finally block
9235 // directly. Note that if a 'catch' ends without resetting the ThreadAbortException, the VM will
9236 // automatically re-raise the exception, using the return address of the catch (that is, the target
9237 // block of the BBJ_EHCATCHRET) as the re-raise address. If this address is in a finally, the VM will
9238 // refuse to do the re-raise, and the ThreadAbortException will get eaten (and lost). On AMD64/ARM64,
9239 // we put the call-to-finally thunk in a special "cloned finally" EH region that does look like a
9240 // finally clause to the VM. Thus, on these platforms, we can't have BBJ_EHCATCHRET target a
9241 // BBJ_CALLFINALLY directly. (Note that on ARM32, we don't mark the thunk specially -- it lives directly
9242 // within the 'try' region protected by the finally, since we generate code in such a way that execution
9243 // never returns to the call-to-finally call, and the finally-protected 'try' region doesn't appear on
9246 assert(step->bbJumpKind == BBJ_ALWAYS || step->bbJumpKind == BBJ_EHCATCHRET);
9248 #if FEATURE_EH_CALLFINALLY_THUNKS
9249 if (step->bbJumpKind == BBJ_EHCATCHRET)
9251 // Need to create another step block in the 'try' region that will actually branch to the
9252 // call-to-finally thunk.
9253 BasicBlock* step2 = fgNewBBinRegion(BBJ_ALWAYS, XTnum + 1, 0, step);
9254 step->bbJumpDest = step2;
9255 step->bbJumpDest->bbRefs++;
9256 step2->setBBWeight(block->bbWeight);
9257 step2->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED;
9262 printf("impImportLeave - jumping out of a finally-protected try (EH#%u), step block is "
9263 "BBJ_EHCATCHRET (BB%02u), new BBJ_ALWAYS step-step block BB%02u\n",
9264 XTnum, step->bbNum, step2->bbNum);
9269 assert(stepType == ST_Catch); // Leave it as catch type for now.
9271 #endif // FEATURE_EH_CALLFINALLY_THUNKS
9273 #if FEATURE_EH_CALLFINALLY_THUNKS
9274 unsigned callFinallyTryIndex =
9275 (HBtab->ebdEnclosingTryIndex == EHblkDsc::NO_ENCLOSING_INDEX) ? 0 : HBtab->ebdEnclosingTryIndex + 1;
9276 unsigned callFinallyHndIndex =
9277 (HBtab->ebdEnclosingHndIndex == EHblkDsc::NO_ENCLOSING_INDEX) ? 0 : HBtab->ebdEnclosingHndIndex + 1;
9278 #else // !FEATURE_EH_CALLFINALLY_THUNKS
9279 unsigned callFinallyTryIndex = XTnum + 1;
9280 unsigned callFinallyHndIndex = 0; // don't care
9281 #endif // !FEATURE_EH_CALLFINALLY_THUNKS
9283 callBlock = fgNewBBinRegion(BBJ_CALLFINALLY, callFinallyTryIndex, callFinallyHndIndex, step);
9284 step->bbJumpDest = callBlock; // the previous call to a finally returns to this call (to the next
9285 // finally in the chain)
9286 step->bbJumpDest->bbRefs++;
9288 #if defined(_TARGET_ARM_)
9289 if (stepType == ST_FinallyReturn)
9291 assert(step->bbJumpKind == BBJ_ALWAYS);
9292 // Mark the target of a finally return
9293 step->bbJumpDest->bbFlags |= BBF_FINALLY_TARGET;
9295 #endif // defined(_TARGET_ARM_)
9297 /* The new block will inherit this block's weight */
9298 callBlock->setBBWeight(block->bbWeight);
9299 callBlock->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED;
9304 printf("impImportLeave - jumping out of a finally-protected try (EH#%u), new BBJ_CALLFINALLY block "
9306 XTnum, callBlock->bbNum);
9311 step = fgNewBBafter(BBJ_ALWAYS, callBlock, true);
9312 stepType = ST_FinallyReturn;
9314 /* The new block will inherit this block's weight */
9315 step->setBBWeight(block->bbWeight);
9316 step->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED | BBF_KEEP_BBJ_ALWAYS;
9321 printf("impImportLeave - jumping out of a finally-protected try (EH#%u), created step (BBJ_ALWAYS) "
9323 XTnum, step->bbNum);
9327 callBlock->bbJumpDest = HBtab->ebdHndBeg; // This callBlock will call the "finally" handler.
9329 invalidatePreds = true;
9331 else if (HBtab->HasCatchHandler() && jitIsBetween(blkAddr, tryBeg, tryEnd) &&
9332 !jitIsBetween(jmpAddr, tryBeg, tryEnd))
9334 // We are jumping out of a catch-protected try.
9336 // If we are returning from a call to a finally, then we must have a step block within a try
9337 // that is protected by a catch. This is so when unwinding from that finally (e.g., if code within the
9338 // finally raises an exception), the VM will find this step block, notice that it is in a protected region,
9339 // and invoke the appropriate catch.
9341 // We also need to handle a special case with the handling of ThreadAbortException. If a try/catch
9342 // catches a ThreadAbortException (which might be because it catches a parent, e.g. System.Exception),
9343 // and the catch doesn't call System.Threading.Thread::ResetAbort(), then when the catch returns to the VM,
9344 // the VM will automatically re-raise the ThreadAbortException. When it does this, it uses the target
9345 // address of the catch return as the new exception address. That is, the re-raised exception appears to
9346 // occur at the catch return address. If this exception return address skips an enclosing try/catch that
9347 // catches ThreadAbortException, then the enclosing try/catch will not catch the exception, as it should.
9352 // // something here raises ThreadAbortException
9353 // LEAVE LABEL_1; // no need to stop at LABEL_2
9354 // } catch (Exception) {
9355 // // This catches ThreadAbortException, but doesn't call System.Threading.Thread::ResetAbort(), so
9356 // // ThreadAbortException is re-raised by the VM at the address specified by the LEAVE opcode.
9357 // // This is bad, since it means the outer try/catch won't get a chance to catch the re-raised
9358 // // ThreadAbortException. So, instead, create step block LABEL_2 and LEAVE to that. We only
9359 // // need to do this transformation if the current EH block is a try/catch that catches
9360 // // ThreadAbortException (or one of its parents), however we might not be able to find that
9361 // // information, so currently we do it for all catch types.
9362 // LEAVE LABEL_1; // Convert this to LEAVE LABEL2;
9364 // LABEL_2: LEAVE LABEL_1; // inserted by this step creation code
9365 // } catch (ThreadAbortException) {
9369 // Note that this pattern isn't theoretical: it occurs in ASP.NET, in IL code generated by the Roslyn C#
9372 if ((stepType == ST_FinallyReturn) || (stepType == ST_Catch))
9374 BasicBlock* catchStep;
9378 if (stepType == ST_FinallyReturn)
9380 assert(step->bbJumpKind == BBJ_ALWAYS);
9384 assert(stepType == ST_Catch);
9385 assert(step->bbJumpKind == BBJ_EHCATCHRET);
9388 /* Create a new exit block in the try region for the existing step block to jump to in this scope */
9389 catchStep = fgNewBBinRegion(BBJ_ALWAYS, XTnum + 1, 0, step);
9390 step->bbJumpDest = catchStep;
9391 step->bbJumpDest->bbRefs++;
9393 #if defined(_TARGET_ARM_)
9394 if (stepType == ST_FinallyReturn)
9396 // Mark the target of a finally return
9397 step->bbJumpDest->bbFlags |= BBF_FINALLY_TARGET;
9399 #endif // defined(_TARGET_ARM_)
9401 /* The new block will inherit this block's weight */
9402 catchStep->setBBWeight(block->bbWeight);
9403 catchStep->bbFlags |= (block->bbFlags & BBF_RUN_RARELY) | BBF_IMPORTED;
9408 if (stepType == ST_FinallyReturn)
9410 printf("impImportLeave - return from finally jumping out of a catch-protected try (EH#%u), new "
9411 "BBJ_ALWAYS block BB%02u\n",
9412 XTnum, catchStep->bbNum);
9416 assert(stepType == ST_Catch);
9417 printf("impImportLeave - return from catch jumping out of a catch-protected try (EH#%u), new "
9418 "BBJ_ALWAYS block BB%02u\n",
9419 XTnum, catchStep->bbNum);
9424 /* This block is the new step */
9428 invalidatePreds = true;
9433 if (step == nullptr)
9435 block->bbJumpKind = BBJ_ALWAYS; // convert the BBJ_LEAVE to a BBJ_ALWAYS
9440 printf("impImportLeave - no enclosing finally-protected try blocks or catch handlers; convert CEE_LEAVE "
9441 "block BB%02u to BBJ_ALWAYS\n",
9448 step->bbJumpDest = leaveTarget; // this is the ultimate destination of the LEAVE
9450 #if defined(_TARGET_ARM_)
9451 if (stepType == ST_FinallyReturn)
9453 assert(step->bbJumpKind == BBJ_ALWAYS);
9454 // Mark the target of a finally return
9455 step->bbJumpDest->bbFlags |= BBF_FINALLY_TARGET;
9457 #endif // defined(_TARGET_ARM_)
9462 printf("impImportLeave - final destination of step blocks set to BB%02u\n", leaveTarget->bbNum);
9466 // Queue up the jump target for importing
9468 impImportBlockPending(leaveTarget);
9471 if (invalidatePreds && fgComputePredsDone)
9473 JITDUMP("\n**** impImportLeave - Removing preds after creating new blocks\n");
9478 fgVerifyHandlerTab();
9482 printf("\nAfter import CEE_LEAVE:\n");
9483 fgDispBasicBlocks();
9489 #endif // FEATURE_EH_FUNCLETS
9491 /*****************************************************************************/
9492 // This is called when reimporting a leave block. It resets the JumpKind,
9493 // JumpDest, and bbNext to the original values
9495 void Compiler::impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr)
9497 #if FEATURE_EH_FUNCLETS
9498 // With EH Funclets, while importing leave opcode we create another block ending with BBJ_ALWAYS (call it B1)
9499 // and the block containing leave (say B0) is marked as BBJ_CALLFINALLY. Say for some reason we reimport B0,
9500 // it is reset (in this routine) by marking as ending with BBJ_LEAVE and further down when B0 is reimported, we
9501 // create another BBJ_ALWAYS (call it B2). In this process B1 gets orphaned and any blocks to which B1 is the
9502 // only predecessor are also considered orphans and attempted to be deleted.
9509 // leave OUTSIDE; // B0 is the block containing this leave, following this would be B1
9514 // In the above nested try-finally example, we create a step block (call it Bstep) which in branches to a block
9515 // where a finally would branch to (and such block is marked as finally target). Block B1 branches to step block.
9516 // Because of re-import of B0, Bstep is also orphaned. Since Bstep is a finally target it cannot be removed. To
9517 // work around this we will duplicate B0 (call it B0Dup) before reseting. B0Dup is marked as BBJ_CALLFINALLY and
9518 // only serves to pair up with B1 (BBJ_ALWAYS) that got orphaned. Now during orphan block deletion B0Dup and B1
9519 // will be treated as pair and handled correctly.
9520 if (block->bbJumpKind == BBJ_CALLFINALLY)
9522 BasicBlock* dupBlock = bbNewBasicBlock(block->bbJumpKind);
9523 dupBlock->bbFlags = block->bbFlags;
9524 dupBlock->bbJumpDest = block->bbJumpDest;
9525 dupBlock->copyEHRegion(block);
9526 dupBlock->bbCatchTyp = block->bbCatchTyp;
9528 // Mark this block as
9529 // a) not referenced by any other block to make sure that it gets deleted
9531 // c) prevent from being imported
9534 dupBlock->bbRefs = 0;
9535 dupBlock->bbWeight = 0;
9536 dupBlock->bbFlags |= BBF_IMPORTED | BBF_INTERNAL | BBF_RUN_RARELY;
9538 // Insert the block right after the block which is getting reset so that BBJ_CALLFINALLY and BBJ_ALWAYS
9539 // will be next to each other.
9540 fgInsertBBafter(block, dupBlock);
9545 printf("New Basic Block BB%02u duplicate of BB%02u created.\n", dupBlock->bbNum, block->bbNum);
9549 #endif // FEATURE_EH_FUNCLETS
9551 block->bbJumpKind = BBJ_LEAVE;
9553 block->bbJumpDest = fgLookupBB(jmpAddr);
9555 // We will leave the BBJ_ALWAYS block we introduced. When it's reimported
9556 // the BBJ_ALWAYS block will be unreachable, and will be removed after. The
9557 // reason we don't want to remove the block at this point is that if we call
9558 // fgInitBBLookup() again we will do it wrong as the BBJ_ALWAYS block won't be
9559 // added and the linked list length will be different than fgBBcount.
9562 /*****************************************************************************/
9563 // Get the first non-prefix opcode. Used for verification of valid combinations
9564 // of prefixes and actual opcodes.
9566 static OPCODE impGetNonPrefixOpcode(const BYTE* codeAddr, const BYTE* codeEndp)
9568 while (codeAddr < codeEndp)
9570 OPCODE opcode = (OPCODE)getU1LittleEndian(codeAddr);
9571 codeAddr += sizeof(__int8);
9573 if (opcode == CEE_PREFIX1)
9575 if (codeAddr >= codeEndp)
9579 opcode = (OPCODE)(getU1LittleEndian(codeAddr) + 256);
9580 codeAddr += sizeof(__int8);
9588 case CEE_CONSTRAINED:
9595 codeAddr += opcodeSizes[opcode];
9601 /*****************************************************************************/
9602 // Checks whether the opcode is a valid opcode for volatile. and unaligned. prefixes
9604 static void impValidateMemoryAccessOpcode(const BYTE* codeAddr, const BYTE* codeEndp, bool volatilePrefix)
9606 OPCODE opcode = impGetNonPrefixOpcode(codeAddr, codeEndp);
9609 // Opcode of all ldind and stdind happen to be in continuous, except stind.i.
9610 ((CEE_LDIND_I1 <= opcode) && (opcode <= CEE_STIND_R8)) || (opcode == CEE_STIND_I) ||
9611 (opcode == CEE_LDFLD) || (opcode == CEE_STFLD) || (opcode == CEE_LDOBJ) || (opcode == CEE_STOBJ) ||
9612 (opcode == CEE_INITBLK) || (opcode == CEE_CPBLK) ||
9613 // volatile. prefix is allowed with the ldsfld and stsfld
9614 (volatilePrefix && ((opcode == CEE_LDSFLD) || (opcode == CEE_STSFLD)))))
9616 BADCODE("Invalid opcode for unaligned. or volatile. prefix");
9620 /*****************************************************************************/
9624 #undef RETURN // undef contracts RETURN macro
9639 const static controlFlow_t controlFlow[] = {
9640 #define OPDEF(c, s, pop, push, args, type, l, s1, s2, flow) flow,
9641 #include "opcode.def"
9647 /*****************************************************************************
9648 * Determine the result type of an arithemetic operation
9649 * On 64-bit inserts upcasts when native int is mixed with int32
9651 var_types Compiler::impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTree** pOp1, GenTree** pOp2)
9653 var_types type = TYP_UNDEF;
9654 GenTree* op1 = *pOp1;
9655 GenTree* op2 = *pOp2;
9657 // Arithemetic operations are generally only allowed with
9658 // primitive types, but certain operations are allowed
9661 if ((oper == GT_SUB) && (genActualType(op1->TypeGet()) == TYP_BYREF || genActualType(op2->TypeGet()) == TYP_BYREF))
9663 if ((genActualType(op1->TypeGet()) == TYP_BYREF) && (genActualType(op2->TypeGet()) == TYP_BYREF))
9665 // byref1-byref2 => gives a native int
9668 else if (genActualTypeIsIntOrI(op1->TypeGet()) && (genActualType(op2->TypeGet()) == TYP_BYREF))
9670 // [native] int - byref => gives a native int
9673 // The reason is that it is possible, in managed C++,
9674 // to have a tree like this:
9681 // const(h) int addr byref
9683 // <BUGNUM> VSW 318822 </BUGNUM>
9685 // So here we decide to make the resulting type to be a native int.
9686 CLANG_FORMAT_COMMENT_ANCHOR;
9688 #ifdef _TARGET_64BIT_
9689 if (genActualType(op1->TypeGet()) != TYP_I_IMPL)
9691 // insert an explicit upcast
9692 op1 = *pOp1 = gtNewCastNode(TYP_I_IMPL, op1, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9694 #endif // _TARGET_64BIT_
9700 // byref - [native] int => gives a byref
9701 assert(genActualType(op1->TypeGet()) == TYP_BYREF && genActualTypeIsIntOrI(op2->TypeGet()));
9703 #ifdef _TARGET_64BIT_
9704 if ((genActualType(op2->TypeGet()) != TYP_I_IMPL))
9706 // insert an explicit upcast
9707 op2 = *pOp2 = gtNewCastNode(TYP_I_IMPL, op2, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9709 #endif // _TARGET_64BIT_
9714 else if ((oper == GT_ADD) &&
9715 (genActualType(op1->TypeGet()) == TYP_BYREF || genActualType(op2->TypeGet()) == TYP_BYREF))
9717 // byref + [native] int => gives a byref
9719 // [native] int + byref => gives a byref
9721 // only one can be a byref : byref op byref not allowed
9722 assert(genActualType(op1->TypeGet()) != TYP_BYREF || genActualType(op2->TypeGet()) != TYP_BYREF);
9723 assert(genActualTypeIsIntOrI(op1->TypeGet()) || genActualTypeIsIntOrI(op2->TypeGet()));
9725 #ifdef _TARGET_64BIT_
9726 if (genActualType(op2->TypeGet()) == TYP_BYREF)
9728 if (genActualType(op1->TypeGet()) != TYP_I_IMPL)
9730 // insert an explicit upcast
9731 op1 = *pOp1 = gtNewCastNode(TYP_I_IMPL, op1, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9734 else if (genActualType(op2->TypeGet()) != TYP_I_IMPL)
9736 // insert an explicit upcast
9737 op2 = *pOp2 = gtNewCastNode(TYP_I_IMPL, op2, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9739 #endif // _TARGET_64BIT_
9743 #ifdef _TARGET_64BIT_
9744 else if (genActualType(op1->TypeGet()) == TYP_I_IMPL || genActualType(op2->TypeGet()) == TYP_I_IMPL)
9746 assert(!varTypeIsFloating(op1->gtType) && !varTypeIsFloating(op2->gtType));
9748 // int + long => gives long
9749 // long + int => gives long
9750 // we get this because in the IL the long isn't Int64, it's just IntPtr
9752 if (genActualType(op1->TypeGet()) != TYP_I_IMPL)
9754 // insert an explicit upcast
9755 op1 = *pOp1 = gtNewCastNode(TYP_I_IMPL, op1, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9757 else if (genActualType(op2->TypeGet()) != TYP_I_IMPL)
9759 // insert an explicit upcast
9760 op2 = *pOp2 = gtNewCastNode(TYP_I_IMPL, op2, fUnsigned, fUnsigned ? TYP_U_IMPL : TYP_I_IMPL);
9765 #else // 32-bit TARGET
9766 else if (genActualType(op1->TypeGet()) == TYP_LONG || genActualType(op2->TypeGet()) == TYP_LONG)
9768 assert(!varTypeIsFloating(op1->gtType) && !varTypeIsFloating(op2->gtType));
9770 // int + long => gives long
9771 // long + int => gives long
9775 #endif // _TARGET_64BIT_
9778 // int + int => gives an int
9779 assert(genActualType(op1->TypeGet()) != TYP_BYREF && genActualType(op2->TypeGet()) != TYP_BYREF);
9781 assert(genActualType(op1->TypeGet()) == genActualType(op2->TypeGet()) ||
9782 varTypeIsFloating(op1->gtType) && varTypeIsFloating(op2->gtType));
9784 type = genActualType(op1->gtType);
9786 #if FEATURE_X87_DOUBLES
9788 // For x87, since we only have 1 size of registers, prefer double
9789 // For everybody else, be more precise
9790 if (type == TYP_FLOAT)
9793 #else // !FEATURE_X87_DOUBLES
9795 // If both operands are TYP_FLOAT, then leave it as TYP_FLOAT.
9796 // Otherwise, turn floats into doubles
9797 if ((type == TYP_FLOAT) && (genActualType(op2->gtType) != TYP_FLOAT))
9799 assert(genActualType(op2->gtType) == TYP_DOUBLE);
9803 #endif // FEATURE_X87_DOUBLES
9806 #if FEATURE_X87_DOUBLES
9807 assert(type == TYP_BYREF || type == TYP_DOUBLE || type == TYP_LONG || type == TYP_INT);
9808 #else // FEATURE_X87_DOUBLES
9809 assert(type == TYP_BYREF || type == TYP_DOUBLE || type == TYP_FLOAT || type == TYP_LONG || type == TYP_INT);
9810 #endif // FEATURE_X87_DOUBLES
9815 //------------------------------------------------------------------------
9816 // impOptimizeCastClassOrIsInst: attempt to resolve a cast when jitting
9819 // op1 - value to cast
9820 // pResolvedToken - resolved token for type to cast to
9821 // isCastClass - true if this is a castclass, false if isinst
9824 // tree representing optimized cast, or null if no optimization possible
9826 GenTree* Compiler::impOptimizeCastClassOrIsInst(GenTree* op1, CORINFO_RESOLVED_TOKEN* pResolvedToken, bool isCastClass)
9828 assert(op1->TypeGet() == TYP_REF);
9830 // Don't optimize for minopts or debug codegen.
9831 if (opts.compDbgCode || opts.MinOpts())
9836 // See what we know about the type of the object being cast.
9837 bool isExact = false;
9838 bool isNonNull = false;
9839 CORINFO_CLASS_HANDLE fromClass = gtGetClassHandle(op1, &isExact, &isNonNull);
9840 GenTree* optResult = nullptr;
9842 if (fromClass != nullptr)
9844 CORINFO_CLASS_HANDLE toClass = pResolvedToken->hClass;
9845 JITDUMP("\nConsidering optimization of %s from %s%p (%s) to %p (%s)\n", isCastClass ? "castclass" : "isinst",
9846 isExact ? "exact " : "", dspPtr(fromClass), info.compCompHnd->getClassName(fromClass), dspPtr(toClass),
9847 info.compCompHnd->getClassName(toClass));
9849 // Perhaps we know if the cast will succeed or fail.
9850 TypeCompareState castResult = info.compCompHnd->compareTypesForCast(fromClass, toClass);
9852 if (castResult == TypeCompareState::Must)
9854 // Cast will succeed, result is simply op1.
9855 JITDUMP("Cast will succeed, optimizing to simply return input\n");
9858 else if (castResult == TypeCompareState::MustNot)
9860 // See if we can sharpen exactness by looking for final classes
9863 DWORD flags = info.compCompHnd->getClassAttribs(fromClass);
9864 DWORD flagsMask = CORINFO_FLG_FINAL | CORINFO_FLG_MARSHAL_BYREF | CORINFO_FLG_CONTEXTFUL |
9865 CORINFO_FLG_VARIANCE | CORINFO_FLG_ARRAY;
9866 isExact = ((flags & flagsMask) == CORINFO_FLG_FINAL);
9869 // Cast to exact type will fail. Handle case where we have
9870 // an exact type (that is, fromClass is not a subtype)
9871 // and we're not going to throw on failure.
9872 if (isExact && !isCastClass)
9874 JITDUMP("Cast will fail, optimizing to return null\n");
9875 GenTree* result = gtNewIconNode(0, TYP_REF);
9877 // If the cast was fed by a box, we can remove that too.
9878 if (op1->IsBoxedValue())
9880 JITDUMP("Also removing upstream box\n");
9881 gtTryRemoveBoxUpstreamEffects(op1);
9888 JITDUMP("Not optimizing failing castclass (yet)\n");
9892 JITDUMP("Can't optimize since fromClass is inexact\n");
9897 JITDUMP("Result of cast unknown, must generate runtime test\n");
9902 JITDUMP("\nCan't optimize since fromClass is unknown\n");
9908 //------------------------------------------------------------------------
9909 // impCastClassOrIsInstToTree: build and import castclass/isinst
9912 // op1 - value to cast
9913 // op2 - type handle for type to cast to
9914 // pResolvedToken - resolved token from the cast operation
9915 // isCastClass - true if this is castclass, false means isinst
9918 // Tree representing the cast
9921 // May expand into a series of runtime checks or a helper call.
9923 GenTree* Compiler::impCastClassOrIsInstToTree(GenTree* op1,
9925 CORINFO_RESOLVED_TOKEN* pResolvedToken,
9928 assert(op1->TypeGet() == TYP_REF);
9930 // Optimistically assume the jit should expand this as an inline test
9931 bool shouldExpandInline = true;
9933 // Profitability check.
9935 // Don't bother with inline expansion when jit is trying to
9936 // generate code quickly, or the cast is in code that won't run very
9937 // often, or the method already is pretty big.
9938 if (compCurBB->isRunRarely() || opts.compDbgCode || opts.MinOpts())
9940 // not worth the code expansion if jitting fast or in a rarely run block
9941 shouldExpandInline = false;
9943 else if ((op1->gtFlags & GTF_GLOB_EFFECT) && lvaHaveManyLocals())
9945 // not worth creating an untracked local variable
9946 shouldExpandInline = false;
9949 // Pessimistically assume the jit cannot expand this as an inline test
9950 bool canExpandInline = false;
9951 const CorInfoHelpFunc helper = info.compCompHnd->getCastingHelper(pResolvedToken, isCastClass);
9955 // Not all classclass/isinst operations can be inline expanded.
9956 // Check legality only if an inline expansion is desirable.
9957 if (shouldExpandInline)
9961 // Jit can only inline expand the normal CHKCASTCLASS helper.
9962 canExpandInline = (helper == CORINFO_HELP_CHKCASTCLASS);
9966 if (helper == CORINFO_HELP_ISINSTANCEOFCLASS)
9968 // Check the class attributes.
9969 DWORD flags = info.compCompHnd->getClassAttribs(pResolvedToken->hClass);
9971 // If the class is final and is not marshal byref or
9972 // contextful, the jit can expand the IsInst check inline.
9973 DWORD flagsMask = CORINFO_FLG_FINAL | CORINFO_FLG_MARSHAL_BYREF | CORINFO_FLG_CONTEXTFUL;
9974 canExpandInline = ((flags & flagsMask) == CORINFO_FLG_FINAL);
9979 const bool expandInline = canExpandInline && shouldExpandInline;
9983 JITDUMP("\nExpanding %s as call because %s\n", isCastClass ? "castclass" : "isinst",
9984 canExpandInline ? "want smaller code or faster jitting" : "inline expansion not legal");
9986 // If we CSE this class handle we prevent assertionProp from making SubType assertions
9987 // so instead we force the CSE logic to not consider CSE-ing this class handle.
9989 op2->gtFlags |= GTF_DONT_CSE;
9991 return gtNewHelperCallNode(helper, TYP_REF, gtNewArgList(op2, op1));
9994 JITDUMP("\nExpanding %s inline\n", isCastClass ? "castclass" : "isinst");
9996 impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("bubbling QMark2"));
10001 // expand the methodtable match:
10003 // condMT ==> GT_NE
10005 // GT_IND op2 (typically CNS_INT)
10010 // This can replace op1 with a GT_COMMA that evaluates op1 into a local
10012 op1 = impCloneExpr(op1, &temp, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL, nullptr DEBUGARG("CASTCLASS eval op1"));
10014 // op1 is now known to be a non-complex tree
10015 // thus we can use gtClone(op1) from now on
10018 GenTree* op2Var = op2;
10021 op2Var = fgInsertCommaFormTemp(&op2);
10022 lvaTable[op2Var->AsLclVarCommon()->GetLclNum()].lvIsCSE = true;
10024 temp = gtNewOperNode(GT_IND, TYP_I_IMPL, temp);
10025 temp->gtFlags |= GTF_EXCEPT;
10026 condMT = gtNewOperNode(GT_NE, TYP_INT, temp, op2);
10030 // expand the null check:
10032 // condNull ==> GT_EQ
10037 condNull = gtNewOperNode(GT_EQ, TYP_INT, gtClone(op1), gtNewIconNode(0, TYP_REF));
10040 // expand the true and false trees for the condMT
10042 GenTree* condFalse = gtClone(op1);
10047 // use the special helper that skips the cases checked by our inlined cast
10049 const CorInfoHelpFunc specialHelper = CORINFO_HELP_CHKCASTCLASS_SPECIAL;
10051 condTrue = gtNewHelperCallNode(specialHelper, TYP_REF, gtNewArgList(op2Var, gtClone(op1)));
10055 condTrue = gtNewIconNode(0, TYP_REF);
10058 #define USE_QMARK_TREES
10060 #ifdef USE_QMARK_TREES
10063 // Generate first QMARK - COLON tree
10065 // qmarkMT ==> GT_QMARK
10069 // condFalse condTrue
10071 temp = new (this, GT_COLON) GenTreeColon(TYP_REF, condTrue, condFalse);
10072 qmarkMT = gtNewQmarkNode(TYP_REF, condMT, temp);
10073 condMT->gtFlags |= GTF_RELOP_QMARK;
10075 GenTree* qmarkNull;
10077 // Generate second QMARK - COLON tree
10079 // qmarkNull ==> GT_QMARK
10081 // condNull GT_COLON
10085 temp = new (this, GT_COLON) GenTreeColon(TYP_REF, gtClone(op1), qmarkMT);
10086 qmarkNull = gtNewQmarkNode(TYP_REF, condNull, temp);
10087 qmarkNull->gtFlags |= GTF_QMARK_CAST_INSTOF;
10088 condNull->gtFlags |= GTF_RELOP_QMARK;
10090 // Make QMark node a top level node by spilling it.
10091 unsigned tmp = lvaGrabTemp(true DEBUGARG("spilling QMark2"));
10092 impAssignTempGen(tmp, qmarkNull, (unsigned)CHECK_SPILL_NONE);
10094 // TODO: Is it possible op1 has a better type?
10095 lvaSetClass(tmp, pResolvedToken->hClass);
10096 return gtNewLclvNode(tmp, TYP_REF);
10101 #define assertImp(cond) ((void)0)
10103 #define assertImp(cond) \
10108 const int cchAssertImpBuf = 600; \
10109 char* assertImpBuf = (char*)alloca(cchAssertImpBuf); \
10110 _snprintf_s(assertImpBuf, cchAssertImpBuf, cchAssertImpBuf - 1, \
10111 "%s : Possibly bad IL with CEE_%s at offset %04Xh (op1=%s op2=%s stkDepth=%d)", #cond, \
10112 impCurOpcName, impCurOpcOffs, op1 ? varTypeName(op1->TypeGet()) : "NULL", \
10113 op2 ? varTypeName(op2->TypeGet()) : "NULL", verCurrentState.esStackDepth); \
10114 assertAbort(assertImpBuf, __FILE__, __LINE__); \
10120 #pragma warning(push)
10121 #pragma warning(disable : 21000) // Suppress PREFast warning about overly large function
10123 /*****************************************************************************
10124 * Import the instr for the given basic block
10126 void Compiler::impImportBlockCode(BasicBlock* block)
10128 #define _impResolveToken(kind) impResolveToken(codeAddr, &resolvedToken, kind)
10134 printf("\nImporting BB%02u (PC=%03u) of '%s'", block->bbNum, block->bbCodeOffs, info.compFullName);
10138 unsigned nxtStmtIndex = impInitBlockLineInfo();
10139 IL_OFFSET nxtStmtOffs;
10141 GenTree* arrayNodeFrom;
10142 GenTree* arrayNodeTo;
10143 GenTree* arrayNodeToIndex;
10144 CorInfoHelpFunc helper;
10145 CorInfoIsAccessAllowedResult accessAllowedResult;
10146 CORINFO_HELPER_DESC calloutHelper;
10147 const BYTE* lastLoadToken = nullptr;
10149 // reject cyclic constraints
10150 if (tiVerificationNeeded)
10152 Verify(!info.hasCircularClassConstraints, "Method parent has circular class type parameter constraints.");
10153 Verify(!info.hasCircularMethodConstraints, "Method has circular method type parameter constraints.");
10156 /* Get the tree list started */
10158 impBeginTreeList();
10160 /* Walk the opcodes that comprise the basic block */
10162 const BYTE* codeAddr = info.compCode + block->bbCodeOffs;
10163 const BYTE* codeEndp = info.compCode + block->bbCodeOffsEnd;
10165 IL_OFFSET opcodeOffs = block->bbCodeOffs;
10166 IL_OFFSET lastSpillOffs = opcodeOffs;
10170 /* remember the start of the delegate creation sequence (used for verification) */
10171 const BYTE* delegateCreateStart = nullptr;
10173 int prefixFlags = 0;
10174 bool explicitTailCall, constraintCall, readonlyCall;
10178 unsigned numArgs = info.compArgsCount;
10180 /* Now process all the opcodes in the block */
10182 var_types callTyp = TYP_COUNT;
10183 OPCODE prevOpcode = CEE_ILLEGAL;
10185 if (block->bbCatchTyp)
10187 if (info.compStmtOffsetsImplicit & ICorDebugInfo::CALL_SITE_BOUNDARIES)
10189 impCurStmtOffsSet(block->bbCodeOffs);
10192 // We will spill the GT_CATCH_ARG and the input of the BB_QMARK block
10193 // to a temp. This is a trade off for code simplicity
10194 impSpillSpecialSideEff();
10197 while (codeAddr < codeEndp)
10199 bool usingReadyToRunHelper = false;
10200 CORINFO_RESOLVED_TOKEN resolvedToken;
10201 CORINFO_RESOLVED_TOKEN constrainedResolvedToken;
10202 CORINFO_CALL_INFO callInfo;
10203 CORINFO_FIELD_INFO fieldInfo;
10205 tiRetVal = typeInfo(); // Default type info
10207 //---------------------------------------------------------------------
10209 /* We need to restrict the max tree depth as many of the Compiler
10210 functions are recursive. We do this by spilling the stack */
10212 if (verCurrentState.esStackDepth)
10214 /* Has it been a while since we last saw a non-empty stack (which
10215 guarantees that the tree depth isnt accumulating. */
10217 if ((opcodeOffs - lastSpillOffs) > MAX_TREE_SIZE && impCanSpillNow(prevOpcode))
10219 impSpillStackEnsure();
10220 lastSpillOffs = opcodeOffs;
10225 lastSpillOffs = opcodeOffs;
10226 impBoxTempInUse = false; // nothing on the stack, box temp OK to use again
10229 /* Compute the current instr offset */
10231 opcodeOffs = (IL_OFFSET)(codeAddr - info.compCode);
10234 if (opts.compDbgInfo)
10237 if (!compIsForInlining())
10240 (nxtStmtIndex < info.compStmtOffsetsCount) ? info.compStmtOffsets[nxtStmtIndex] : BAD_IL_OFFSET;
10242 /* Have we reached the next stmt boundary ? */
10244 if (nxtStmtOffs != BAD_IL_OFFSET && opcodeOffs >= nxtStmtOffs)
10246 assert(nxtStmtOffs == info.compStmtOffsets[nxtStmtIndex]);
10248 if (verCurrentState.esStackDepth != 0 && opts.compDbgCode)
10250 /* We need to provide accurate IP-mapping at this point.
10251 So spill anything on the stack so that it will form
10252 gtStmts with the correct stmt offset noted */
10254 impSpillStackEnsure(true);
10257 // Has impCurStmtOffs been reported in any tree?
10259 if (impCurStmtOffs != BAD_IL_OFFSET && opts.compDbgCode)
10261 GenTree* placeHolder = new (this, GT_NO_OP) GenTree(GT_NO_OP, TYP_VOID);
10262 impAppendTree(placeHolder, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
10264 assert(impCurStmtOffs == BAD_IL_OFFSET);
10267 if (impCurStmtOffs == BAD_IL_OFFSET)
10269 /* Make sure that nxtStmtIndex is in sync with opcodeOffs.
10270 If opcodeOffs has gone past nxtStmtIndex, catch up */
10272 while ((nxtStmtIndex + 1) < info.compStmtOffsetsCount &&
10273 info.compStmtOffsets[nxtStmtIndex + 1] <= opcodeOffs)
10278 /* Go to the new stmt */
10280 impCurStmtOffsSet(info.compStmtOffsets[nxtStmtIndex]);
10282 /* Update the stmt boundary index */
10285 assert(nxtStmtIndex <= info.compStmtOffsetsCount);
10287 /* Are there any more line# entries after this one? */
10289 if (nxtStmtIndex < info.compStmtOffsetsCount)
10291 /* Remember where the next line# starts */
10293 nxtStmtOffs = info.compStmtOffsets[nxtStmtIndex];
10297 /* No more line# entries */
10299 nxtStmtOffs = BAD_IL_OFFSET;
10303 else if ((info.compStmtOffsetsImplicit & ICorDebugInfo::STACK_EMPTY_BOUNDARIES) &&
10304 (verCurrentState.esStackDepth == 0))
10306 /* At stack-empty locations, we have already added the tree to
10307 the stmt list with the last offset. We just need to update
10311 impCurStmtOffsSet(opcodeOffs);
10313 else if ((info.compStmtOffsetsImplicit & ICorDebugInfo::CALL_SITE_BOUNDARIES) &&
10314 impOpcodeIsCallSiteBoundary(prevOpcode))
10316 /* Make sure we have a type cached */
10317 assert(callTyp != TYP_COUNT);
10319 if (callTyp == TYP_VOID)
10321 impCurStmtOffsSet(opcodeOffs);
10323 else if (opts.compDbgCode)
10325 impSpillStackEnsure(true);
10326 impCurStmtOffsSet(opcodeOffs);
10329 else if ((info.compStmtOffsetsImplicit & ICorDebugInfo::NOP_BOUNDARIES) && (prevOpcode == CEE_NOP))
10331 if (opts.compDbgCode)
10333 impSpillStackEnsure(true);
10336 impCurStmtOffsSet(opcodeOffs);
10339 assert(impCurStmtOffs == BAD_IL_OFFSET || nxtStmtOffs == BAD_IL_OFFSET ||
10340 jitGetILoffs(impCurStmtOffs) <= nxtStmtOffs);
10344 CORINFO_CLASS_HANDLE clsHnd = DUMMY_INIT(NULL);
10345 CORINFO_CLASS_HANDLE ldelemClsHnd = DUMMY_INIT(NULL);
10346 CORINFO_CLASS_HANDLE stelemClsHnd = DUMMY_INIT(NULL);
10348 var_types lclTyp, ovflType = TYP_UNKNOWN;
10349 GenTree* op1 = DUMMY_INIT(NULL);
10350 GenTree* op2 = DUMMY_INIT(NULL);
10351 GenTreeArgList* args = nullptr; // What good do these "DUMMY_INIT"s do?
10352 GenTree* newObjThisPtr = DUMMY_INIT(NULL);
10353 bool uns = DUMMY_INIT(false);
10354 bool isLocal = false;
10356 /* Get the next opcode and the size of its parameters */
10358 OPCODE opcode = (OPCODE)getU1LittleEndian(codeAddr);
10359 codeAddr += sizeof(__int8);
10362 impCurOpcOffs = (IL_OFFSET)(codeAddr - info.compCode - 1);
10363 JITDUMP("\n [%2u] %3u (0x%03x) ", verCurrentState.esStackDepth, impCurOpcOffs, impCurOpcOffs);
10368 // Return if any previous code has caused inline to fail.
10369 if (compDonotInline())
10374 /* Get the size of additional parameters */
10376 signed int sz = opcodeSizes[opcode];
10379 clsHnd = NO_CLASS_HANDLE;
10380 lclTyp = TYP_COUNT;
10381 callTyp = TYP_COUNT;
10383 impCurOpcOffs = (IL_OFFSET)(codeAddr - info.compCode - 1);
10384 impCurOpcName = opcodeNames[opcode];
10386 if (verbose && (opcode != CEE_PREFIX1))
10388 printf("%s", impCurOpcName);
10391 /* Use assertImp() to display the opcode */
10393 op1 = op2 = nullptr;
10396 /* See what kind of an opcode we have, then */
10398 unsigned mflags = 0;
10399 unsigned clsFlags = 0;
10412 CORINFO_SIG_INFO sig;
10414 bool ovfl, unordered, callNode;
10416 CORINFO_CLASS_HANDLE tokenType;
10426 opcode = (OPCODE)(getU1LittleEndian(codeAddr) + 256);
10427 opcodeOffs = (IL_OFFSET)(codeAddr - info.compCode);
10428 codeAddr += sizeof(__int8);
10429 goto DECODE_OPCODE;
10433 // We need to call impSpillLclRefs() for a struct type lclVar.
10434 // This is done for non-block assignments in the handling of stloc.
10435 if ((op1->OperGet() == GT_ASG) && varTypeIsStruct(op1->gtOp.gtOp1) &&
10436 (op1->gtOp.gtOp1->gtOper == GT_LCL_VAR))
10438 impSpillLclRefs(op1->gtOp.gtOp1->AsLclVarCommon()->gtLclNum);
10441 /* Append 'op1' to the list of statements */
10442 impAppendTree(op1, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
10447 /* Append 'op1' to the list of statements */
10449 impAppendTree(op1, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
10455 // Remember at which BC offset the tree was finished
10456 impNoteLastILoffs();
10461 impPushNullObjRefOnStack();
10464 case CEE_LDC_I4_M1:
10474 cval.intVal = (opcode - CEE_LDC_I4_0);
10475 assert(-1 <= cval.intVal && cval.intVal <= 8);
10479 cval.intVal = getI1LittleEndian(codeAddr);
10482 cval.intVal = getI4LittleEndian(codeAddr);
10485 JITDUMP(" %d", cval.intVal);
10486 impPushOnStack(gtNewIconNode(cval.intVal), typeInfo(TI_INT));
10490 cval.lngVal = getI8LittleEndian(codeAddr);
10491 JITDUMP(" 0x%016llx", cval.lngVal);
10492 impPushOnStack(gtNewLconNode(cval.lngVal), typeInfo(TI_LONG));
10496 cval.dblVal = getR8LittleEndian(codeAddr);
10497 JITDUMP(" %#.17g", cval.dblVal);
10498 impPushOnStack(gtNewDconNode(cval.dblVal), typeInfo(TI_DOUBLE));
10502 cval.dblVal = getR4LittleEndian(codeAddr);
10503 JITDUMP(" %#.17g", cval.dblVal);
10505 GenTree* cnsOp = gtNewDconNode(cval.dblVal);
10506 #if !FEATURE_X87_DOUBLES
10507 // X87 stack doesn't differentiate between float/double
10508 // so R4 is treated as R8, but everybody else does
10509 cnsOp->gtType = TYP_FLOAT;
10510 #endif // FEATURE_X87_DOUBLES
10511 impPushOnStack(cnsOp, typeInfo(TI_DOUBLE));
10517 if (compIsForInlining())
10519 if (impInlineInfo->inlineCandidateInfo->dwRestrictions & INLINE_NO_CALLEE_LDSTR)
10521 compInlineResult->NoteFatal(InlineObservation::CALLSITE_HAS_LDSTR_RESTRICTION);
10526 val = getU4LittleEndian(codeAddr);
10527 JITDUMP(" %08X", val);
10528 if (tiVerificationNeeded)
10530 Verify(info.compCompHnd->isValidStringRef(info.compScopeHnd, val), "bad string");
10531 tiRetVal = typeInfo(TI_REF, impGetStringClass());
10533 impPushOnStack(gtNewSconNode(val, info.compScopeHnd), tiRetVal);
10538 lclNum = getU2LittleEndian(codeAddr);
10539 JITDUMP(" %u", lclNum);
10540 impLoadArg(lclNum, opcodeOffs + sz + 1);
10544 lclNum = getU1LittleEndian(codeAddr);
10545 JITDUMP(" %u", lclNum);
10546 impLoadArg(lclNum, opcodeOffs + sz + 1);
10553 lclNum = (opcode - CEE_LDARG_0);
10554 assert(lclNum >= 0 && lclNum < 4);
10555 impLoadArg(lclNum, opcodeOffs + sz + 1);
10559 lclNum = getU2LittleEndian(codeAddr);
10560 JITDUMP(" %u", lclNum);
10561 impLoadLoc(lclNum, opcodeOffs + sz + 1);
10565 lclNum = getU1LittleEndian(codeAddr);
10566 JITDUMP(" %u", lclNum);
10567 impLoadLoc(lclNum, opcodeOffs + sz + 1);
10574 lclNum = (opcode - CEE_LDLOC_0);
10575 assert(lclNum >= 0 && lclNum < 4);
10576 impLoadLoc(lclNum, opcodeOffs + sz + 1);
10580 lclNum = getU2LittleEndian(codeAddr);
10584 lclNum = getU1LittleEndian(codeAddr);
10586 JITDUMP(" %u", lclNum);
10588 if (tiVerificationNeeded)
10590 Verify(lclNum < info.compILargsCount, "bad arg num");
10593 if (compIsForInlining())
10595 op1 = impInlineFetchArg(lclNum, impInlineInfo->inlArgInfo, impInlineInfo->lclVarInfo);
10596 noway_assert(op1->gtOper == GT_LCL_VAR);
10597 lclNum = op1->AsLclVar()->gtLclNum;
10602 lclNum = compMapILargNum(lclNum); // account for possible hidden param
10603 assertImp(lclNum < numArgs);
10605 if (lclNum == info.compThisArg)
10607 lclNum = lvaArg0Var;
10610 // We should have seen this arg write in the prescan
10611 assert(lvaTable[lclNum].lvHasILStoreOp);
10613 if (tiVerificationNeeded)
10615 typeInfo& tiLclVar = lvaTable[lclNum].lvVerTypeInfo;
10616 Verify(tiCompatibleWith(impStackTop().seTypeInfo, NormaliseForStack(tiLclVar), true),
10619 if (verTrackObjCtorInitState && (verCurrentState.thisInitialized != TIS_Init))
10621 Verify(!tiLclVar.IsThisPtr(), "storing to uninit this ptr");
10628 lclNum = getU2LittleEndian(codeAddr);
10630 JITDUMP(" %u", lclNum);
10634 lclNum = getU1LittleEndian(codeAddr);
10636 JITDUMP(" %u", lclNum);
10644 lclNum = (opcode - CEE_STLOC_0);
10645 assert(lclNum >= 0 && lclNum < 4);
10648 if (tiVerificationNeeded)
10650 Verify(lclNum < info.compMethodInfo->locals.numArgs, "bad local num");
10651 Verify(tiCompatibleWith(impStackTop().seTypeInfo,
10652 NormaliseForStack(lvaTable[lclNum + numArgs].lvVerTypeInfo), true),
10656 if (compIsForInlining())
10658 lclTyp = impInlineInfo->lclVarInfo[lclNum + impInlineInfo->argCnt].lclTypeInfo;
10660 /* Have we allocated a temp for this local? */
10662 lclNum = impInlineFetchLocal(lclNum DEBUGARG("Inline stloc first use temp"));
10671 if (lclNum >= info.compLocalsCount && lclNum != lvaArg0Var)
10673 assert(!tiVerificationNeeded); // We should have thrown the VerificationException before.
10679 /* if it is a struct assignment, make certain we don't overflow the buffer */
10680 assert(lclTyp != TYP_STRUCT || lvaLclSize(lclNum) >= info.compCompHnd->getClassSize(clsHnd));
10682 if (lvaTable[lclNum].lvNormalizeOnLoad())
10684 lclTyp = lvaGetRealType(lclNum);
10688 lclTyp = lvaGetActualType(lclNum);
10692 /* Pop the value being assigned */
10695 StackEntry se = impPopStack();
10696 clsHnd = se.seTypeInfo.GetClassHandle();
10698 tiRetVal = se.seTypeInfo;
10701 #ifdef FEATURE_SIMD
10702 if (varTypeIsSIMD(lclTyp) && (lclTyp != op1->TypeGet()))
10704 assert(op1->TypeGet() == TYP_STRUCT);
10705 op1->gtType = lclTyp;
10707 #endif // FEATURE_SIMD
10709 op1 = impImplicitIorI4Cast(op1, lclTyp);
10711 #ifdef _TARGET_64BIT_
10712 // Downcast the TYP_I_IMPL into a 32-bit Int for x86 JIT compatiblity
10713 if (varTypeIsI(op1->TypeGet()) && (genActualType(lclTyp) == TYP_INT))
10715 assert(!tiVerificationNeeded); // We should have thrown the VerificationException before.
10716 op1 = gtNewCastNode(TYP_INT, op1, false, TYP_INT);
10718 #endif // _TARGET_64BIT_
10720 // We had better assign it a value of the correct type
10722 genActualType(lclTyp) == genActualType(op1->gtType) ||
10723 genActualType(lclTyp) == TYP_I_IMPL && op1->IsVarAddr() ||
10724 (genActualType(lclTyp) == TYP_I_IMPL && (op1->gtType == TYP_BYREF || op1->gtType == TYP_REF)) ||
10725 (genActualType(op1->gtType) == TYP_I_IMPL && lclTyp == TYP_BYREF) ||
10726 (varTypeIsFloating(lclTyp) && varTypeIsFloating(op1->TypeGet())) ||
10727 ((genActualType(lclTyp) == TYP_BYREF) && genActualType(op1->TypeGet()) == TYP_REF));
10729 /* If op1 is "&var" then its type is the transient "*" and it can
10730 be used either as TYP_BYREF or TYP_I_IMPL */
10732 if (op1->IsVarAddr())
10734 assertImp(genActualType(lclTyp) == TYP_I_IMPL || lclTyp == TYP_BYREF);
10736 /* When "&var" is created, we assume it is a byref. If it is
10737 being assigned to a TYP_I_IMPL var, change the type to
10738 prevent unnecessary GC info */
10740 if (genActualType(lclTyp) == TYP_I_IMPL)
10742 op1->gtType = TYP_I_IMPL;
10746 // If this is a local and the local is a ref type, see
10747 // if we can improve type information based on the
10748 // value being assigned.
10749 if (isLocal && (lclTyp == TYP_REF))
10751 // We should have seen a stloc in our IL prescan.
10752 assert(lvaTable[lclNum].lvHasILStoreOp);
10754 const bool isSingleILStoreLocal =
10755 !lvaTable[lclNum].lvHasMultipleILStoreOp && !lvaTable[lclNum].lvHasLdAddrOp;
10757 // Conservative check that there is just one
10758 // definition that reaches this store.
10759 const bool hasSingleReachingDef = (block->bbStackDepthOnEntry() == 0);
10761 if (isSingleILStoreLocal && hasSingleReachingDef)
10763 lvaUpdateClass(lclNum, op1, clsHnd);
10767 /* Filter out simple assignments to itself */
10769 if (op1->gtOper == GT_LCL_VAR && lclNum == op1->gtLclVarCommon.gtLclNum)
10771 if (opts.compDbgCode)
10773 op1 = gtNewNothingNode();
10782 /* Create the assignment node */
10784 op2 = gtNewLclvNode(lclNum, lclTyp, opcodeOffs + sz + 1);
10786 /* If the local is aliased or pinned, we need to spill calls and
10787 indirections from the stack. */
10789 if ((lvaTable[lclNum].lvAddrExposed || lvaTable[lclNum].lvHasLdAddrOp || lvaTable[lclNum].lvPinned) &&
10790 (verCurrentState.esStackDepth > 0))
10792 impSpillSideEffects(false,
10793 (unsigned)CHECK_SPILL_ALL DEBUGARG("Local could be aliased or is pinned"));
10796 /* Spill any refs to the local from the stack */
10798 impSpillLclRefs(lclNum);
10800 #if !FEATURE_X87_DOUBLES
10801 // We can generate an assignment to a TYP_FLOAT from a TYP_DOUBLE
10802 // We insert a cast to the dest 'op2' type
10804 if ((op1->TypeGet() != op2->TypeGet()) && varTypeIsFloating(op1->gtType) &&
10805 varTypeIsFloating(op2->gtType))
10807 op1 = gtNewCastNode(op2->TypeGet(), op1, false, op2->TypeGet());
10809 #endif // !FEATURE_X87_DOUBLES
10811 if (varTypeIsStruct(lclTyp))
10813 op1 = impAssignStruct(op2, op1, clsHnd, (unsigned)CHECK_SPILL_ALL);
10817 // The code generator generates GC tracking information
10818 // based on the RHS of the assignment. Later the LHS (which is
10819 // is a BYREF) gets used and the emitter checks that that variable
10820 // is being tracked. It is not (since the RHS was an int and did
10821 // not need tracking). To keep this assert happy, we change the RHS
10822 if (lclTyp == TYP_BYREF && !varTypeIsGC(op1->gtType))
10824 op1->gtType = TYP_BYREF;
10826 op1 = gtNewAssignNode(op2, op1);
10832 lclNum = getU2LittleEndian(codeAddr);
10836 lclNum = getU1LittleEndian(codeAddr);
10838 JITDUMP(" %u", lclNum);
10839 if (tiVerificationNeeded)
10841 Verify(lclNum < info.compMethodInfo->locals.numArgs, "bad local num");
10842 Verify(info.compInitMem, "initLocals not set");
10845 if (compIsForInlining())
10847 // Get the local type
10848 lclTyp = impInlineInfo->lclVarInfo[lclNum + impInlineInfo->argCnt].lclTypeInfo;
10850 /* Have we allocated a temp for this local? */
10852 lclNum = impInlineFetchLocal(lclNum DEBUGARG("Inline ldloca(s) first use temp"));
10854 op1 = gtNewLclvNode(lclNum, lvaGetActualType(lclNum));
10860 assertImp(lclNum < info.compLocalsCount);
10864 lclNum = getU2LittleEndian(codeAddr);
10868 lclNum = getU1LittleEndian(codeAddr);
10870 JITDUMP(" %u", lclNum);
10871 Verify(lclNum < info.compILargsCount, "bad arg num");
10873 if (compIsForInlining())
10875 // In IL, LDARGA(_S) is used to load the byref managed pointer of struct argument,
10876 // followed by a ldfld to load the field.
10878 op1 = impInlineFetchArg(lclNum, impInlineInfo->inlArgInfo, impInlineInfo->lclVarInfo);
10879 if (op1->gtOper != GT_LCL_VAR)
10881 compInlineResult->NoteFatal(InlineObservation::CALLSITE_LDARGA_NOT_LOCAL_VAR);
10885 assert(op1->gtOper == GT_LCL_VAR);
10890 lclNum = compMapILargNum(lclNum); // account for possible hidden param
10891 assertImp(lclNum < numArgs);
10893 if (lclNum == info.compThisArg)
10895 lclNum = lvaArg0Var;
10902 op1 = gtNewLclvNode(lclNum, lvaGetActualType(lclNum), opcodeOffs + sz + 1);
10905 assert(op1->gtOper == GT_LCL_VAR);
10907 /* Note that this is supposed to create the transient type "*"
10908 which may be used as a TYP_I_IMPL. However we catch places
10909 where it is used as a TYP_I_IMPL and change the node if needed.
10910 Thus we are pessimistic and may report byrefs in the GC info
10911 where it was not absolutely needed, but it is safer this way.
10913 op1 = gtNewOperNode(GT_ADDR, TYP_BYREF, op1);
10915 // &aliasedVar doesnt need GTF_GLOB_REF, though alisasedVar does
10916 assert((op1->gtFlags & GTF_GLOB_REF) == 0);
10918 tiRetVal = lvaTable[lclNum].lvVerTypeInfo;
10919 if (tiVerificationNeeded)
10921 // Don't allow taking address of uninit this ptr.
10922 if (verTrackObjCtorInitState && (verCurrentState.thisInitialized != TIS_Init))
10924 Verify(!tiRetVal.IsThisPtr(), "address of uninit this ptr");
10927 if (!tiRetVal.IsByRef())
10929 tiRetVal.MakeByRef();
10933 Verify(false, "byref to byref");
10937 impPushOnStack(op1, tiRetVal);
10942 if (!info.compIsVarArgs)
10944 BADCODE("arglist in non-vararg method");
10947 if (tiVerificationNeeded)
10949 tiRetVal = typeInfo(TI_STRUCT, impGetRuntimeArgumentHandle());
10951 assertImp((info.compMethodInfo->args.callConv & CORINFO_CALLCONV_MASK) == CORINFO_CALLCONV_VARARG);
10953 /* The ARGLIST cookie is a hidden 'last' parameter, we have already
10954 adjusted the arg count cos this is like fetching the last param */
10955 assertImp(0 < numArgs);
10956 assert(lvaTable[lvaVarargsHandleArg].lvAddrExposed);
10957 lclNum = lvaVarargsHandleArg;
10958 op1 = gtNewLclvNode(lclNum, TYP_I_IMPL, opcodeOffs + sz + 1);
10959 op1 = gtNewOperNode(GT_ADDR, TYP_BYREF, op1);
10960 impPushOnStack(op1, tiRetVal);
10963 case CEE_ENDFINALLY:
10965 if (compIsForInlining())
10967 assert(!"Shouldn't have exception handlers in the inliner!");
10968 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_ENDFINALLY);
10972 if (verCurrentState.esStackDepth > 0)
10974 impEvalSideEffects();
10977 if (info.compXcptnsCount == 0)
10979 BADCODE("endfinally outside finally");
10982 assert(verCurrentState.esStackDepth == 0);
10984 op1 = gtNewOperNode(GT_RETFILT, TYP_VOID, nullptr);
10987 case CEE_ENDFILTER:
10989 if (compIsForInlining())
10991 assert(!"Shouldn't have exception handlers in the inliner!");
10992 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_ENDFILTER);
10996 block->bbSetRunRarely(); // filters are rare
10998 if (info.compXcptnsCount == 0)
11000 BADCODE("endfilter outside filter");
11003 if (tiVerificationNeeded)
11005 Verify(impStackTop().seTypeInfo.IsType(TI_INT), "bad endfilt arg");
11008 op1 = impPopStack().val;
11009 assertImp(op1->gtType == TYP_INT);
11010 if (!bbInFilterILRange(block))
11012 BADCODE("EndFilter outside a filter handler");
11015 /* Mark current bb as end of filter */
11017 assert(compCurBB->bbFlags & BBF_DONT_REMOVE);
11018 assert(compCurBB->bbJumpKind == BBJ_EHFILTERRET);
11020 /* Mark catch handler as successor */
11022 op1 = gtNewOperNode(GT_RETFILT, op1->TypeGet(), op1);
11023 if (verCurrentState.esStackDepth != 0)
11025 verRaiseVerifyException(INDEBUG("stack must be 1 on end of filter") DEBUGARG(__FILE__)
11026 DEBUGARG(__LINE__));
11031 prefixFlags &= ~PREFIX_TAILCALL; // ret without call before it
11033 if (!impReturnInstruction(block, prefixFlags, opcode))
11044 assert(!compIsForInlining());
11046 if (tiVerificationNeeded)
11048 Verify(false, "Invalid opcode: CEE_JMP");
11051 if ((info.compFlags & CORINFO_FLG_SYNCH) || block->hasTryIndex() || block->hasHndIndex())
11053 /* CEE_JMP does not make sense in some "protected" regions. */
11055 BADCODE("Jmp not allowed in protected region");
11058 if (verCurrentState.esStackDepth != 0)
11060 BADCODE("Stack must be empty after CEE_JMPs");
11063 _impResolveToken(CORINFO_TOKENKIND_Method);
11065 JITDUMP(" %08X", resolvedToken.token);
11067 /* The signature of the target has to be identical to ours.
11068 At least check that argCnt and returnType match */
11070 eeGetMethodSig(resolvedToken.hMethod, &sig);
11071 if (sig.numArgs != info.compMethodInfo->args.numArgs ||
11072 sig.retType != info.compMethodInfo->args.retType ||
11073 sig.callConv != info.compMethodInfo->args.callConv)
11075 BADCODE("Incompatible target for CEE_JMPs");
11078 op1 = new (this, GT_JMP) GenTreeVal(GT_JMP, TYP_VOID, (size_t)resolvedToken.hMethod);
11080 /* Mark the basic block as being a JUMP instead of RETURN */
11082 block->bbFlags |= BBF_HAS_JMP;
11084 /* Set this flag to make sure register arguments have a location assigned
11085 * even if we don't use them inside the method */
11087 compJmpOpUsed = true;
11089 fgNoStructPromotion = true;
11094 assertImp(sz == sizeof(unsigned));
11096 _impResolveToken(CORINFO_TOKENKIND_Class);
11098 JITDUMP(" %08X", resolvedToken.token);
11100 ldelemClsHnd = resolvedToken.hClass;
11102 if (tiVerificationNeeded)
11104 typeInfo tiArray = impStackTop(1).seTypeInfo;
11105 typeInfo tiIndex = impStackTop().seTypeInfo;
11107 // As per ECMA 'index' specified can be either int32 or native int.
11108 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11110 typeInfo arrayElemType = verMakeTypeInfo(ldelemClsHnd);
11111 Verify(tiArray.IsNullObjRef() ||
11112 typeInfo::AreEquivalent(verGetArrayElemType(tiArray), arrayElemType),
11115 tiRetVal = arrayElemType;
11116 tiRetVal.MakeByRef();
11117 if (prefixFlags & PREFIX_READONLY)
11119 tiRetVal.SetIsReadonlyByRef();
11122 // an array interior pointer is always in the heap
11123 tiRetVal.SetIsPermanentHomeByRef();
11126 // If it's a value class array we just do a simple address-of
11127 if (eeIsValueClass(ldelemClsHnd))
11129 CorInfoType cit = info.compCompHnd->getTypeForPrimitiveValueClass(ldelemClsHnd);
11130 if (cit == CORINFO_TYPE_UNDEF)
11132 lclTyp = TYP_STRUCT;
11136 lclTyp = JITtype2varType(cit);
11138 goto ARR_LD_POST_VERIFY;
11141 // Similarly, if its a readonly access, we can do a simple address-of
11142 // without doing a runtime type-check
11143 if (prefixFlags & PREFIX_READONLY)
11146 goto ARR_LD_POST_VERIFY;
11149 // Otherwise we need the full helper function with run-time type check
11150 op1 = impTokenToHandle(&resolvedToken);
11151 if (op1 == nullptr)
11152 { // compDonotInline()
11156 args = gtNewArgList(op1); // Type
11157 args = gtNewListNode(impPopStack().val, args); // index
11158 args = gtNewListNode(impPopStack().val, args); // array
11159 op1 = gtNewHelperCallNode(CORINFO_HELP_LDELEMA_REF, TYP_BYREF, args);
11161 impPushOnStack(op1, tiRetVal);
11164 // ldelem for reference and value types
11166 assertImp(sz == sizeof(unsigned));
11168 _impResolveToken(CORINFO_TOKENKIND_Class);
11170 JITDUMP(" %08X", resolvedToken.token);
11172 ldelemClsHnd = resolvedToken.hClass;
11174 if (tiVerificationNeeded)
11176 typeInfo tiArray = impStackTop(1).seTypeInfo;
11177 typeInfo tiIndex = impStackTop().seTypeInfo;
11179 // As per ECMA 'index' specified can be either int32 or native int.
11180 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11181 tiRetVal = verMakeTypeInfo(ldelemClsHnd);
11183 Verify(tiArray.IsNullObjRef() || tiCompatibleWith(verGetArrayElemType(tiArray), tiRetVal, false),
11184 "type of array incompatible with type operand");
11185 tiRetVal.NormaliseForStack();
11188 // If it's a reference type or generic variable type
11189 // then just generate code as though it's a ldelem.ref instruction
11190 if (!eeIsValueClass(ldelemClsHnd))
11193 opcode = CEE_LDELEM_REF;
11197 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(ldelemClsHnd);
11198 lclTyp = JITtype2varType(jitTyp);
11199 tiRetVal = verMakeTypeInfo(ldelemClsHnd); // precise type always needed for struct
11200 tiRetVal.NormaliseForStack();
11202 goto ARR_LD_POST_VERIFY;
11204 case CEE_LDELEM_I1:
11207 case CEE_LDELEM_I2:
11208 lclTyp = TYP_SHORT;
11211 lclTyp = TYP_I_IMPL;
11214 // Should be UINT, but since no platform widens 4->8 bytes it doesn't matter
11215 // and treating it as TYP_INT avoids other asserts.
11216 case CEE_LDELEM_U4:
11220 case CEE_LDELEM_I4:
11223 case CEE_LDELEM_I8:
11226 case CEE_LDELEM_REF:
11229 case CEE_LDELEM_R4:
11230 lclTyp = TYP_FLOAT;
11232 case CEE_LDELEM_R8:
11233 lclTyp = TYP_DOUBLE;
11235 case CEE_LDELEM_U1:
11236 lclTyp = TYP_UBYTE;
11238 case CEE_LDELEM_U2:
11239 lclTyp = TYP_USHORT;
11244 if (tiVerificationNeeded)
11246 typeInfo tiArray = impStackTop(1).seTypeInfo;
11247 typeInfo tiIndex = impStackTop().seTypeInfo;
11249 // As per ECMA 'index' specified can be either int32 or native int.
11250 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11251 if (tiArray.IsNullObjRef())
11253 if (lclTyp == TYP_REF)
11254 { // we will say a deref of a null array yields a null ref
11255 tiRetVal = typeInfo(TI_NULL);
11259 tiRetVal = typeInfo(lclTyp);
11264 tiRetVal = verGetArrayElemType(tiArray);
11265 typeInfo arrayElemTi = typeInfo(lclTyp);
11266 #ifdef _TARGET_64BIT_
11267 if (opcode == CEE_LDELEM_I)
11269 arrayElemTi = typeInfo::nativeInt();
11272 if (lclTyp != TYP_REF && lclTyp != TYP_STRUCT)
11274 Verify(typeInfo::AreEquivalent(tiRetVal, arrayElemTi), "bad array");
11277 #endif // _TARGET_64BIT_
11279 Verify(tiRetVal.IsType(arrayElemTi.GetType()), "bad array");
11282 tiRetVal.NormaliseForStack();
11284 ARR_LD_POST_VERIFY:
11286 /* Pull the index value and array address */
11287 op2 = impPopStack().val;
11288 op1 = impPopStack().val;
11289 assertImp(op1->gtType == TYP_REF);
11291 /* Check for null pointer - in the inliner case we simply abort */
11293 if (compIsForInlining())
11295 if (op1->gtOper == GT_CNS_INT)
11297 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_NULL_FOR_LDELEM);
11302 op1 = impCheckForNullPointer(op1);
11304 /* Mark the block as containing an index expression */
11306 if (op1->gtOper == GT_LCL_VAR)
11308 if (op2->gtOper == GT_LCL_VAR || op2->gtOper == GT_CNS_INT || op2->gtOper == GT_ADD)
11310 block->bbFlags |= BBF_HAS_IDX_LEN;
11311 optMethodFlags |= OMF_HAS_ARRAYREF;
11315 /* Create the index node and push it on the stack */
11317 op1 = gtNewIndexRef(lclTyp, op1, op2);
11319 ldstruct = (opcode == CEE_LDELEM && lclTyp == TYP_STRUCT);
11321 if ((opcode == CEE_LDELEMA) || ldstruct ||
11322 (ldelemClsHnd != DUMMY_INIT(NULL) && eeIsValueClass(ldelemClsHnd)))
11324 assert(ldelemClsHnd != DUMMY_INIT(NULL));
11326 // remember the element size
11327 if (lclTyp == TYP_REF)
11329 op1->gtIndex.gtIndElemSize = TARGET_POINTER_SIZE;
11333 // If ldElemClass is precisely a primitive type, use that, otherwise, preserve the struct type.
11334 if (info.compCompHnd->getTypeForPrimitiveValueClass(ldelemClsHnd) == CORINFO_TYPE_UNDEF)
11336 op1->gtIndex.gtStructElemClass = ldelemClsHnd;
11338 assert(lclTyp != TYP_STRUCT || op1->gtIndex.gtStructElemClass != nullptr);
11339 if (lclTyp == TYP_STRUCT)
11341 size = info.compCompHnd->getClassSize(ldelemClsHnd);
11342 op1->gtIndex.gtIndElemSize = size;
11343 op1->gtType = lclTyp;
11347 if ((opcode == CEE_LDELEMA) || ldstruct)
11350 lclTyp = TYP_BYREF;
11352 op1 = gtNewOperNode(GT_ADDR, lclTyp, op1);
11356 assert(lclTyp != TYP_STRUCT);
11362 // Create an OBJ for the result
11363 op1 = gtNewObjNode(ldelemClsHnd, op1);
11364 op1->gtFlags |= GTF_EXCEPT;
11366 impPushOnStack(op1, tiRetVal);
11369 // stelem for reference and value types
11372 assertImp(sz == sizeof(unsigned));
11374 _impResolveToken(CORINFO_TOKENKIND_Class);
11376 JITDUMP(" %08X", resolvedToken.token);
11378 stelemClsHnd = resolvedToken.hClass;
11380 if (tiVerificationNeeded)
11382 typeInfo tiArray = impStackTop(2).seTypeInfo;
11383 typeInfo tiIndex = impStackTop(1).seTypeInfo;
11384 typeInfo tiValue = impStackTop().seTypeInfo;
11386 // As per ECMA 'index' specified can be either int32 or native int.
11387 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11388 typeInfo arrayElem = verMakeTypeInfo(stelemClsHnd);
11390 Verify(tiArray.IsNullObjRef() || tiCompatibleWith(arrayElem, verGetArrayElemType(tiArray), false),
11391 "type operand incompatible with array element type");
11392 arrayElem.NormaliseForStack();
11393 Verify(tiCompatibleWith(tiValue, arrayElem, true), "value incompatible with type operand");
11396 // If it's a reference type just behave as though it's a stelem.ref instruction
11397 if (!eeIsValueClass(stelemClsHnd))
11399 goto STELEM_REF_POST_VERIFY;
11402 // Otherwise extract the type
11404 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(stelemClsHnd);
11405 lclTyp = JITtype2varType(jitTyp);
11406 goto ARR_ST_POST_VERIFY;
11409 case CEE_STELEM_REF:
11411 if (tiVerificationNeeded)
11413 typeInfo tiArray = impStackTop(2).seTypeInfo;
11414 typeInfo tiIndex = impStackTop(1).seTypeInfo;
11415 typeInfo tiValue = impStackTop().seTypeInfo;
11417 // As per ECMA 'index' specified can be either int32 or native int.
11418 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11419 Verify(tiValue.IsObjRef(), "bad value");
11421 // we only check that it is an object referece, The helper does additional checks
11422 Verify(tiArray.IsNullObjRef() || verGetArrayElemType(tiArray).IsType(TI_REF), "bad array");
11425 STELEM_REF_POST_VERIFY:
11427 arrayNodeTo = impStackTop(2).val;
11428 arrayNodeToIndex = impStackTop(1).val;
11429 arrayNodeFrom = impStackTop().val;
11432 // Note that it is not legal to optimize away CORINFO_HELP_ARRADDR_ST in a
11433 // lot of cases because of covariance. ie. foo[] can be cast to object[].
11436 // Check for assignment to same array, ie. arrLcl[i] = arrLcl[j]
11437 // This does not need CORINFO_HELP_ARRADDR_ST
11438 if (arrayNodeFrom->OperGet() == GT_INDEX && arrayNodeFrom->gtOp.gtOp1->gtOper == GT_LCL_VAR &&
11439 arrayNodeTo->gtOper == GT_LCL_VAR &&
11440 arrayNodeTo->gtLclVarCommon.gtLclNum == arrayNodeFrom->gtOp.gtOp1->gtLclVarCommon.gtLclNum &&
11441 !lvaTable[arrayNodeTo->gtLclVarCommon.gtLclNum].lvAddrExposed)
11443 JITDUMP("\nstelem of ref from same array: skipping covariant store check\n");
11445 goto ARR_ST_POST_VERIFY;
11448 // Check for assignment of NULL. This does not need CORINFO_HELP_ARRADDR_ST
11449 if (arrayNodeFrom->OperGet() == GT_CNS_INT)
11451 JITDUMP("\nstelem of null: skipping covariant store check\n");
11452 assert(arrayNodeFrom->gtType == TYP_REF && arrayNodeFrom->gtIntCon.gtIconVal == 0);
11454 goto ARR_ST_POST_VERIFY;
11457 /* Call a helper function to do the assignment */
11458 op1 = gtNewHelperCallNode(CORINFO_HELP_ARRADDR_ST, TYP_VOID, impPopList(3, nullptr));
11462 case CEE_STELEM_I1:
11465 case CEE_STELEM_I2:
11466 lclTyp = TYP_SHORT;
11469 lclTyp = TYP_I_IMPL;
11471 case CEE_STELEM_I4:
11474 case CEE_STELEM_I8:
11477 case CEE_STELEM_R4:
11478 lclTyp = TYP_FLOAT;
11480 case CEE_STELEM_R8:
11481 lclTyp = TYP_DOUBLE;
11486 if (tiVerificationNeeded)
11488 typeInfo tiArray = impStackTop(2).seTypeInfo;
11489 typeInfo tiIndex = impStackTop(1).seTypeInfo;
11490 typeInfo tiValue = impStackTop().seTypeInfo;
11492 // As per ECMA 'index' specified can be either int32 or native int.
11493 Verify(tiIndex.IsIntOrNativeIntType(), "bad index");
11494 typeInfo arrayElem = typeInfo(lclTyp);
11495 #ifdef _TARGET_64BIT_
11496 if (opcode == CEE_STELEM_I)
11498 arrayElem = typeInfo::nativeInt();
11500 #endif // _TARGET_64BIT_
11501 Verify(tiArray.IsNullObjRef() || typeInfo::AreEquivalent(verGetArrayElemType(tiArray), arrayElem),
11504 Verify(tiCompatibleWith(NormaliseForStack(tiValue), arrayElem.NormaliseForStack(), true),
11508 ARR_ST_POST_VERIFY:
11509 /* The strict order of evaluation is LHS-operands, RHS-operands,
11510 range-check, and then assignment. However, codegen currently
11511 does the range-check before evaluation the RHS-operands. So to
11512 maintain strict ordering, we spill the stack. */
11514 if (impStackTop().val->gtFlags & GTF_SIDE_EFFECT)
11516 impSpillSideEffects(false, (unsigned)CHECK_SPILL_ALL DEBUGARG(
11517 "Strict ordering of exceptions for Array store"));
11520 /* Pull the new value from the stack */
11521 op2 = impPopStack().val;
11523 /* Pull the index value */
11524 op1 = impPopStack().val;
11526 /* Pull the array address */
11527 op3 = impPopStack().val;
11529 assertImp(op3->gtType == TYP_REF);
11530 if (op2->IsVarAddr())
11532 op2->gtType = TYP_I_IMPL;
11535 op3 = impCheckForNullPointer(op3);
11537 // Mark the block as containing an index expression
11539 if (op3->gtOper == GT_LCL_VAR)
11541 if (op1->gtOper == GT_LCL_VAR || op1->gtOper == GT_CNS_INT || op1->gtOper == GT_ADD)
11543 block->bbFlags |= BBF_HAS_IDX_LEN;
11544 optMethodFlags |= OMF_HAS_ARRAYREF;
11548 /* Create the index node */
11550 op1 = gtNewIndexRef(lclTyp, op3, op1);
11552 /* Create the assignment node and append it */
11554 if (lclTyp == TYP_STRUCT)
11556 assert(stelemClsHnd != DUMMY_INIT(NULL));
11558 op1->gtIndex.gtStructElemClass = stelemClsHnd;
11559 op1->gtIndex.gtIndElemSize = info.compCompHnd->getClassSize(stelemClsHnd);
11561 if (varTypeIsStruct(op1))
11563 op1 = impAssignStruct(op1, op2, stelemClsHnd, (unsigned)CHECK_SPILL_ALL);
11567 op2 = impImplicitR4orR8Cast(op2, op1->TypeGet());
11568 op1 = gtNewAssignNode(op1, op2);
11571 /* Mark the expression as containing an assignment */
11573 op1->gtFlags |= GTF_ASG;
11584 case CEE_ADD_OVF_UN:
11592 goto MATH_OP2_FLAGS;
11601 case CEE_SUB_OVF_UN:
11609 goto MATH_OP2_FLAGS;
11613 goto MATH_MAYBE_CALL_NO_OVF;
11618 case CEE_MUL_OVF_UN:
11625 goto MATH_MAYBE_CALL_OVF;
11627 // Other binary math operations
11631 goto MATH_MAYBE_CALL_NO_OVF;
11635 goto MATH_MAYBE_CALL_NO_OVF;
11639 goto MATH_MAYBE_CALL_NO_OVF;
11643 goto MATH_MAYBE_CALL_NO_OVF;
11645 MATH_MAYBE_CALL_NO_OVF:
11647 MATH_MAYBE_CALL_OVF:
11648 // Morpher has some complex logic about when to turn different
11649 // typed nodes on different platforms into helper calls. We
11650 // need to either duplicate that logic here, or just
11651 // pessimistically make all the nodes large enough to become
11652 // call nodes. Since call nodes aren't that much larger and
11653 // these opcodes are infrequent enough I chose the latter.
11655 goto MATH_OP2_FLAGS;
11667 MATH_OP2: // For default values of 'ovfl' and 'callNode'
11672 MATH_OP2_FLAGS: // If 'ovfl' and 'callNode' have already been set
11674 /* Pull two values and push back the result */
11676 if (tiVerificationNeeded)
11678 const typeInfo& tiOp1 = impStackTop(1).seTypeInfo;
11679 const typeInfo& tiOp2 = impStackTop().seTypeInfo;
11681 Verify(tiCompatibleWith(tiOp1, tiOp2, true), "different arg type");
11682 if (oper == GT_ADD || oper == GT_DIV || oper == GT_SUB || oper == GT_MUL || oper == GT_MOD)
11684 Verify(tiOp1.IsNumberType(), "not number");
11688 Verify(tiOp1.IsIntegerType(), "not integer");
11691 Verify(!ovfl || tiOp1.IsIntegerType(), "not integer");
11695 #ifdef _TARGET_64BIT_
11696 if (tiOp2.IsNativeIntType())
11700 #endif // _TARGET_64BIT_
11703 op2 = impPopStack().val;
11704 op1 = impPopStack().val;
11706 #if !CPU_HAS_FP_SUPPORT
11707 if (varTypeIsFloating(op1->gtType))
11712 /* Can't do arithmetic with references */
11713 assertImp(genActualType(op1->TypeGet()) != TYP_REF && genActualType(op2->TypeGet()) != TYP_REF);
11715 // Change both to TYP_I_IMPL (impBashVarAddrsToI won't change if its a true byref, only
11716 // if it is in the stack)
11717 impBashVarAddrsToI(op1, op2);
11719 type = impGetByRefResultType(oper, uns, &op1, &op2);
11721 assert(!ovfl || !varTypeIsFloating(op1->gtType));
11723 /* Special case: "int+0", "int-0", "int*1", "int/1" */
11725 if (op2->gtOper == GT_CNS_INT)
11727 if ((op2->IsIntegralConst(0) && (oper == GT_ADD || oper == GT_SUB)) ||
11728 (op2->IsIntegralConst(1) && (oper == GT_MUL || oper == GT_DIV)))
11731 impPushOnStack(op1, tiRetVal);
11736 #if !FEATURE_X87_DOUBLES
11737 // We can generate a TYP_FLOAT operation that has a TYP_DOUBLE operand
11739 if (varTypeIsFloating(type) && varTypeIsFloating(op1->gtType) && varTypeIsFloating(op2->gtType))
11741 if (op1->TypeGet() != type)
11743 // We insert a cast of op1 to 'type'
11744 op1 = gtNewCastNode(type, op1, false, type);
11746 if (op2->TypeGet() != type)
11748 // We insert a cast of op2 to 'type'
11749 op2 = gtNewCastNode(type, op2, false, type);
11752 #endif // !FEATURE_X87_DOUBLES
11754 #if SMALL_TREE_NODES
11757 /* These operators can later be transformed into 'GT_CALL' */
11759 assert(GenTree::s_gtNodeSizes[GT_CALL] > GenTree::s_gtNodeSizes[GT_MUL]);
11760 #ifndef _TARGET_ARM_
11761 assert(GenTree::s_gtNodeSizes[GT_CALL] > GenTree::s_gtNodeSizes[GT_DIV]);
11762 assert(GenTree::s_gtNodeSizes[GT_CALL] > GenTree::s_gtNodeSizes[GT_UDIV]);
11763 assert(GenTree::s_gtNodeSizes[GT_CALL] > GenTree::s_gtNodeSizes[GT_MOD]);
11764 assert(GenTree::s_gtNodeSizes[GT_CALL] > GenTree::s_gtNodeSizes[GT_UMOD]);
11766 // It's tempting to use LargeOpOpcode() here, but this logic is *not* saying
11767 // that we'll need to transform into a general large node, but rather specifically
11768 // to a call: by doing it this way, things keep working if there are multiple sizes,
11769 // and a CALL is no longer the largest.
11770 // That said, as of now it *is* a large node, so we'll do this with an assert rather
11772 assert(GenTree::s_gtNodeSizes[GT_CALL] == TREE_NODE_SZ_LARGE);
11773 op1 = new (this, GT_CALL) GenTreeOp(oper, type, op1, op2 DEBUGARG(/*largeNode*/ true));
11776 #endif // SMALL_TREE_NODES
11778 op1 = gtNewOperNode(oper, type, op1, op2);
11781 /* Special case: integer/long division may throw an exception */
11783 if (varTypeIsIntegral(op1->TypeGet()) && op1->OperMayThrow(this))
11785 op1->gtFlags |= GTF_EXCEPT;
11790 assert(oper == GT_ADD || oper == GT_SUB || oper == GT_MUL);
11791 if (ovflType != TYP_UNKNOWN)
11793 op1->gtType = ovflType;
11795 op1->gtFlags |= (GTF_EXCEPT | GTF_OVERFLOW);
11798 op1->gtFlags |= GTF_UNSIGNED;
11802 impPushOnStack(op1, tiRetVal);
11817 if (tiVerificationNeeded)
11819 const typeInfo& tiVal = impStackTop(1).seTypeInfo;
11820 const typeInfo& tiShift = impStackTop(0).seTypeInfo;
11821 Verify(tiVal.IsIntegerType() && tiShift.IsType(TI_INT), "Bad shift args");
11824 op2 = impPopStack().val;
11825 op1 = impPopStack().val; // operand to be shifted
11826 impBashVarAddrsToI(op1, op2);
11828 type = genActualType(op1->TypeGet());
11829 op1 = gtNewOperNode(oper, type, op1, op2);
11831 impPushOnStack(op1, tiRetVal);
11835 if (tiVerificationNeeded)
11837 tiRetVal = impStackTop().seTypeInfo;
11838 Verify(tiRetVal.IsIntegerType(), "bad int value");
11841 op1 = impPopStack().val;
11842 impBashVarAddrsToI(op1, nullptr);
11843 type = genActualType(op1->TypeGet());
11844 impPushOnStack(gtNewOperNode(GT_NOT, type, op1), tiRetVal);
11848 if (tiVerificationNeeded)
11850 tiRetVal = impStackTop().seTypeInfo;
11851 Verify(tiRetVal.IsType(TI_DOUBLE), "bad R value");
11853 op1 = impPopStack().val;
11854 type = op1->TypeGet();
11855 op1 = gtNewOperNode(GT_CKFINITE, type, op1);
11856 op1->gtFlags |= GTF_EXCEPT;
11858 impPushOnStack(op1, tiRetVal);
11863 val = getI4LittleEndian(codeAddr); // jump distance
11864 jmpAddr = (IL_OFFSET)((codeAddr - info.compCode + sizeof(__int32)) + val);
11868 val = getI1LittleEndian(codeAddr); // jump distance
11869 jmpAddr = (IL_OFFSET)((codeAddr - info.compCode + sizeof(__int8)) + val);
11873 if (compIsForInlining())
11875 compInlineResult->NoteFatal(InlineObservation::CALLEE_HAS_LEAVE);
11879 JITDUMP(" %04X", jmpAddr);
11880 if (block->bbJumpKind != BBJ_LEAVE)
11882 impResetLeaveBlock(block, jmpAddr);
11885 assert(jmpAddr == block->bbJumpDest->bbCodeOffs);
11886 impImportLeave(block);
11887 impNoteBranchOffs();
11893 jmpDist = (sz == 1) ? getI1LittleEndian(codeAddr) : getI4LittleEndian(codeAddr);
11895 if (compIsForInlining() && jmpDist == 0)
11900 impNoteBranchOffs();
11906 case CEE_BRFALSE_S:
11908 /* Pop the comparand (now there's a neat term) from the stack */
11909 if (tiVerificationNeeded)
11911 typeInfo& tiVal = impStackTop().seTypeInfo;
11912 Verify(tiVal.IsObjRef() || tiVal.IsByRef() || tiVal.IsIntegerType() || tiVal.IsMethod(),
11916 op1 = impPopStack().val;
11917 type = op1->TypeGet();
11919 // brfalse and brtrue is only allowed on I4, refs, and byrefs.
11920 if (!opts.MinOpts() && !opts.compDbgCode && block->bbJumpDest == block->bbNext)
11922 block->bbJumpKind = BBJ_NONE;
11924 if (op1->gtFlags & GTF_GLOB_EFFECT)
11926 op1 = gtUnusedValNode(op1);
11935 if (op1->OperIsCompare())
11937 if (opcode == CEE_BRFALSE || opcode == CEE_BRFALSE_S)
11939 // Flip the sense of the compare
11941 op1 = gtReverseCond(op1);
11946 /* We'll compare against an equally-sized integer 0 */
11947 /* For small types, we always compare against int */
11948 op2 = gtNewZeroConNode(genActualType(op1->gtType));
11950 /* Create the comparison operator and try to fold it */
11952 oper = (opcode == CEE_BRTRUE || opcode == CEE_BRTRUE_S) ? GT_NE : GT_EQ;
11953 op1 = gtNewOperNode(oper, TYP_INT, op1, op2);
11960 /* Fold comparison if we can */
11962 op1 = gtFoldExpr(op1);
11964 /* Try to fold the really simple cases like 'iconst *, ifne/ifeq'*/
11965 /* Don't make any blocks unreachable in import only mode */
11967 if ((op1->gtOper == GT_CNS_INT) && !compIsForImportOnly())
11969 /* gtFoldExpr() should prevent this as we don't want to make any blocks
11970 unreachable under compDbgCode */
11971 assert(!opts.compDbgCode);
11973 BBjumpKinds foldedJumpKind = (BBjumpKinds)(op1->gtIntCon.gtIconVal ? BBJ_ALWAYS : BBJ_NONE);
11974 assertImp((block->bbJumpKind == BBJ_COND) // normal case
11975 || (block->bbJumpKind == foldedJumpKind)); // this can happen if we are reimporting the
11976 // block for the second time
11978 block->bbJumpKind = foldedJumpKind;
11982 if (op1->gtIntCon.gtIconVal)
11984 printf("\nThe conditional jump becomes an unconditional jump to BB%02u\n",
11985 block->bbJumpDest->bbNum);
11989 printf("\nThe block falls through into the next BB%02u\n", block->bbNext->bbNum);
11996 op1 = gtNewOperNode(GT_JTRUE, TYP_VOID, op1);
11998 /* GT_JTRUE is handled specially for non-empty stacks. See 'addStmt'
11999 in impImportBlock(block). For correct line numbers, spill stack. */
12001 if (opts.compDbgCode && impCurStmtOffs != BAD_IL_OFFSET)
12003 impSpillStackEnsure(true);
12030 if (tiVerificationNeeded)
12032 verVerifyCond(impStackTop(1).seTypeInfo, impStackTop().seTypeInfo, opcode);
12033 tiRetVal = typeInfo(TI_INT);
12036 op2 = impPopStack().val;
12037 op1 = impPopStack().val;
12039 #ifdef _TARGET_64BIT_
12040 if (varTypeIsI(op1->TypeGet()) && (genActualType(op2->TypeGet()) == TYP_INT))
12042 op2 = gtNewCastNode(TYP_I_IMPL, op2, uns, uns ? TYP_U_IMPL : TYP_I_IMPL);
12044 else if (varTypeIsI(op2->TypeGet()) && (genActualType(op1->TypeGet()) == TYP_INT))
12046 op1 = gtNewCastNode(TYP_I_IMPL, op1, uns, uns ? TYP_U_IMPL : TYP_I_IMPL);
12048 #endif // _TARGET_64BIT_
12050 assertImp(genActualType(op1->TypeGet()) == genActualType(op2->TypeGet()) ||
12051 varTypeIsI(op1->TypeGet()) && varTypeIsI(op2->TypeGet()) ||
12052 varTypeIsFloating(op1->gtType) && varTypeIsFloating(op2->gtType));
12054 /* Create the comparison node */
12056 op1 = gtNewOperNode(oper, TYP_INT, op1, op2);
12058 /* TODO: setting both flags when only one is appropriate */
12059 if (opcode == CEE_CGT_UN || opcode == CEE_CLT_UN)
12061 op1->gtFlags |= GTF_RELOP_NAN_UN | GTF_UNSIGNED;
12064 // Fold result, if possible.
12065 op1 = gtFoldExpr(op1);
12067 impPushOnStack(op1, tiRetVal);
12073 goto CMP_2_OPs_AND_BR;
12078 goto CMP_2_OPs_AND_BR;
12083 goto CMP_2_OPs_AND_BR_UN;
12088 goto CMP_2_OPs_AND_BR;
12093 goto CMP_2_OPs_AND_BR_UN;
12098 goto CMP_2_OPs_AND_BR;
12103 goto CMP_2_OPs_AND_BR_UN;
12108 goto CMP_2_OPs_AND_BR;
12113 goto CMP_2_OPs_AND_BR_UN;
12118 goto CMP_2_OPs_AND_BR_UN;
12120 CMP_2_OPs_AND_BR_UN:
12123 goto CMP_2_OPs_AND_BR_ALL;
12127 goto CMP_2_OPs_AND_BR_ALL;
12128 CMP_2_OPs_AND_BR_ALL:
12130 if (tiVerificationNeeded)
12132 verVerifyCond(impStackTop(1).seTypeInfo, impStackTop().seTypeInfo, opcode);
12135 /* Pull two values */
12136 op2 = impPopStack().val;
12137 op1 = impPopStack().val;
12139 #ifdef _TARGET_64BIT_
12140 if ((op1->TypeGet() == TYP_I_IMPL) && (genActualType(op2->TypeGet()) == TYP_INT))
12142 op2 = gtNewCastNode(TYP_I_IMPL, op2, uns, uns ? TYP_U_IMPL : TYP_I_IMPL);
12144 else if ((op2->TypeGet() == TYP_I_IMPL) && (genActualType(op1->TypeGet()) == TYP_INT))
12146 op1 = gtNewCastNode(TYP_I_IMPL, op1, uns, uns ? TYP_U_IMPL : TYP_I_IMPL);
12148 #endif // _TARGET_64BIT_
12150 assertImp(genActualType(op1->TypeGet()) == genActualType(op2->TypeGet()) ||
12151 varTypeIsI(op1->TypeGet()) && varTypeIsI(op2->TypeGet()) ||
12152 varTypeIsFloating(op1->gtType) && varTypeIsFloating(op2->gtType));
12154 if (!opts.MinOpts() && !opts.compDbgCode && block->bbJumpDest == block->bbNext)
12156 block->bbJumpKind = BBJ_NONE;
12158 if (op1->gtFlags & GTF_GLOB_EFFECT)
12160 impSpillSideEffects(false, (unsigned)CHECK_SPILL_ALL DEBUGARG(
12161 "Branch to next Optimization, op1 side effect"));
12162 impAppendTree(gtUnusedValNode(op1), (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
12164 if (op2->gtFlags & GTF_GLOB_EFFECT)
12166 impSpillSideEffects(false, (unsigned)CHECK_SPILL_ALL DEBUGARG(
12167 "Branch to next Optimization, op2 side effect"));
12168 impAppendTree(gtUnusedValNode(op2), (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
12172 if ((op1->gtFlags | op2->gtFlags) & GTF_GLOB_EFFECT)
12174 impNoteLastILoffs();
12179 #if !FEATURE_X87_DOUBLES
12180 // We can generate an compare of different sized floating point op1 and op2
12181 // We insert a cast
12183 if (varTypeIsFloating(op1->TypeGet()))
12185 if (op1->TypeGet() != op2->TypeGet())
12187 assert(varTypeIsFloating(op2->TypeGet()));
12189 // say op1=double, op2=float. To avoid loss of precision
12190 // while comparing, op2 is converted to double and double
12191 // comparison is done.
12192 if (op1->TypeGet() == TYP_DOUBLE)
12194 // We insert a cast of op2 to TYP_DOUBLE
12195 op2 = gtNewCastNode(TYP_DOUBLE, op2, false, TYP_DOUBLE);
12197 else if (op2->TypeGet() == TYP_DOUBLE)
12199 // We insert a cast of op1 to TYP_DOUBLE
12200 op1 = gtNewCastNode(TYP_DOUBLE, op1, false, TYP_DOUBLE);
12204 #endif // !FEATURE_X87_DOUBLES
12206 /* Create and append the operator */
12208 op1 = gtNewOperNode(oper, TYP_INT, op1, op2);
12212 op1->gtFlags |= GTF_UNSIGNED;
12217 op1->gtFlags |= GTF_RELOP_NAN_UN;
12223 assert(!compIsForInlining());
12225 if (tiVerificationNeeded)
12227 Verify(impStackTop().seTypeInfo.IsType(TI_INT), "Bad switch val");
12229 /* Pop the switch value off the stack */
12230 op1 = impPopStack().val;
12231 assertImp(genActualTypeIsIntOrI(op1->TypeGet()));
12233 /* We can create a switch node */
12235 op1 = gtNewOperNode(GT_SWITCH, TYP_VOID, op1);
12237 val = (int)getU4LittleEndian(codeAddr);
12238 codeAddr += 4 + val * 4; // skip over the switch-table
12242 /************************** Casting OPCODES ***************************/
12244 case CEE_CONV_OVF_I1:
12247 case CEE_CONV_OVF_I2:
12248 lclTyp = TYP_SHORT;
12250 case CEE_CONV_OVF_I:
12251 lclTyp = TYP_I_IMPL;
12253 case CEE_CONV_OVF_I4:
12256 case CEE_CONV_OVF_I8:
12260 case CEE_CONV_OVF_U1:
12261 lclTyp = TYP_UBYTE;
12263 case CEE_CONV_OVF_U2:
12264 lclTyp = TYP_USHORT;
12266 case CEE_CONV_OVF_U:
12267 lclTyp = TYP_U_IMPL;
12269 case CEE_CONV_OVF_U4:
12272 case CEE_CONV_OVF_U8:
12273 lclTyp = TYP_ULONG;
12276 case CEE_CONV_OVF_I1_UN:
12279 case CEE_CONV_OVF_I2_UN:
12280 lclTyp = TYP_SHORT;
12282 case CEE_CONV_OVF_I_UN:
12283 lclTyp = TYP_I_IMPL;
12285 case CEE_CONV_OVF_I4_UN:
12288 case CEE_CONV_OVF_I8_UN:
12292 case CEE_CONV_OVF_U1_UN:
12293 lclTyp = TYP_UBYTE;
12295 case CEE_CONV_OVF_U2_UN:
12296 lclTyp = TYP_USHORT;
12298 case CEE_CONV_OVF_U_UN:
12299 lclTyp = TYP_U_IMPL;
12301 case CEE_CONV_OVF_U4_UN:
12304 case CEE_CONV_OVF_U8_UN:
12305 lclTyp = TYP_ULONG;
12310 goto CONV_OVF_COMMON;
12313 goto CONV_OVF_COMMON;
12323 lclTyp = TYP_SHORT;
12326 lclTyp = TYP_I_IMPL;
12336 lclTyp = TYP_UBYTE;
12339 lclTyp = TYP_USHORT;
12341 #if (REGSIZE_BYTES == 8)
12343 lclTyp = TYP_U_IMPL;
12347 lclTyp = TYP_U_IMPL;
12354 lclTyp = TYP_ULONG;
12358 lclTyp = TYP_FLOAT;
12361 lclTyp = TYP_DOUBLE;
12364 case CEE_CONV_R_UN:
12365 lclTyp = TYP_DOUBLE;
12379 // just check that we have a number on the stack
12380 if (tiVerificationNeeded)
12382 const typeInfo& tiVal = impStackTop().seTypeInfo;
12383 Verify(tiVal.IsNumberType(), "bad arg");
12385 #ifdef _TARGET_64BIT_
12386 bool isNative = false;
12390 case CEE_CONV_OVF_I:
12391 case CEE_CONV_OVF_I_UN:
12393 case CEE_CONV_OVF_U:
12394 case CEE_CONV_OVF_U_UN:
12398 // leave 'isNative' = false;
12403 tiRetVal = typeInfo::nativeInt();
12406 #endif // _TARGET_64BIT_
12408 tiRetVal = typeInfo(lclTyp).NormaliseForStack();
12412 // only converts from FLOAT or DOUBLE to an integer type
12413 // and converts from ULONG (or LONG on ARM) to DOUBLE are morphed to calls
12415 if (varTypeIsFloating(lclTyp))
12417 callNode = varTypeIsLong(impStackTop().val) || uns // uint->dbl gets turned into uint->long->dbl
12418 #ifdef _TARGET_64BIT_
12419 // TODO-ARM64-Bug?: This was AMD64; I enabled it for ARM64 also. OK?
12420 // TYP_BYREF could be used as TYP_I_IMPL which is long.
12421 // TODO-CQ: remove this when we lower casts long/ulong --> float/double
12422 // and generate SSE2 code instead of going through helper calls.
12423 || (impStackTop().val->TypeGet() == TYP_BYREF)
12429 callNode = varTypeIsFloating(impStackTop().val->TypeGet());
12432 // At this point uns, ovf, callNode all set
12434 op1 = impPopStack().val;
12435 impBashVarAddrsToI(op1);
12437 if (varTypeIsSmall(lclTyp) && !ovfl && op1->gtType == TYP_INT && op1->gtOper == GT_AND)
12439 op2 = op1->gtOp.gtOp2;
12441 if (op2->gtOper == GT_CNS_INT)
12443 ssize_t ival = op2->gtIntCon.gtIconVal;
12444 ssize_t mask, umask;
12460 assert(!"unexpected type");
12464 if (((ival & umask) == ival) || ((ival & mask) == ival && uns))
12466 /* Toss the cast, it's a waste of time */
12468 impPushOnStack(op1, tiRetVal);
12471 else if (ival == mask)
12473 /* Toss the masking, it's a waste of time, since
12474 we sign-extend from the small value anyways */
12476 op1 = op1->gtOp.gtOp1;
12481 /* The 'op2' sub-operand of a cast is the 'real' type number,
12482 since the result of a cast to one of the 'small' integer
12483 types is an integer.
12486 type = genActualType(lclTyp);
12488 #if SMALL_TREE_NODES
12491 op1 = gtNewCastNodeL(type, op1, uns, lclTyp);
12494 #endif // SMALL_TREE_NODES
12496 op1 = gtNewCastNode(type, op1, uns, lclTyp);
12501 op1->gtFlags |= (GTF_OVERFLOW | GTF_EXCEPT);
12503 impPushOnStack(op1, tiRetVal);
12507 if (tiVerificationNeeded)
12509 tiRetVal = impStackTop().seTypeInfo;
12510 Verify(tiRetVal.IsNumberType(), "Bad arg");
12513 op1 = impPopStack().val;
12514 impBashVarAddrsToI(op1, nullptr);
12515 impPushOnStack(gtNewOperNode(GT_NEG, genActualType(op1->gtType), op1), tiRetVal);
12520 /* Pull the top value from the stack */
12522 StackEntry se = impPopStack();
12523 clsHnd = se.seTypeInfo.GetClassHandle();
12526 /* Get hold of the type of the value being duplicated */
12528 lclTyp = genActualType(op1->gtType);
12530 /* Does the value have any side effects? */
12532 if ((op1->gtFlags & GTF_SIDE_EFFECT) || opts.compDbgCode)
12534 // Since we are throwing away the value, just normalize
12535 // it to its address. This is more efficient.
12537 if (varTypeIsStruct(op1))
12539 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
12540 // Non-calls, such as obj or ret_expr, have to go through this.
12541 // Calls with large struct return value have to go through this.
12542 // Helper calls with small struct return value also have to go
12543 // through this since they do not follow Unix calling convention.
12544 if (op1->gtOper != GT_CALL || !IsMultiRegReturnedType(clsHnd) ||
12545 op1->AsCall()->gtCallType == CT_HELPER)
12546 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
12548 op1 = impGetStructAddr(op1, clsHnd, (unsigned)CHECK_SPILL_ALL, false);
12552 // If op1 is non-overflow cast, throw it away since it is useless.
12553 // Another reason for throwing away the useless cast is in the context of
12554 // implicit tail calls when the operand of pop is GT_CAST(GT_CALL(..)).
12555 // The cast gets added as part of importing GT_CALL, which gets in the way
12556 // of fgMorphCall() on the forms of tail call nodes that we assert.
12557 if ((op1->gtOper == GT_CAST) && !op1->gtOverflow())
12559 op1 = op1->gtOp.gtOp1;
12562 // If 'op1' is an expression, create an assignment node.
12563 // Helps analyses (like CSE) to work fine.
12565 if (op1->gtOper != GT_CALL)
12567 op1 = gtUnusedValNode(op1);
12570 /* Append the value to the tree list */
12574 /* No side effects - just throw the <BEEP> thing away */
12580 if (tiVerificationNeeded)
12582 // Dup could start the begining of delegate creation sequence, remember that
12583 delegateCreateStart = codeAddr - 1;
12587 // If the expression to dup is simple, just clone it.
12588 // Otherwise spill it to a temp, and reload the temp
12590 StackEntry se = impPopStack();
12591 GenTree* tree = se.val;
12592 tiRetVal = se.seTypeInfo;
12595 if (!opts.compDbgCode && !op1->IsIntegralConst(0) && !op1->IsFPZero() && !op1->IsLocal())
12597 const unsigned tmpNum = lvaGrabTemp(true DEBUGARG("dup spill"));
12598 impAssignTempGen(tmpNum, op1, tiRetVal.GetClassHandle(), (unsigned)CHECK_SPILL_ALL);
12599 var_types type = genActualType(lvaTable[tmpNum].TypeGet());
12600 op1 = gtNewLclvNode(tmpNum, type);
12602 // Propagate type info to the temp from the stack and the original tree
12603 if (type == TYP_REF)
12605 lvaSetClass(tmpNum, tree, tiRetVal.GetClassHandle());
12609 op1 = impCloneExpr(op1, &op2, tiRetVal.GetClassHandle(), (unsigned)CHECK_SPILL_ALL,
12610 nullptr DEBUGARG("DUP instruction"));
12612 assert(!(op1->gtFlags & GTF_GLOB_EFFECT) && !(op2->gtFlags & GTF_GLOB_EFFECT));
12613 impPushOnStack(op1, tiRetVal);
12614 impPushOnStack(op2, tiRetVal);
12622 lclTyp = TYP_SHORT;
12631 lclTyp = TYP_I_IMPL;
12633 case CEE_STIND_REF:
12637 lclTyp = TYP_FLOAT;
12640 lclTyp = TYP_DOUBLE;
12644 if (tiVerificationNeeded)
12646 typeInfo instrType(lclTyp);
12647 #ifdef _TARGET_64BIT_
12648 if (opcode == CEE_STIND_I)
12650 instrType = typeInfo::nativeInt();
12652 #endif // _TARGET_64BIT_
12653 verVerifySTIND(impStackTop(1).seTypeInfo, impStackTop(0).seTypeInfo, instrType);
12657 compUnsafeCastUsed = true; // Have to go conservative
12662 op2 = impPopStack().val; // value to store
12663 op1 = impPopStack().val; // address to store to
12665 // you can indirect off of a TYP_I_IMPL (if we are in C) or a BYREF
12666 assertImp(genActualType(op1->gtType) == TYP_I_IMPL || op1->gtType == TYP_BYREF);
12668 impBashVarAddrsToI(op1, op2);
12670 op2 = impImplicitR4orR8Cast(op2, lclTyp);
12672 #ifdef _TARGET_64BIT_
12673 // Automatic upcast for a GT_CNS_INT into TYP_I_IMPL
12674 if ((op2->OperGet() == GT_CNS_INT) && varTypeIsI(lclTyp) && !varTypeIsI(op2->gtType))
12676 op2->gtType = TYP_I_IMPL;
12680 // Allow a downcast of op2 from TYP_I_IMPL into a 32-bit Int for x86 JIT compatiblity
12682 if (varTypeIsI(op2->gtType) && (genActualType(lclTyp) == TYP_INT))
12684 assert(!tiVerificationNeeded); // We should have thrown the VerificationException before.
12685 op2 = gtNewCastNode(TYP_INT, op2, false, TYP_INT);
12687 // Allow an upcast of op2 from a 32-bit Int into TYP_I_IMPL for x86 JIT compatiblity
12689 if (varTypeIsI(lclTyp) && (genActualType(op2->gtType) == TYP_INT))
12691 assert(!tiVerificationNeeded); // We should have thrown the VerificationException before.
12692 op2 = gtNewCastNode(TYP_I_IMPL, op2, false, TYP_I_IMPL);
12695 #endif // _TARGET_64BIT_
12697 if (opcode == CEE_STIND_REF)
12699 // STIND_REF can be used to store TYP_INT, TYP_I_IMPL, TYP_REF, or TYP_BYREF
12700 assertImp(varTypeIsIntOrI(op2->gtType) || varTypeIsGC(op2->gtType));
12701 lclTyp = genActualType(op2->TypeGet());
12704 // Check target type.
12706 if (op2->gtType == TYP_BYREF || lclTyp == TYP_BYREF)
12708 if (op2->gtType == TYP_BYREF)
12710 assertImp(lclTyp == TYP_BYREF || lclTyp == TYP_I_IMPL);
12712 else if (lclTyp == TYP_BYREF)
12714 assertImp(op2->gtType == TYP_BYREF || varTypeIsIntOrI(op2->gtType));
12719 assertImp(genActualType(op2->gtType) == genActualType(lclTyp) ||
12720 ((lclTyp == TYP_I_IMPL) && (genActualType(op2->gtType) == TYP_INT)) ||
12721 (varTypeIsFloating(op2->gtType) && varTypeIsFloating(lclTyp)));
12725 op1 = gtNewOperNode(GT_IND, lclTyp, op1);
12727 // stind could point anywhere, example a boxed class static int
12728 op1->gtFlags |= GTF_IND_TGTANYWHERE;
12730 if (prefixFlags & PREFIX_VOLATILE)
12732 assert(op1->OperGet() == GT_IND);
12733 op1->gtFlags |= GTF_DONT_CSE; // Can't CSE a volatile
12734 op1->gtFlags |= GTF_ORDER_SIDEEFF; // Prevent this from being reordered
12735 op1->gtFlags |= GTF_IND_VOLATILE;
12738 if ((prefixFlags & PREFIX_UNALIGNED) && !varTypeIsByte(lclTyp))
12740 assert(op1->OperGet() == GT_IND);
12741 op1->gtFlags |= GTF_IND_UNALIGNED;
12744 op1 = gtNewAssignNode(op1, op2);
12745 op1->gtFlags |= GTF_EXCEPT | GTF_GLOB_REF;
12747 // Spill side-effects AND global-data-accesses
12748 if (verCurrentState.esStackDepth > 0)
12750 impSpillSideEffects(true, (unsigned)CHECK_SPILL_ALL DEBUGARG("spill side effects before STIND"));
12759 lclTyp = TYP_SHORT;
12768 case CEE_LDIND_REF:
12772 lclTyp = TYP_I_IMPL;
12775 lclTyp = TYP_FLOAT;
12778 lclTyp = TYP_DOUBLE;
12781 lclTyp = TYP_UBYTE;
12784 lclTyp = TYP_USHORT;
12788 if (tiVerificationNeeded)
12790 typeInfo lclTiType(lclTyp);
12791 #ifdef _TARGET_64BIT_
12792 if (opcode == CEE_LDIND_I)
12794 lclTiType = typeInfo::nativeInt();
12796 #endif // _TARGET_64BIT_
12797 tiRetVal = verVerifyLDIND(impStackTop().seTypeInfo, lclTiType);
12798 tiRetVal.NormaliseForStack();
12802 compUnsafeCastUsed = true; // Have to go conservative
12807 op1 = impPopStack().val; // address to load from
12808 impBashVarAddrsToI(op1);
12810 #ifdef _TARGET_64BIT_
12811 // Allow an upcast of op1 from a 32-bit Int into TYP_I_IMPL for x86 JIT compatiblity
12813 if (genActualType(op1->gtType) == TYP_INT)
12815 assert(!tiVerificationNeeded); // We should have thrown the VerificationException before.
12816 op1 = gtNewCastNode(TYP_I_IMPL, op1, false, TYP_I_IMPL);
12820 assertImp(genActualType(op1->gtType) == TYP_I_IMPL || op1->gtType == TYP_BYREF);
12822 op1 = gtNewOperNode(GT_IND, lclTyp, op1);
12824 // ldind could point anywhere, example a boxed class static int
12825 op1->gtFlags |= (GTF_EXCEPT | GTF_GLOB_REF | GTF_IND_TGTANYWHERE);
12827 if (prefixFlags & PREFIX_VOLATILE)
12829 assert(op1->OperGet() == GT_IND);
12830 op1->gtFlags |= GTF_DONT_CSE; // Can't CSE a volatile
12831 op1->gtFlags |= GTF_ORDER_SIDEEFF; // Prevent this from being reordered
12832 op1->gtFlags |= GTF_IND_VOLATILE;
12835 if ((prefixFlags & PREFIX_UNALIGNED) && !varTypeIsByte(lclTyp))
12837 assert(op1->OperGet() == GT_IND);
12838 op1->gtFlags |= GTF_IND_UNALIGNED;
12841 impPushOnStack(op1, tiRetVal);
12845 case CEE_UNALIGNED:
12848 val = getU1LittleEndian(codeAddr);
12850 JITDUMP(" %u", val);
12851 if ((val != 1) && (val != 2) && (val != 4))
12853 BADCODE("Alignment unaligned. must be 1, 2, or 4");
12856 Verify(!(prefixFlags & PREFIX_UNALIGNED), "Multiple unaligned. prefixes");
12857 prefixFlags |= PREFIX_UNALIGNED;
12859 impValidateMemoryAccessOpcode(codeAddr, codeEndp, false);
12862 opcode = (OPCODE)getU1LittleEndian(codeAddr);
12863 opcodeOffs = (IL_OFFSET)(codeAddr - info.compCode);
12864 codeAddr += sizeof(__int8);
12865 goto DECODE_OPCODE;
12869 Verify(!(prefixFlags & PREFIX_VOLATILE), "Multiple volatile. prefixes");
12870 prefixFlags |= PREFIX_VOLATILE;
12872 impValidateMemoryAccessOpcode(codeAddr, codeEndp, true);
12879 // Need to do a lookup here so that we perform an access check
12880 // and do a NOWAY if protections are violated
12881 _impResolveToken(CORINFO_TOKENKIND_Method);
12883 JITDUMP(" %08X", resolvedToken.token);
12885 eeGetCallInfo(&resolvedToken, nullptr /* constraint typeRef*/,
12886 addVerifyFlag(combine(CORINFO_CALLINFO_SECURITYCHECKS, CORINFO_CALLINFO_LDFTN)),
12889 // This check really only applies to intrinsic Array.Address methods
12890 if (callInfo.sig.callConv & CORINFO_CALLCONV_PARAMTYPE)
12892 NO_WAY("Currently do not support LDFTN of Parameterized functions");
12895 // Do this before DO_LDFTN since CEE_LDVIRTFN does it on its own.
12896 impHandleAccessAllowed(callInfo.accessAllowed, &callInfo.callsiteCalloutHelper);
12898 if (tiVerificationNeeded)
12900 // LDFTN could start the begining of delegate creation sequence, remember that
12901 delegateCreateStart = codeAddr - 2;
12903 // check any constraints on the callee's class and type parameters
12904 VerifyOrReturn(info.compCompHnd->satisfiesClassConstraints(resolvedToken.hClass),
12905 "method has unsatisfied class constraints");
12906 VerifyOrReturn(info.compCompHnd->satisfiesMethodConstraints(resolvedToken.hClass,
12907 resolvedToken.hMethod),
12908 "method has unsatisfied method constraints");
12910 mflags = callInfo.verMethodFlags;
12911 Verify(!(mflags & CORINFO_FLG_CONSTRUCTOR), "LDFTN on a constructor");
12915 op1 = impMethodPointer(&resolvedToken, &callInfo);
12916 if (compDonotInline())
12921 CORINFO_RESOLVED_TOKEN* heapToken = impAllocateToken(resolvedToken);
12922 impPushOnStack(op1, typeInfo(heapToken));
12927 case CEE_LDVIRTFTN:
12929 /* Get the method token */
12931 _impResolveToken(CORINFO_TOKENKIND_Method);
12933 JITDUMP(" %08X", resolvedToken.token);
12935 eeGetCallInfo(&resolvedToken, nullptr /* constraint typeRef */,
12936 addVerifyFlag(combine(combine(CORINFO_CALLINFO_SECURITYCHECKS, CORINFO_CALLINFO_LDFTN),
12937 CORINFO_CALLINFO_CALLVIRT)),
12940 // This check really only applies to intrinsic Array.Address methods
12941 if (callInfo.sig.callConv & CORINFO_CALLCONV_PARAMTYPE)
12943 NO_WAY("Currently do not support LDFTN of Parameterized functions");
12946 mflags = callInfo.methodFlags;
12948 impHandleAccessAllowed(callInfo.accessAllowed, &callInfo.callsiteCalloutHelper);
12950 if (compIsForInlining())
12952 if (mflags & (CORINFO_FLG_FINAL | CORINFO_FLG_STATIC) || !(mflags & CORINFO_FLG_VIRTUAL))
12954 compInlineResult->NoteFatal(InlineObservation::CALLSITE_LDVIRTFN_ON_NON_VIRTUAL);
12959 CORINFO_SIG_INFO& ftnSig = callInfo.sig;
12961 if (tiVerificationNeeded)
12964 Verify(ftnSig.hasThis(), "ldvirtftn on a static method");
12965 Verify(!(mflags & CORINFO_FLG_CONSTRUCTOR), "LDVIRTFTN on a constructor");
12967 // JIT32 verifier rejects verifiable ldvirtftn pattern
12968 typeInfo declType =
12969 verMakeTypeInfo(resolvedToken.hClass, true); // Change TI_STRUCT to TI_REF when necessary
12971 typeInfo arg = impStackTop().seTypeInfo;
12972 Verify((arg.IsType(TI_REF) || arg.IsType(TI_NULL)) && tiCompatibleWith(arg, declType, true),
12975 CORINFO_CLASS_HANDLE instanceClassHnd = info.compClassHnd;
12976 if (!(arg.IsType(TI_NULL) || (mflags & CORINFO_FLG_STATIC)))
12978 instanceClassHnd = arg.GetClassHandleForObjRef();
12981 // check any constraints on the method's class and type parameters
12982 VerifyOrReturn(info.compCompHnd->satisfiesClassConstraints(resolvedToken.hClass),
12983 "method has unsatisfied class constraints");
12984 VerifyOrReturn(info.compCompHnd->satisfiesMethodConstraints(resolvedToken.hClass,
12985 resolvedToken.hMethod),
12986 "method has unsatisfied method constraints");
12988 if (mflags & CORINFO_FLG_PROTECTED)
12990 Verify(info.compCompHnd->canAccessFamily(info.compMethodHnd, instanceClassHnd),
12991 "Accessing protected method through wrong type.");
12995 /* Get the object-ref */
12996 op1 = impPopStack().val;
12997 assertImp(op1->gtType == TYP_REF);
12999 if (opts.IsReadyToRun())
13001 if (callInfo.kind != CORINFO_VIRTUALCALL_LDVIRTFTN)
13003 if (op1->gtFlags & GTF_SIDE_EFFECT)
13005 op1 = gtUnusedValNode(op1);
13006 impAppendTree(op1, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
13011 else if (mflags & (CORINFO_FLG_FINAL | CORINFO_FLG_STATIC) || !(mflags & CORINFO_FLG_VIRTUAL))
13013 if (op1->gtFlags & GTF_SIDE_EFFECT)
13015 op1 = gtUnusedValNode(op1);
13016 impAppendTree(op1, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
13021 GenTree* fptr = impImportLdvirtftn(op1, &resolvedToken, &callInfo);
13022 if (compDonotInline())
13027 CORINFO_RESOLVED_TOKEN* heapToken = impAllocateToken(resolvedToken);
13028 assert(heapToken->tokenType == CORINFO_TOKENKIND_Method);
13029 heapToken->tokenType = CORINFO_TOKENKIND_Ldvirtftn;
13030 impPushOnStack(fptr, typeInfo(heapToken));
13035 case CEE_CONSTRAINED:
13037 assertImp(sz == sizeof(unsigned));
13038 impResolveToken(codeAddr, &constrainedResolvedToken, CORINFO_TOKENKIND_Constrained);
13039 codeAddr += sizeof(unsigned); // prefix instructions must increment codeAddr manually
13040 JITDUMP(" (%08X) ", constrainedResolvedToken.token);
13042 Verify(!(prefixFlags & PREFIX_CONSTRAINED), "Multiple constrained. prefixes");
13043 prefixFlags |= PREFIX_CONSTRAINED;
13046 OPCODE actualOpcode = impGetNonPrefixOpcode(codeAddr, codeEndp);
13047 if (actualOpcode != CEE_CALLVIRT)
13049 BADCODE("constrained. has to be followed by callvirt");
13056 JITDUMP(" readonly.");
13058 Verify(!(prefixFlags & PREFIX_READONLY), "Multiple readonly. prefixes");
13059 prefixFlags |= PREFIX_READONLY;
13062 OPCODE actualOpcode = impGetNonPrefixOpcode(codeAddr, codeEndp);
13063 if (actualOpcode != CEE_LDELEMA && !impOpcodeIsCallOpcode(actualOpcode))
13065 BADCODE("readonly. has to be followed by ldelema or call");
13075 Verify(!(prefixFlags & PREFIX_TAILCALL_EXPLICIT), "Multiple tailcall. prefixes");
13076 prefixFlags |= PREFIX_TAILCALL_EXPLICIT;
13079 OPCODE actualOpcode = impGetNonPrefixOpcode(codeAddr, codeEndp);
13080 if (!impOpcodeIsCallOpcode(actualOpcode))
13082 BADCODE("tailcall. has to be followed by call, callvirt or calli");
13090 /* Since we will implicitly insert newObjThisPtr at the start of the
13091 argument list, spill any GTF_ORDER_SIDEEFF */
13092 impSpillSpecialSideEff();
13094 /* NEWOBJ does not respond to TAIL */
13095 prefixFlags &= ~PREFIX_TAILCALL_EXPLICIT;
13097 /* NEWOBJ does not respond to CONSTRAINED */
13098 prefixFlags &= ~PREFIX_CONSTRAINED;
13100 _impResolveToken(CORINFO_TOKENKIND_NewObj);
13102 eeGetCallInfo(&resolvedToken, nullptr /* constraint typeRef*/,
13103 addVerifyFlag(combine(CORINFO_CALLINFO_SECURITYCHECKS, CORINFO_CALLINFO_ALLOWINSTPARAM)),
13106 if (compIsForInlining())
13108 if (impInlineInfo->inlineCandidateInfo->dwRestrictions & INLINE_RESPECT_BOUNDARY)
13110 // Check to see if this call violates the boundary.
13111 compInlineResult->NoteFatal(InlineObservation::CALLSITE_CROSS_BOUNDARY_SECURITY);
13116 mflags = callInfo.methodFlags;
13118 if ((mflags & (CORINFO_FLG_STATIC | CORINFO_FLG_ABSTRACT)) != 0)
13120 BADCODE("newobj on static or abstract method");
13123 // Insert the security callout before any actual code is generated
13124 impHandleAccessAllowed(callInfo.accessAllowed, &callInfo.callsiteCalloutHelper);
13126 // There are three different cases for new
13127 // Object size is variable (depends on arguments)
13128 // 1) Object is an array (arrays treated specially by the EE)
13129 // 2) Object is some other variable sized object (e.g. String)
13130 // 3) Class Size can be determined beforehand (normal case)
13131 // In the first case, we need to call a NEWOBJ helper (multinewarray)
13132 // in the second case we call the constructor with a '0' this pointer
13133 // In the third case we alloc the memory, then call the constuctor
13135 clsFlags = callInfo.classFlags;
13136 if (clsFlags & CORINFO_FLG_ARRAY)
13138 if (tiVerificationNeeded)
13140 CORINFO_CLASS_HANDLE elemTypeHnd;
13141 INDEBUG(CorInfoType corType =)
13142 info.compCompHnd->getChildType(resolvedToken.hClass, &elemTypeHnd);
13143 assert(!(elemTypeHnd == nullptr && corType == CORINFO_TYPE_VALUECLASS));
13144 Verify(elemTypeHnd == nullptr ||
13145 !(info.compCompHnd->getClassAttribs(elemTypeHnd) & CORINFO_FLG_CONTAINS_STACK_PTR),
13146 "newarr of byref-like objects");
13147 verVerifyCall(opcode, &resolvedToken, nullptr, ((prefixFlags & PREFIX_TAILCALL_EXPLICIT) != 0),
13148 ((prefixFlags & PREFIX_READONLY) != 0), delegateCreateStart, codeAddr - 1,
13149 &callInfo DEBUGARG(info.compFullName));
13151 // Arrays need to call the NEWOBJ helper.
13152 assertImp(clsFlags & CORINFO_FLG_VAROBJSIZE);
13154 impImportNewObjArray(&resolvedToken, &callInfo);
13155 if (compDonotInline())
13163 // At present this can only be String
13164 else if (clsFlags & CORINFO_FLG_VAROBJSIZE)
13166 if (IsTargetAbi(CORINFO_CORERT_ABI))
13168 // The dummy argument does not exist in CoreRT
13169 newObjThisPtr = nullptr;
13173 // This is the case for variable-sized objects that are not
13174 // arrays. In this case, call the constructor with a null 'this'
13176 newObjThisPtr = gtNewIconNode(0, TYP_REF);
13179 /* Remember that this basic block contains 'new' of an object */
13180 block->bbFlags |= BBF_HAS_NEWOBJ;
13181 optMethodFlags |= OMF_HAS_NEWOBJ;
13185 // This is the normal case where the size of the object is
13186 // fixed. Allocate the memory and call the constructor.
13188 // Note: We cannot add a peep to avoid use of temp here
13189 // becase we don't have enough interference info to detect when
13190 // sources and destination interfere, example: s = new S(ref);
13192 // TODO: We find the correct place to introduce a general
13193 // reverse copy prop for struct return values from newobj or
13194 // any function returning structs.
13196 /* get a temporary for the new object */
13197 lclNum = lvaGrabTemp(true DEBUGARG("NewObj constructor temp"));
13198 if (compDonotInline())
13200 // Fail fast if lvaGrabTemp fails with CALLSITE_TOO_MANY_LOCALS.
13201 assert(compInlineResult->GetObservation() == InlineObservation::CALLSITE_TOO_MANY_LOCALS);
13205 // In the value class case we only need clsHnd for size calcs.
13207 // The lookup of the code pointer will be handled by CALL in this case
13208 if (clsFlags & CORINFO_FLG_VALUECLASS)
13210 if (compIsForInlining())
13212 // If value class has GC fields, inform the inliner. It may choose to
13213 // bail out on the inline.
13214 DWORD typeFlags = info.compCompHnd->getClassAttribs(resolvedToken.hClass);
13215 if ((typeFlags & CORINFO_FLG_CONTAINS_GC_PTR) != 0)
13217 compInlineResult->Note(InlineObservation::CALLEE_HAS_GC_STRUCT);
13218 if (compInlineResult->IsFailure())
13223 // Do further notification in the case where the call site is rare;
13224 // some policies do not track the relative hotness of call sites for
13225 // "always" inline cases.
13226 if (impInlineInfo->iciBlock->isRunRarely())
13228 compInlineResult->Note(InlineObservation::CALLSITE_RARE_GC_STRUCT);
13229 if (compInlineResult->IsFailure())
13237 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(resolvedToken.hClass);
13238 unsigned size = info.compCompHnd->getClassSize(resolvedToken.hClass);
13240 if (impIsPrimitive(jitTyp))
13242 lvaTable[lclNum].lvType = JITtype2varType(jitTyp);
13246 // The local variable itself is the allocated space.
13247 // Here we need unsafe value cls check, since the address of struct is taken for further use
13248 // and potentially exploitable.
13249 lvaSetStruct(lclNum, resolvedToken.hClass, true /* unsafe value cls check */);
13251 if (compIsForInlining() || fgStructTempNeedsExplicitZeroInit(lvaTable + lclNum, block))
13253 // Append a tree to zero-out the temp
13254 newObjThisPtr = gtNewLclvNode(lclNum, lvaTable[lclNum].TypeGet());
13256 newObjThisPtr = gtNewBlkOpNode(newObjThisPtr, // Dest
13257 gtNewIconNode(0), // Value
13259 false, // isVolatile
13260 false); // not copyBlock
13261 impAppendTree(newObjThisPtr, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
13264 // Obtain the address of the temp
13266 gtNewOperNode(GT_ADDR, TYP_BYREF, gtNewLclvNode(lclNum, lvaTable[lclNum].TypeGet()));
13270 #ifdef FEATURE_READYTORUN_COMPILER
13271 if (opts.IsReadyToRun())
13273 op1 = impReadyToRunHelperToTree(&resolvedToken, CORINFO_HELP_READYTORUN_NEW, TYP_REF);
13274 usingReadyToRunHelper = (op1 != nullptr);
13277 if (!usingReadyToRunHelper)
13280 op1 = impParentClassTokenToHandle(&resolvedToken, nullptr, TRUE);
13281 if (op1 == nullptr)
13282 { // compDonotInline()
13286 // TODO: ReadyToRun: When generic dictionary lookups are necessary, replace the lookup call
13287 // and the newfast call with a single call to a dynamic R2R cell that will:
13288 // 1) Load the context
13289 // 2) Perform the generic dictionary lookup and caching, and generate the appropriate
13291 // 3) Allocate and return the new object
13292 // Reason: performance (today, we'll always use the slow helper for the R2R generics case)
13294 op1 = gtNewAllocObjNode(info.compCompHnd->getNewHelper(&resolvedToken, info.compMethodHnd),
13295 resolvedToken.hClass, TYP_REF, op1);
13298 // Remember that this basic block contains 'new' of an object
13299 block->bbFlags |= BBF_HAS_NEWOBJ;
13300 optMethodFlags |= OMF_HAS_NEWOBJ;
13302 // Append the assignment to the temp/local. Dont need to spill
13303 // at all as we are just calling an EE-Jit helper which can only
13304 // cause an (async) OutOfMemoryException.
13306 // We assign the newly allocated object (by a GT_ALLOCOBJ node)
13307 // to a temp. Note that the pattern "temp = allocObj" is required
13308 // by ObjectAllocator phase to be able to determine GT_ALLOCOBJ nodes
13309 // without exhaustive walk over all expressions.
13311 impAssignTempGen(lclNum, op1, (unsigned)CHECK_SPILL_NONE);
13312 lvaSetClass(lclNum, resolvedToken.hClass, true /* is Exact */);
13314 newObjThisPtr = gtNewLclvNode(lclNum, TYP_REF);
13321 /* CALLI does not respond to CONSTRAINED */
13322 prefixFlags &= ~PREFIX_CONSTRAINED;
13324 if (compIsForInlining())
13326 // CALLI doesn't have a method handle, so assume the worst.
13327 if (impInlineInfo->inlineCandidateInfo->dwRestrictions & INLINE_RESPECT_BOUNDARY)
13329 compInlineResult->NoteFatal(InlineObservation::CALLSITE_CROSS_BOUNDARY_CALLI);
13339 // We can't call getCallInfo on the token from a CALLI, but we need it in
13340 // many other places. We unfortunately embed that knowledge here.
13341 if (opcode != CEE_CALLI)
13343 _impResolveToken(CORINFO_TOKENKIND_Method);
13345 eeGetCallInfo(&resolvedToken,
13346 (prefixFlags & PREFIX_CONSTRAINED) ? &constrainedResolvedToken : nullptr,
13347 // this is how impImportCall invokes getCallInfo
13349 combine(combine(CORINFO_CALLINFO_ALLOWINSTPARAM, CORINFO_CALLINFO_SECURITYCHECKS),
13350 (opcode == CEE_CALLVIRT) ? CORINFO_CALLINFO_CALLVIRT
13351 : CORINFO_CALLINFO_NONE)),
13356 // Suppress uninitialized use warning.
13357 memset(&resolvedToken, 0, sizeof(resolvedToken));
13358 memset(&callInfo, 0, sizeof(callInfo));
13360 resolvedToken.token = getU4LittleEndian(codeAddr);
13363 CALL: // memberRef should be set.
13364 // newObjThisPtr should be set for CEE_NEWOBJ
13366 JITDUMP(" %08X", resolvedToken.token);
13367 constraintCall = (prefixFlags & PREFIX_CONSTRAINED) != 0;
13369 bool newBBcreatedForTailcallStress;
13371 newBBcreatedForTailcallStress = false;
13373 if (compIsForInlining())
13375 if (compDonotInline())
13379 // We rule out inlinees with explicit tail calls in fgMakeBasicBlocks.
13380 assert((prefixFlags & PREFIX_TAILCALL_EXPLICIT) == 0);
13384 if (compTailCallStress())
13386 // Have we created a new BB after the "call" instruction in fgMakeBasicBlocks()?
13387 // Tail call stress only recognizes call+ret patterns and forces them to be
13388 // explicit tail prefixed calls. Also fgMakeBasicBlocks() under tail call stress
13389 // doesn't import 'ret' opcode following the call into the basic block containing
13390 // the call instead imports it to a new basic block. Note that fgMakeBasicBlocks()
13391 // is already checking that there is an opcode following call and hence it is
13392 // safe here to read next opcode without bounds check.
13393 newBBcreatedForTailcallStress =
13394 impOpcodeIsCallOpcode(opcode) && // Current opcode is a CALL, (not a CEE_NEWOBJ). So, don't
13395 // make it jump to RET.
13396 (OPCODE)getU1LittleEndian(codeAddr + sz) == CEE_RET; // Next opcode is a CEE_RET
13398 if (newBBcreatedForTailcallStress &&
13399 !(prefixFlags & PREFIX_TAILCALL_EXPLICIT) && // User hasn't set "tail." prefix yet.
13400 verCheckTailCallConstraint(opcode, &resolvedToken,
13401 constraintCall ? &constrainedResolvedToken : nullptr,
13402 true) // Is it legal to do tailcall?
13405 // Stress the tailcall.
13406 JITDUMP(" (Tailcall stress: prefixFlags |= PREFIX_TAILCALL_EXPLICIT)");
13407 prefixFlags |= PREFIX_TAILCALL_EXPLICIT;
13412 // This is split up to avoid goto flow warnings.
13414 isRecursive = !compIsForInlining() && (callInfo.hMethod == info.compMethodHnd);
13416 // Note that when running under tail call stress, a call will be marked as explicit tail prefixed
13417 // hence will not be considered for implicit tail calling.
13418 if (impIsImplicitTailCallCandidate(opcode, codeAddr + sz, codeEndp, prefixFlags, isRecursive))
13420 if (compIsForInlining())
13422 #if FEATURE_TAILCALL_OPT_SHARED_RETURN
13423 // Are we inlining at an implicit tail call site? If so the we can flag
13424 // implicit tail call sites in the inline body. These call sites
13425 // often end up in non BBJ_RETURN blocks, so only flag them when
13426 // we're able to handle shared returns.
13427 if (impInlineInfo->iciCall->IsImplicitTailCall())
13429 JITDUMP(" (Inline Implicit Tail call: prefixFlags |= PREFIX_TAILCALL_IMPLICIT)");
13430 prefixFlags |= PREFIX_TAILCALL_IMPLICIT;
13432 #endif // FEATURE_TAILCALL_OPT_SHARED_RETURN
13436 JITDUMP(" (Implicit Tail call: prefixFlags |= PREFIX_TAILCALL_IMPLICIT)");
13437 prefixFlags |= PREFIX_TAILCALL_IMPLICIT;
13441 // Treat this call as tail call for verification only if "tail" prefixed (i.e. explicit tail call).
13442 explicitTailCall = (prefixFlags & PREFIX_TAILCALL_EXPLICIT) != 0;
13443 readonlyCall = (prefixFlags & PREFIX_READONLY) != 0;
13445 if (opcode != CEE_CALLI && opcode != CEE_NEWOBJ)
13447 // All calls and delegates need a security callout.
13448 // For delegates, this is the call to the delegate constructor, not the access check on the
13450 impHandleAccessAllowed(callInfo.accessAllowed, &callInfo.callsiteCalloutHelper);
13452 #if 0 // DevDiv 410397 - This breaks too many obfuscated apps to do this in an in-place release
13454 // DevDiv 291703 - we need to check for accessibility between the caller of InitializeArray
13455 // and the field it is reading, thus it is now unverifiable to not immediately precede with
13456 // ldtoken <filed token>, and we now check accessibility
13457 if ((callInfo.methodFlags & CORINFO_FLG_INTRINSIC) &&
13458 (info.compCompHnd->getIntrinsicID(callInfo.hMethod) == CORINFO_INTRINSIC_InitializeArray))
13460 if (prevOpcode != CEE_LDTOKEN)
13462 Verify(prevOpcode == CEE_LDTOKEN, "Need ldtoken for InitializeArray");
13466 assert(lastLoadToken != NULL);
13467 // Now that we know we have a token, verify that it is accessible for loading
13468 CORINFO_RESOLVED_TOKEN resolvedLoadField;
13469 impResolveToken(lastLoadToken, &resolvedLoadField, CORINFO_TOKENKIND_Field);
13470 eeGetFieldInfo(&resolvedLoadField, CORINFO_ACCESS_INIT_ARRAY, &fieldInfo);
13471 impHandleAccessAllowed(fieldInfo.accessAllowed, &fieldInfo.accessCalloutHelper);
13475 #endif // DevDiv 410397
13478 if (tiVerificationNeeded)
13480 verVerifyCall(opcode, &resolvedToken, constraintCall ? &constrainedResolvedToken : nullptr,
13481 explicitTailCall, readonlyCall, delegateCreateStart, codeAddr - 1,
13482 &callInfo DEBUGARG(info.compFullName));
13485 // Insert delegate callout here.
13486 if (opcode == CEE_NEWOBJ && (mflags & CORINFO_FLG_CONSTRUCTOR) && (clsFlags & CORINFO_FLG_DELEGATE))
13489 // We should do this only if verification is enabled
13490 // If verification is disabled, delegateCreateStart will not be initialized correctly
13491 if (tiVerificationNeeded)
13493 mdMemberRef delegateMethodRef = mdMemberRefNil;
13494 // We should get here only for well formed delegate creation.
13495 assert(verCheckDelegateCreation(delegateCreateStart, codeAddr - 1, delegateMethodRef));
13500 callTyp = impImportCall(opcode, &resolvedToken, constraintCall ? &constrainedResolvedToken : nullptr,
13501 newObjThisPtr, prefixFlags, &callInfo, opcodeOffs);
13502 if (compDonotInline())
13504 // We do not check fails after lvaGrabTemp. It is covered with CoreCLR_13272 issue.
13505 assert((callTyp == TYP_UNDEF) ||
13506 (compInlineResult->GetObservation() == InlineObservation::CALLSITE_TOO_MANY_LOCALS));
13510 if (explicitTailCall || newBBcreatedForTailcallStress) // If newBBcreatedForTailcallStress is true, we
13511 // have created a new BB after the "call"
13512 // instruction in fgMakeBasicBlocks(). So we need to jump to RET regardless.
13514 assert(!compIsForInlining());
13526 BOOL isLoadAddress = (opcode == CEE_LDFLDA || opcode == CEE_LDSFLDA);
13527 BOOL isLoadStatic = (opcode == CEE_LDSFLD || opcode == CEE_LDSFLDA);
13529 /* Get the CP_Fieldref index */
13530 assertImp(sz == sizeof(unsigned));
13532 _impResolveToken(CORINFO_TOKENKIND_Field);
13534 JITDUMP(" %08X", resolvedToken.token);
13536 int aflags = isLoadAddress ? CORINFO_ACCESS_ADDRESS : CORINFO_ACCESS_GET;
13538 GenTree* obj = nullptr;
13539 typeInfo* tiObj = nullptr;
13540 CORINFO_CLASS_HANDLE objType = nullptr; // used for fields
13542 if (opcode == CEE_LDFLD || opcode == CEE_LDFLDA)
13544 tiObj = &impStackTop().seTypeInfo;
13545 StackEntry se = impPopStack();
13546 objType = se.seTypeInfo.GetClassHandle();
13549 if (impIsThis(obj))
13551 aflags |= CORINFO_ACCESS_THIS;
13553 // An optimization for Contextful classes:
13554 // we unwrap the proxy when we have a 'this reference'
13556 if (info.compUnwrapContextful)
13558 aflags |= CORINFO_ACCESS_UNWRAP;
13563 eeGetFieldInfo(&resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo);
13565 // Figure out the type of the member. We always call canAccessField, so you always need this
13567 CorInfoType ciType = fieldInfo.fieldType;
13568 clsHnd = fieldInfo.structType;
13570 lclTyp = JITtype2varType(ciType);
13572 #ifdef _TARGET_AMD64
13573 noway_assert(varTypeIsIntegralOrI(lclTyp) || varTypeIsFloating(lclTyp) || lclTyp == TYP_STRUCT);
13574 #endif // _TARGET_AMD64
13576 if (compIsForInlining())
13578 switch (fieldInfo.fieldAccessor)
13580 case CORINFO_FIELD_INSTANCE_HELPER:
13581 case CORINFO_FIELD_INSTANCE_ADDR_HELPER:
13582 case CORINFO_FIELD_STATIC_ADDR_HELPER:
13583 case CORINFO_FIELD_STATIC_TLS:
13585 compInlineResult->NoteFatal(InlineObservation::CALLEE_LDFLD_NEEDS_HELPER);
13588 case CORINFO_FIELD_STATIC_GENERICS_STATIC_HELPER:
13589 case CORINFO_FIELD_STATIC_READYTORUN_HELPER:
13590 /* We may be able to inline the field accessors in specific instantiations of generic
13592 compInlineResult->NoteFatal(InlineObservation::CALLSITE_LDFLD_NEEDS_HELPER);
13599 if (!isLoadAddress && (fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC) && lclTyp == TYP_STRUCT &&
13602 if ((info.compCompHnd->getTypeForPrimitiveValueClass(clsHnd) == CORINFO_TYPE_UNDEF) &&
13603 !(info.compFlags & CORINFO_FLG_FORCEINLINE))
13605 // Loading a static valuetype field usually will cause a JitHelper to be called
13606 // for the static base. This will bloat the code.
13607 compInlineResult->Note(InlineObservation::CALLEE_LDFLD_STATIC_VALUECLASS);
13609 if (compInlineResult->IsFailure())
13617 tiRetVal = verMakeTypeInfo(ciType, clsHnd);
13620 tiRetVal.MakeByRef();
13624 tiRetVal.NormaliseForStack();
13627 // Perform this check always to ensure that we get field access exceptions even with
13628 // SkipVerification.
13629 impHandleAccessAllowed(fieldInfo.accessAllowed, &fieldInfo.accessCalloutHelper);
13631 if (tiVerificationNeeded)
13633 // You can also pass the unboxed struct to LDFLD
13634 BOOL bAllowPlainValueTypeAsThis = FALSE;
13635 if (opcode == CEE_LDFLD && impIsValueType(tiObj))
13637 bAllowPlainValueTypeAsThis = TRUE;
13640 verVerifyField(&resolvedToken, fieldInfo, tiObj, isLoadAddress, bAllowPlainValueTypeAsThis);
13642 // If we're doing this on a heap object or from a 'safe' byref
13643 // then the result is a safe byref too
13644 if (isLoadAddress) // load address
13646 if (fieldInfo.fieldFlags &
13647 CORINFO_FLG_FIELD_STATIC) // statics marked as safe will have permanent home
13649 if (fieldInfo.fieldFlags & CORINFO_FLG_FIELD_SAFESTATIC_BYREF_RETURN)
13651 tiRetVal.SetIsPermanentHomeByRef();
13654 else if (tiObj->IsObjRef() || tiObj->IsPermanentHomeByRef())
13656 // ldflda of byref is safe if done on a gc object or on a
13658 tiRetVal.SetIsPermanentHomeByRef();
13664 // tiVerificationNeeded is false.
13665 // Raise InvalidProgramException if static load accesses non-static field
13666 if (isLoadStatic && ((fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC) == 0))
13668 BADCODE("static access on an instance field");
13672 // We are using ldfld/a on a static field. We allow it, but need to get side-effect from obj.
13673 if ((fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC) && obj != nullptr)
13675 if (obj->gtFlags & GTF_SIDE_EFFECT)
13677 obj = gtUnusedValNode(obj);
13678 impAppendTree(obj, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
13683 /* Preserve 'small' int types */
13684 if (!varTypeIsSmall(lclTyp))
13686 lclTyp = genActualType(lclTyp);
13689 bool usesHelper = false;
13691 switch (fieldInfo.fieldAccessor)
13693 case CORINFO_FIELD_INSTANCE:
13694 #ifdef FEATURE_READYTORUN_COMPILER
13695 case CORINFO_FIELD_INSTANCE_WITH_BASE:
13698 bool nullcheckNeeded = false;
13700 obj = impCheckForNullPointer(obj);
13702 if (isLoadAddress && (obj->gtType == TYP_BYREF) && fgAddrCouldBeNull(obj))
13704 nullcheckNeeded = true;
13707 // If the object is a struct, what we really want is
13708 // for the field to operate on the address of the struct.
13709 if (!varTypeGCtype(obj->TypeGet()) && impIsValueType(tiObj))
13711 assert(opcode == CEE_LDFLD && objType != nullptr);
13713 obj = impGetStructAddr(obj, objType, (unsigned)CHECK_SPILL_ALL, true);
13716 /* Create the data member node */
13717 op1 = gtNewFieldRef(lclTyp, resolvedToken.hField, obj, fieldInfo.offset, nullcheckNeeded);
13719 #ifdef FEATURE_READYTORUN_COMPILER
13720 if (fieldInfo.fieldAccessor == CORINFO_FIELD_INSTANCE_WITH_BASE)
13722 op1->gtField.gtFieldLookup = fieldInfo.fieldLookup;
13726 op1->gtFlags |= (obj->gtFlags & GTF_GLOB_EFFECT);
13728 if (fgAddrCouldBeNull(obj))
13730 op1->gtFlags |= GTF_EXCEPT;
13733 // If gtFldObj is a BYREF then our target is a value class and
13734 // it could point anywhere, example a boxed class static int
13735 if (obj->gtType == TYP_BYREF)
13737 op1->gtFlags |= GTF_IND_TGTANYWHERE;
13740 DWORD typeFlags = info.compCompHnd->getClassAttribs(resolvedToken.hClass);
13741 if (StructHasOverlappingFields(typeFlags))
13743 op1->gtField.gtFldMayOverlap = true;
13746 // wrap it in a address of operator if necessary
13749 op1 = gtNewOperNode(GT_ADDR,
13750 (var_types)(varTypeIsGC(obj->TypeGet()) ? TYP_BYREF : TYP_I_IMPL), op1);
13754 if (compIsForInlining() &&
13755 impInlineIsGuaranteedThisDerefBeforeAnySideEffects(nullptr, obj,
13756 impInlineInfo->inlArgInfo))
13758 impInlineInfo->thisDereferencedFirst = true;
13764 case CORINFO_FIELD_STATIC_TLS:
13765 #ifdef _TARGET_X86_
13766 // Legacy TLS access is implemented as intrinsic on x86 only
13768 /* Create the data member node */
13769 op1 = gtNewFieldRef(lclTyp, resolvedToken.hField, NULL, fieldInfo.offset);
13770 op1->gtFlags |= GTF_IND_TLS_REF; // fgMorphField will handle the transformation
13774 op1 = gtNewOperNode(GT_ADDR, (var_types)TYP_I_IMPL, op1);
13778 fieldInfo.fieldAccessor = CORINFO_FIELD_STATIC_ADDR_HELPER;
13783 case CORINFO_FIELD_STATIC_ADDR_HELPER:
13784 case CORINFO_FIELD_INSTANCE_HELPER:
13785 case CORINFO_FIELD_INSTANCE_ADDR_HELPER:
13786 op1 = gtNewRefCOMfield(obj, &resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo, lclTyp,
13791 case CORINFO_FIELD_STATIC_ADDRESS:
13792 // Replace static read-only fields with constant if possible
13793 if ((aflags & CORINFO_ACCESS_GET) && (fieldInfo.fieldFlags & CORINFO_FLG_FIELD_FINAL) &&
13794 !(fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC_IN_HEAP) &&
13795 (varTypeIsIntegral(lclTyp) || varTypeIsFloating(lclTyp)))
13797 CorInfoInitClassResult initClassResult =
13798 info.compCompHnd->initClass(resolvedToken.hField, info.compMethodHnd,
13799 impTokenLookupContextHandle);
13801 if (initClassResult & CORINFO_INITCLASS_INITIALIZED)
13803 void** pFldAddr = nullptr;
13805 info.compCompHnd->getFieldAddress(resolvedToken.hField, (void**)&pFldAddr);
13807 // We should always be able to access this static's address directly
13808 assert(pFldAddr == nullptr);
13810 op1 = impImportStaticReadOnlyField(fldAddr, lclTyp);
13817 case CORINFO_FIELD_STATIC_RVA_ADDRESS:
13818 case CORINFO_FIELD_STATIC_SHARED_STATIC_HELPER:
13819 case CORINFO_FIELD_STATIC_GENERICS_STATIC_HELPER:
13820 case CORINFO_FIELD_STATIC_READYTORUN_HELPER:
13821 op1 = impImportStaticFieldAccess(&resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo,
13825 case CORINFO_FIELD_INTRINSIC_ZERO:
13827 assert(aflags & CORINFO_ACCESS_GET);
13828 op1 = gtNewIconNode(0, lclTyp);
13833 case CORINFO_FIELD_INTRINSIC_EMPTY_STRING:
13835 assert(aflags & CORINFO_ACCESS_GET);
13838 InfoAccessType iat = info.compCompHnd->emptyStringLiteral(&pValue);
13839 op1 = gtNewStringLiteralNode(iat, pValue);
13844 case CORINFO_FIELD_INTRINSIC_ISLITTLEENDIAN:
13846 assert(aflags & CORINFO_ACCESS_GET);
13848 op1 = gtNewIconNode(0, lclTyp);
13850 op1 = gtNewIconNode(1, lclTyp);
13857 assert(!"Unexpected fieldAccessor");
13860 if (!isLoadAddress)
13863 if (prefixFlags & PREFIX_VOLATILE)
13865 op1->gtFlags |= GTF_DONT_CSE; // Can't CSE a volatile
13866 op1->gtFlags |= GTF_ORDER_SIDEEFF; // Prevent this from being reordered
13870 assert((op1->OperGet() == GT_FIELD) || (op1->OperGet() == GT_IND) ||
13871 (op1->OperGet() == GT_OBJ));
13872 op1->gtFlags |= GTF_IND_VOLATILE;
13876 if ((prefixFlags & PREFIX_UNALIGNED) && !varTypeIsByte(lclTyp))
13880 assert((op1->OperGet() == GT_FIELD) || (op1->OperGet() == GT_IND) ||
13881 (op1->OperGet() == GT_OBJ));
13882 op1->gtFlags |= GTF_IND_UNALIGNED;
13887 /* Check if the class needs explicit initialization */
13889 if (fieldInfo.fieldFlags & CORINFO_FLG_FIELD_INITCLASS)
13891 GenTree* helperNode = impInitClass(&resolvedToken);
13892 if (compDonotInline())
13896 if (helperNode != nullptr)
13898 op1 = gtNewOperNode(GT_COMMA, op1->TypeGet(), helperNode, op1);
13903 impPushOnStack(op1, tiRetVal);
13911 BOOL isStoreStatic = (opcode == CEE_STSFLD);
13913 CORINFO_CLASS_HANDLE fieldClsHnd; // class of the field (if it's a ref type)
13915 /* Get the CP_Fieldref index */
13917 assertImp(sz == sizeof(unsigned));
13919 _impResolveToken(CORINFO_TOKENKIND_Field);
13921 JITDUMP(" %08X", resolvedToken.token);
13923 int aflags = CORINFO_ACCESS_SET;
13924 GenTree* obj = nullptr;
13925 typeInfo* tiObj = nullptr;
13928 /* Pull the value from the stack */
13929 StackEntry se = impPopStack();
13931 tiVal = se.seTypeInfo;
13932 clsHnd = tiVal.GetClassHandle();
13934 if (opcode == CEE_STFLD)
13936 tiObj = &impStackTop().seTypeInfo;
13937 obj = impPopStack().val;
13939 if (impIsThis(obj))
13941 aflags |= CORINFO_ACCESS_THIS;
13943 // An optimization for Contextful classes:
13944 // we unwrap the proxy when we have a 'this reference'
13946 if (info.compUnwrapContextful)
13948 aflags |= CORINFO_ACCESS_UNWRAP;
13953 eeGetFieldInfo(&resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo);
13955 // Figure out the type of the member. We always call canAccessField, so you always need this
13957 CorInfoType ciType = fieldInfo.fieldType;
13958 fieldClsHnd = fieldInfo.structType;
13960 lclTyp = JITtype2varType(ciType);
13962 if (compIsForInlining())
13964 /* Is this a 'special' (COM) field? or a TLS ref static field?, field stored int GC heap? or
13965 * per-inst static? */
13967 switch (fieldInfo.fieldAccessor)
13969 case CORINFO_FIELD_INSTANCE_HELPER:
13970 case CORINFO_FIELD_INSTANCE_ADDR_HELPER:
13971 case CORINFO_FIELD_STATIC_ADDR_HELPER:
13972 case CORINFO_FIELD_STATIC_TLS:
13974 compInlineResult->NoteFatal(InlineObservation::CALLEE_STFLD_NEEDS_HELPER);
13977 case CORINFO_FIELD_STATIC_GENERICS_STATIC_HELPER:
13978 case CORINFO_FIELD_STATIC_READYTORUN_HELPER:
13979 /* We may be able to inline the field accessors in specific instantiations of generic
13981 compInlineResult->NoteFatal(InlineObservation::CALLSITE_STFLD_NEEDS_HELPER);
13989 impHandleAccessAllowed(fieldInfo.accessAllowed, &fieldInfo.accessCalloutHelper);
13991 if (tiVerificationNeeded)
13993 verVerifyField(&resolvedToken, fieldInfo, tiObj, TRUE);
13994 typeInfo fieldType = verMakeTypeInfo(ciType, fieldClsHnd);
13995 Verify(tiCompatibleWith(tiVal, fieldType.NormaliseForStack(), true), "type mismatch");
13999 // tiVerificationNeed is false.
14000 // Raise InvalidProgramException if static store accesses non-static field
14001 if (isStoreStatic && ((fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC) == 0))
14003 BADCODE("static access on an instance field");
14007 // We are using stfld on a static field.
14008 // We allow it, but need to eval any side-effects for obj
14009 if ((fieldInfo.fieldFlags & CORINFO_FLG_FIELD_STATIC) && obj != nullptr)
14011 if (obj->gtFlags & GTF_SIDE_EFFECT)
14013 obj = gtUnusedValNode(obj);
14014 impAppendTree(obj, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
14019 /* Preserve 'small' int types */
14020 if (!varTypeIsSmall(lclTyp))
14022 lclTyp = genActualType(lclTyp);
14025 switch (fieldInfo.fieldAccessor)
14027 case CORINFO_FIELD_INSTANCE:
14028 #ifdef FEATURE_READYTORUN_COMPILER
14029 case CORINFO_FIELD_INSTANCE_WITH_BASE:
14032 obj = impCheckForNullPointer(obj);
14034 /* Create the data member node */
14035 op1 = gtNewFieldRef(lclTyp, resolvedToken.hField, obj, fieldInfo.offset);
14036 DWORD typeFlags = info.compCompHnd->getClassAttribs(resolvedToken.hClass);
14037 if (StructHasOverlappingFields(typeFlags))
14039 op1->gtField.gtFldMayOverlap = true;
14042 #ifdef FEATURE_READYTORUN_COMPILER
14043 if (fieldInfo.fieldAccessor == CORINFO_FIELD_INSTANCE_WITH_BASE)
14045 op1->gtField.gtFieldLookup = fieldInfo.fieldLookup;
14049 op1->gtFlags |= (obj->gtFlags & GTF_GLOB_EFFECT);
14051 if (fgAddrCouldBeNull(obj))
14053 op1->gtFlags |= GTF_EXCEPT;
14056 // If gtFldObj is a BYREF then our target is a value class and
14057 // it could point anywhere, example a boxed class static int
14058 if (obj->gtType == TYP_BYREF)
14060 op1->gtFlags |= GTF_IND_TGTANYWHERE;
14063 if (compIsForInlining() &&
14064 impInlineIsGuaranteedThisDerefBeforeAnySideEffects(op2, obj, impInlineInfo->inlArgInfo))
14066 impInlineInfo->thisDereferencedFirst = true;
14071 case CORINFO_FIELD_STATIC_TLS:
14072 #ifdef _TARGET_X86_
14073 // Legacy TLS access is implemented as intrinsic on x86 only
14075 /* Create the data member node */
14076 op1 = gtNewFieldRef(lclTyp, resolvedToken.hField, NULL, fieldInfo.offset);
14077 op1->gtFlags |= GTF_IND_TLS_REF; // fgMorphField will handle the transformation
14081 fieldInfo.fieldAccessor = CORINFO_FIELD_STATIC_ADDR_HELPER;
14086 case CORINFO_FIELD_STATIC_ADDR_HELPER:
14087 case CORINFO_FIELD_INSTANCE_HELPER:
14088 case CORINFO_FIELD_INSTANCE_ADDR_HELPER:
14089 op1 = gtNewRefCOMfield(obj, &resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo, lclTyp,
14093 case CORINFO_FIELD_STATIC_ADDRESS:
14094 case CORINFO_FIELD_STATIC_RVA_ADDRESS:
14095 case CORINFO_FIELD_STATIC_SHARED_STATIC_HELPER:
14096 case CORINFO_FIELD_STATIC_GENERICS_STATIC_HELPER:
14097 case CORINFO_FIELD_STATIC_READYTORUN_HELPER:
14098 op1 = impImportStaticFieldAccess(&resolvedToken, (CORINFO_ACCESS_FLAGS)aflags, &fieldInfo,
14103 assert(!"Unexpected fieldAccessor");
14106 // Create the member assignment, unless we have a struct.
14107 // TODO-1stClassStructs: This could be limited to TYP_STRUCT, to avoid extra copies.
14108 bool deferStructAssign = varTypeIsStruct(lclTyp);
14110 if (!deferStructAssign)
14112 if (prefixFlags & PREFIX_VOLATILE)
14114 assert((op1->OperGet() == GT_FIELD) || (op1->OperGet() == GT_IND));
14115 op1->gtFlags |= GTF_DONT_CSE; // Can't CSE a volatile
14116 op1->gtFlags |= GTF_ORDER_SIDEEFF; // Prevent this from being reordered
14117 op1->gtFlags |= GTF_IND_VOLATILE;
14119 if ((prefixFlags & PREFIX_UNALIGNED) && !varTypeIsByte(lclTyp))
14121 assert((op1->OperGet() == GT_FIELD) || (op1->OperGet() == GT_IND));
14122 op1->gtFlags |= GTF_IND_UNALIGNED;
14125 /* V4.0 allows assignment of i4 constant values to i8 type vars when IL verifier is bypassed (full
14126 trust apps). The reason this works is that JIT stores an i4 constant in Gentree union during
14127 importation and reads from the union as if it were a long during code generation. Though this
14128 can potentially read garbage, one can get lucky to have this working correctly.
14130 This code pattern is generated by Dev10 MC++ compiler while storing to fields when compiled with
14131 /O2 switch (default when compiling retail configs in Dev10) and a customer app has taken a
14132 dependency on it. To be backward compatible, we will explicitly add an upward cast here so that
14133 it works correctly always.
14135 Note that this is limited to x86 alone as there is no back compat to be addressed for Arm JIT
14138 CLANG_FORMAT_COMMENT_ANCHOR;
14140 #ifndef _TARGET_64BIT_
14141 // In UWP6.0 and beyond (post-.NET Core 2.0), we decided to let this cast from int to long be
14142 // generated for ARM as well as x86, so the following IR will be accepted:
14144 // | /--* CNS_INT int 2
14146 // \--* CLS_VAR long
14148 if ((op1->TypeGet() != op2->TypeGet()) && op2->OperIsConst() && varTypeIsIntOrI(op2->TypeGet()) &&
14149 varTypeIsLong(op1->TypeGet()))
14151 op2 = gtNewCastNode(op1->TypeGet(), op2, false, op1->TypeGet());
14155 #ifdef _TARGET_64BIT_
14156 // Automatic upcast for a GT_CNS_INT into TYP_I_IMPL
14157 if ((op2->OperGet() == GT_CNS_INT) && varTypeIsI(lclTyp) && !varTypeIsI(op2->gtType))
14159 op2->gtType = TYP_I_IMPL;
14163 // Allow a downcast of op2 from TYP_I_IMPL into a 32-bit Int for x86 JIT compatiblity
14165 if (varTypeIsI(op2->gtType) && (genActualType(lclTyp) == TYP_INT))
14167 op2 = gtNewCastNode(TYP_INT, op2, false, TYP_INT);
14169 // Allow an upcast of op2 from a 32-bit Int into TYP_I_IMPL for x86 JIT compatiblity
14171 if (varTypeIsI(lclTyp) && (genActualType(op2->gtType) == TYP_INT))
14173 op2 = gtNewCastNode(TYP_I_IMPL, op2, false, TYP_I_IMPL);
14178 #if !FEATURE_X87_DOUBLES
14179 // We can generate an assignment to a TYP_FLOAT from a TYP_DOUBLE
14180 // We insert a cast to the dest 'op1' type
14182 if ((op1->TypeGet() != op2->TypeGet()) && varTypeIsFloating(op1->gtType) &&
14183 varTypeIsFloating(op2->gtType))
14185 op2 = gtNewCastNode(op1->TypeGet(), op2, false, op1->TypeGet());
14187 #endif // !FEATURE_X87_DOUBLES
14189 op1 = gtNewAssignNode(op1, op2);
14191 /* Mark the expression as containing an assignment */
14193 op1->gtFlags |= GTF_ASG;
14196 /* Check if the class needs explicit initialization */
14198 if (fieldInfo.fieldFlags & CORINFO_FLG_FIELD_INITCLASS)
14200 GenTree* helperNode = impInitClass(&resolvedToken);
14201 if (compDonotInline())
14205 if (helperNode != nullptr)
14207 op1 = gtNewOperNode(GT_COMMA, op1->TypeGet(), helperNode, op1);
14211 /* stfld can interfere with value classes (consider the sequence
14212 ldloc, ldloca, ..., stfld, stloc). We will be conservative and
14213 spill all value class references from the stack. */
14215 if (obj && ((obj->gtType == TYP_BYREF) || (obj->gtType == TYP_I_IMPL)))
14219 if (impIsValueType(tiObj))
14221 impSpillEvalStack();
14225 impSpillValueClasses();
14229 /* Spill any refs to the same member from the stack */
14231 impSpillLclRefs((ssize_t)resolvedToken.hField);
14233 /* stsfld also interferes with indirect accesses (for aliased
14234 statics) and calls. But don't need to spill other statics
14235 as we have explicitly spilled this particular static field. */
14237 impSpillSideEffects(false, (unsigned)CHECK_SPILL_ALL DEBUGARG("spill side effects before STFLD"));
14239 if (deferStructAssign)
14241 op1 = impAssignStruct(op1, op2, clsHnd, (unsigned)CHECK_SPILL_ALL);
14249 /* Get the class type index operand */
14251 _impResolveToken(CORINFO_TOKENKIND_Newarr);
14253 JITDUMP(" %08X", resolvedToken.token);
14255 if (!opts.IsReadyToRun())
14257 // Need to restore array classes before creating array objects on the heap
14258 op1 = impTokenToHandle(&resolvedToken, nullptr, TRUE /*mustRestoreHandle*/);
14259 if (op1 == nullptr)
14260 { // compDonotInline()
14265 if (tiVerificationNeeded)
14267 // As per ECMA 'numElems' specified can be either int32 or native int.
14268 Verify(impStackTop().seTypeInfo.IsIntOrNativeIntType(), "bad bound");
14270 CORINFO_CLASS_HANDLE elemTypeHnd;
14271 info.compCompHnd->getChildType(resolvedToken.hClass, &elemTypeHnd);
14272 Verify(elemTypeHnd == nullptr ||
14273 !(info.compCompHnd->getClassAttribs(elemTypeHnd) & CORINFO_FLG_CONTAINS_STACK_PTR),
14274 "array of byref-like type");
14277 tiRetVal = verMakeTypeInfo(resolvedToken.hClass);
14279 accessAllowedResult =
14280 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
14281 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
14283 /* Form the arglist: array class handle, size */
14284 op2 = impPopStack().val;
14285 assertImp(genActualTypeIsIntOrI(op2->gtType));
14287 #ifdef FEATURE_READYTORUN_COMPILER
14288 if (opts.IsReadyToRun())
14290 op1 = impReadyToRunHelperToTree(&resolvedToken, CORINFO_HELP_READYTORUN_NEWARR_1, TYP_REF,
14291 gtNewArgList(op2));
14292 usingReadyToRunHelper = (op1 != nullptr);
14294 if (!usingReadyToRunHelper)
14296 // TODO: ReadyToRun: When generic dictionary lookups are necessary, replace the lookup call
14297 // and the newarr call with a single call to a dynamic R2R cell that will:
14298 // 1) Load the context
14299 // 2) Perform the generic dictionary lookup and caching, and generate the appropriate stub
14300 // 3) Allocate the new array
14301 // Reason: performance (today, we'll always use the slow helper for the R2R generics case)
14303 // Need to restore array classes before creating array objects on the heap
14304 op1 = impTokenToHandle(&resolvedToken, nullptr, TRUE /*mustRestoreHandle*/);
14305 if (op1 == nullptr)
14306 { // compDonotInline()
14312 if (!usingReadyToRunHelper)
14315 args = gtNewArgList(op1, op2);
14317 /* Create a call to 'new' */
14319 // Note that this only works for shared generic code because the same helper is used for all
14320 // reference array types
14321 op1 = gtNewHelperCallNode(info.compCompHnd->getNewArrHelper(resolvedToken.hClass), TYP_REF, args);
14324 op1->gtCall.compileTimeHelperArgumentHandle = (CORINFO_GENERIC_HANDLE)resolvedToken.hClass;
14326 /* Remember that this basic block contains 'new' of an sd array */
14328 block->bbFlags |= BBF_HAS_NEWARRAY;
14329 optMethodFlags |= OMF_HAS_NEWARRAY;
14331 /* Push the result of the call on the stack */
14333 impPushOnStack(op1, tiRetVal);
14340 if (tiVerificationNeeded)
14342 Verify(false, "bad opcode");
14345 // We don't allow locallocs inside handlers
14346 if (block->hasHndIndex())
14348 BADCODE("Localloc can't be inside handler");
14351 setNeedsGSSecurityCookie();
14353 // Get the size to allocate
14355 op2 = impPopStack().val;
14356 assertImp(genActualTypeIsIntOrI(op2->gtType));
14358 if (verCurrentState.esStackDepth != 0)
14360 BADCODE("Localloc can only be used when the stack is empty");
14363 // If the localloc is not in a loop and its size is a small constant,
14364 // create a new local var of TYP_BLK and return its address.
14366 bool convertedToLocal = false;
14368 // Need to aggressively fold here, as even fixed-size locallocs
14369 // will have casts in the way.
14370 op2 = gtFoldExpr(op2);
14372 if (op2->IsIntegralConst())
14374 const ssize_t allocSize = op2->AsIntCon()->IconValue();
14376 if (allocSize == 0)
14378 // Result is nullptr
14379 JITDUMP("Converting stackalloc of 0 bytes to push null unmanaged pointer\n");
14380 op1 = gtNewIconNode(0, TYP_I_IMPL);
14381 convertedToLocal = true;
14383 else if ((allocSize > 0) && ((compCurBB->bbFlags & BBF_BACKWARD_JUMP) == 0))
14385 // Get the size threshold for local conversion
14386 ssize_t maxSize = DEFAULT_MAX_LOCALLOC_TO_LOCAL_SIZE;
14389 // Optionally allow this to be modified
14390 maxSize = JitConfig.JitStackAllocToLocalSize();
14393 if (allocSize <= maxSize)
14395 const unsigned stackallocAsLocal = lvaGrabTemp(false DEBUGARG("stackallocLocal"));
14396 JITDUMP("Converting stackalloc of %lld bytes to new local V%02u\n", allocSize,
14397 stackallocAsLocal);
14398 lvaTable[stackallocAsLocal].lvType = TYP_BLK;
14399 lvaTable[stackallocAsLocal].lvExactSize = (unsigned)allocSize;
14400 lvaTable[stackallocAsLocal].lvIsUnsafeBuffer = true;
14401 op1 = gtNewLclvNode(stackallocAsLocal, TYP_BLK);
14402 op1 = gtNewOperNode(GT_ADDR, TYP_I_IMPL, op1);
14403 convertedToLocal = true;
14404 compGSReorderStackLayout = true;
14409 if (!convertedToLocal)
14411 // Bail out if inlining and the localloc was not converted.
14413 // Note we might consider allowing the inline, if the call
14414 // site is not in a loop.
14415 if (compIsForInlining())
14417 InlineObservation obs = op2->IsIntegralConst()
14418 ? InlineObservation::CALLEE_LOCALLOC_TOO_LARGE
14419 : InlineObservation::CALLSITE_LOCALLOC_SIZE_UNKNOWN;
14420 compInlineResult->NoteFatal(obs);
14424 op1 = gtNewOperNode(GT_LCLHEAP, TYP_I_IMPL, op2);
14425 // May throw a stack overflow exception. Obviously, we don't want locallocs to be CSE'd.
14426 op1->gtFlags |= (GTF_EXCEPT | GTF_DONT_CSE);
14428 /* The FP register may not be back to the original value at the end
14429 of the method, even if the frame size is 0, as localloc may
14430 have modified it. So we will HAVE to reset it */
14431 compLocallocUsed = true;
14435 compLocallocOptimized = true;
14439 impPushOnStack(op1, tiRetVal);
14444 /* Get the type token */
14445 assertImp(sz == sizeof(unsigned));
14447 _impResolveToken(CORINFO_TOKENKIND_Casting);
14449 JITDUMP(" %08X", resolvedToken.token);
14451 if (!opts.IsReadyToRun())
14453 op2 = impTokenToHandle(&resolvedToken, nullptr, FALSE);
14454 if (op2 == nullptr)
14455 { // compDonotInline()
14460 if (tiVerificationNeeded)
14462 Verify(impStackTop().seTypeInfo.IsObjRef(), "obj reference needed");
14463 // Even if this is a value class, we know it is boxed.
14464 tiRetVal = typeInfo(TI_REF, resolvedToken.hClass);
14466 accessAllowedResult =
14467 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
14468 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
14470 op1 = impPopStack().val;
14472 GenTree* optTree = impOptimizeCastClassOrIsInst(op1, &resolvedToken, false);
14474 if (optTree != nullptr)
14476 impPushOnStack(optTree, tiRetVal);
14481 #ifdef FEATURE_READYTORUN_COMPILER
14482 if (opts.IsReadyToRun())
14484 GenTreeCall* opLookup =
14485 impReadyToRunHelperToTree(&resolvedToken, CORINFO_HELP_READYTORUN_ISINSTANCEOF, TYP_REF,
14486 gtNewArgList(op1));
14487 usingReadyToRunHelper = (opLookup != nullptr);
14488 op1 = (usingReadyToRunHelper ? opLookup : op1);
14490 if (!usingReadyToRunHelper)
14492 // TODO: ReadyToRun: When generic dictionary lookups are necessary, replace the lookup call
14493 // and the isinstanceof_any call with a single call to a dynamic R2R cell that will:
14494 // 1) Load the context
14495 // 2) Perform the generic dictionary lookup and caching, and generate the appropriate
14497 // 3) Perform the 'is instance' check on the input object
14498 // Reason: performance (today, we'll always use the slow helper for the R2R generics case)
14500 op2 = impTokenToHandle(&resolvedToken, nullptr, FALSE);
14501 if (op2 == nullptr)
14502 { // compDonotInline()
14508 if (!usingReadyToRunHelper)
14511 op1 = impCastClassOrIsInstToTree(op1, op2, &resolvedToken, false);
14513 if (compDonotInline())
14518 impPushOnStack(op1, tiRetVal);
14523 case CEE_REFANYVAL:
14525 // get the class handle and make a ICON node out of it
14527 _impResolveToken(CORINFO_TOKENKIND_Class);
14529 JITDUMP(" %08X", resolvedToken.token);
14531 op2 = impTokenToHandle(&resolvedToken);
14532 if (op2 == nullptr)
14533 { // compDonotInline()
14537 if (tiVerificationNeeded)
14539 Verify(typeInfo::AreEquivalent(impStackTop().seTypeInfo, verMakeTypeInfo(impGetRefAnyClass())),
14541 tiRetVal = verMakeTypeInfo(resolvedToken.hClass).MakeByRef();
14544 op1 = impPopStack().val;
14545 // make certain it is normalized;
14546 op1 = impNormStructVal(op1, impGetRefAnyClass(), (unsigned)CHECK_SPILL_ALL);
14548 // Call helper GETREFANY(classHandle, op1);
14549 args = gtNewArgList(op2, op1);
14550 op1 = gtNewHelperCallNode(CORINFO_HELP_GETREFANY, TYP_BYREF, args);
14552 impPushOnStack(op1, tiRetVal);
14555 case CEE_REFANYTYPE:
14557 if (tiVerificationNeeded)
14559 Verify(typeInfo::AreEquivalent(impStackTop().seTypeInfo, verMakeTypeInfo(impGetRefAnyClass())),
14563 op1 = impPopStack().val;
14565 // make certain it is normalized;
14566 op1 = impNormStructVal(op1, impGetRefAnyClass(), (unsigned)CHECK_SPILL_ALL);
14568 if (op1->gtOper == GT_OBJ)
14570 // Get the address of the refany
14571 op1 = op1->gtOp.gtOp1;
14573 // Fetch the type from the correct slot
14574 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1,
14575 gtNewIconNode(offsetof(CORINFO_RefAny, type), TYP_I_IMPL));
14576 op1 = gtNewOperNode(GT_IND, TYP_BYREF, op1);
14580 assertImp(op1->gtOper == GT_MKREFANY);
14582 // The pointer may have side-effects
14583 if (op1->gtOp.gtOp1->gtFlags & GTF_SIDE_EFFECT)
14585 impAppendTree(op1->gtOp.gtOp1, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
14587 impNoteLastILoffs();
14591 // We already have the class handle
14592 op1 = op1->gtOp.gtOp2;
14595 // convert native TypeHandle to RuntimeTypeHandle
14597 GenTreeArgList* helperArgs = gtNewArgList(op1);
14599 op1 = gtNewHelperCallNode(CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE_MAYBENULL, TYP_STRUCT, helperArgs);
14601 // The handle struct is returned in register
14602 op1->gtCall.gtReturnType = GetRuntimeHandleUnderlyingType();
14604 tiRetVal = typeInfo(TI_STRUCT, impGetTypeHandleClass());
14607 impPushOnStack(op1, tiRetVal);
14612 /* Get the Class index */
14613 assertImp(sz == sizeof(unsigned));
14614 lastLoadToken = codeAddr;
14615 _impResolveToken(CORINFO_TOKENKIND_Ldtoken);
14617 tokenType = info.compCompHnd->getTokenTypeAsHandle(&resolvedToken);
14619 op1 = impTokenToHandle(&resolvedToken, nullptr, TRUE);
14620 if (op1 == nullptr)
14621 { // compDonotInline()
14625 helper = CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE;
14626 assert(resolvedToken.hClass != nullptr);
14628 if (resolvedToken.hMethod != nullptr)
14630 helper = CORINFO_HELP_METHODDESC_TO_STUBRUNTIMEMETHOD;
14632 else if (resolvedToken.hField != nullptr)
14634 helper = CORINFO_HELP_FIELDDESC_TO_STUBRUNTIMEFIELD;
14637 GenTreeArgList* helperArgs = gtNewArgList(op1);
14639 op1 = gtNewHelperCallNode(helper, TYP_STRUCT, helperArgs);
14641 // The handle struct is returned in register
14642 op1->gtCall.gtReturnType = GetRuntimeHandleUnderlyingType();
14644 tiRetVal = verMakeTypeInfo(tokenType);
14645 impPushOnStack(op1, tiRetVal);
14650 case CEE_UNBOX_ANY:
14652 /* Get the Class index */
14653 assertImp(sz == sizeof(unsigned));
14655 _impResolveToken(CORINFO_TOKENKIND_Class);
14657 JITDUMP(" %08X", resolvedToken.token);
14659 BOOL runtimeLookup;
14660 op2 = impTokenToHandle(&resolvedToken, &runtimeLookup);
14661 if (op2 == nullptr)
14663 assert(compDonotInline());
14667 // Run this always so we can get access exceptions even with SkipVerification.
14668 accessAllowedResult =
14669 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
14670 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
14672 if (opcode == CEE_UNBOX_ANY && !eeIsValueClass(resolvedToken.hClass))
14674 if (tiVerificationNeeded)
14676 typeInfo tiUnbox = impStackTop().seTypeInfo;
14677 Verify(tiUnbox.IsObjRef(), "bad unbox.any arg");
14678 tiRetVal = verMakeTypeInfo(resolvedToken.hClass);
14679 tiRetVal.NormaliseForStack();
14681 JITDUMP("\n Importing UNBOX.ANY(refClass) as CASTCLASS\n");
14682 op1 = impPopStack().val;
14686 /* Pop the object and create the unbox helper call */
14687 /* You might think that for UNBOX_ANY we need to push a different */
14688 /* (non-byref) type, but here we're making the tiRetVal that is used */
14689 /* for the intermediate pointer which we then transfer onto the OBJ */
14690 /* instruction. OBJ then creates the appropriate tiRetVal. */
14691 if (tiVerificationNeeded)
14693 typeInfo tiUnbox = impStackTop().seTypeInfo;
14694 Verify(tiUnbox.IsObjRef(), "Bad unbox arg");
14696 tiRetVal = verMakeTypeInfo(resolvedToken.hClass);
14697 Verify(tiRetVal.IsValueClass(), "not value class");
14698 tiRetVal.MakeByRef();
14700 // We always come from an objref, so this is safe byref
14701 tiRetVal.SetIsPermanentHomeByRef();
14702 tiRetVal.SetIsReadonlyByRef();
14705 op1 = impPopStack().val;
14706 assertImp(op1->gtType == TYP_REF);
14708 helper = info.compCompHnd->getUnBoxHelper(resolvedToken.hClass);
14709 assert(helper == CORINFO_HELP_UNBOX || helper == CORINFO_HELP_UNBOX_NULLABLE);
14711 // Check legality and profitability of inline expansion for unboxing.
14712 const bool canExpandInline = (helper == CORINFO_HELP_UNBOX);
14713 const bool shouldExpandInline = !(compCurBB->isRunRarely() || opts.compDbgCode || opts.MinOpts());
14715 if (canExpandInline && shouldExpandInline)
14717 JITDUMP("\n Importing %s as inline sequence\n", opcode == CEE_UNBOX ? "UNBOX" : "UNBOX.ANY");
14718 // we are doing normal unboxing
14719 // inline the common case of the unbox helper
14720 // UNBOX(exp) morphs into
14721 // clone = pop(exp);
14722 // ((*clone == typeToken) ? nop : helper(clone, typeToken));
14723 // push(clone + TARGET_POINTER_SIZE)
14725 GenTree* cloneOperand;
14726 op1 = impCloneExpr(op1, &cloneOperand, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
14727 nullptr DEBUGARG("inline UNBOX clone1"));
14728 op1 = gtNewOperNode(GT_IND, TYP_I_IMPL, op1);
14730 GenTree* condBox = gtNewOperNode(GT_EQ, TYP_INT, op1, op2);
14732 op1 = impCloneExpr(cloneOperand, &cloneOperand, NO_CLASS_HANDLE, (unsigned)CHECK_SPILL_ALL,
14733 nullptr DEBUGARG("inline UNBOX clone2"));
14734 op2 = impTokenToHandle(&resolvedToken);
14735 if (op2 == nullptr)
14736 { // compDonotInline()
14739 args = gtNewArgList(op2, op1);
14740 op1 = gtNewHelperCallNode(helper, TYP_VOID, args);
14742 op1 = new (this, GT_COLON) GenTreeColon(TYP_VOID, gtNewNothingNode(), op1);
14743 op1 = gtNewQmarkNode(TYP_VOID, condBox, op1);
14744 condBox->gtFlags |= GTF_RELOP_QMARK;
14746 // QMARK nodes cannot reside on the evaluation stack. Because there
14747 // may be other trees on the evaluation stack that side-effect the
14748 // sources of the UNBOX operation we must spill the stack.
14750 impAppendTree(op1, (unsigned)CHECK_SPILL_ALL, impCurStmtOffs);
14752 // Create the address-expression to reference past the object header
14753 // to the beginning of the value-type. Today this means adjusting
14754 // past the base of the objects vtable field which is pointer sized.
14756 op2 = gtNewIconNode(TARGET_POINTER_SIZE, TYP_I_IMPL);
14757 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, cloneOperand, op2);
14761 JITDUMP("\n Importing %s as helper call because %s\n", opcode == CEE_UNBOX ? "UNBOX" : "UNBOX.ANY",
14762 canExpandInline ? "want smaller code or faster jitting" : "inline expansion not legal");
14764 // Don't optimize, just call the helper and be done with it
14765 args = gtNewArgList(op2, op1);
14767 gtNewHelperCallNode(helper,
14768 (var_types)((helper == CORINFO_HELP_UNBOX) ? TYP_BYREF : TYP_STRUCT), args);
14771 assert(helper == CORINFO_HELP_UNBOX && op1->gtType == TYP_BYREF || // Unbox helper returns a byref.
14772 helper == CORINFO_HELP_UNBOX_NULLABLE &&
14773 varTypeIsStruct(op1) // UnboxNullable helper returns a struct.
14777 ----------------------------------------------------------------------
14780 | \ | CORINFO_HELP_UNBOX | CORINFO_HELP_UNBOX_NULLABLE |
14781 | \ | (which returns a BYREF) | (which returns a STRUCT) | |
14783 |---------------------------------------------------------------------
14784 | UNBOX | push the BYREF | spill the STRUCT to a local, |
14785 | | | push the BYREF to this local |
14786 |---------------------------------------------------------------------
14787 | UNBOX_ANY | push a GT_OBJ of | push the STRUCT |
14788 | | the BYREF | For Linux when the |
14789 | | | struct is returned in two |
14790 | | | registers create a temp |
14791 | | | which address is passed to |
14792 | | | the unbox_nullable helper. |
14793 |---------------------------------------------------------------------
14796 if (opcode == CEE_UNBOX)
14798 if (helper == CORINFO_HELP_UNBOX_NULLABLE)
14800 // Unbox nullable helper returns a struct type.
14801 // We need to spill it to a temp so than can take the address of it.
14802 // Here we need unsafe value cls check, since the address of struct is taken to be used
14803 // further along and potetially be exploitable.
14805 unsigned tmp = lvaGrabTemp(true DEBUGARG("UNBOXing a nullable"));
14806 lvaSetStruct(tmp, resolvedToken.hClass, true /* unsafe value cls check */);
14808 op2 = gtNewLclvNode(tmp, TYP_STRUCT);
14809 op1 = impAssignStruct(op2, op1, resolvedToken.hClass, (unsigned)CHECK_SPILL_ALL);
14810 assert(op1->gtType == TYP_VOID); // We must be assigning the return struct to the temp.
14812 op2 = gtNewLclvNode(tmp, TYP_STRUCT);
14813 op2 = gtNewOperNode(GT_ADDR, TYP_BYREF, op2);
14814 op1 = gtNewOperNode(GT_COMMA, TYP_BYREF, op1, op2);
14817 assert(op1->gtType == TYP_BYREF);
14818 assert(!tiVerificationNeeded || tiRetVal.IsByRef());
14822 assert(opcode == CEE_UNBOX_ANY);
14824 if (helper == CORINFO_HELP_UNBOX)
14826 // Normal unbox helper returns a TYP_BYREF.
14827 impPushOnStack(op1, tiRetVal);
14832 assert(helper == CORINFO_HELP_UNBOX_NULLABLE && "Make sure the helper is nullable!");
14834 #if FEATURE_MULTIREG_RET
14836 if (varTypeIsStruct(op1) && IsMultiRegReturnedType(resolvedToken.hClass))
14838 // Unbox nullable helper returns a TYP_STRUCT.
14839 // For the multi-reg case we need to spill it to a temp so that
14840 // we can pass the address to the unbox_nullable jit helper.
14842 unsigned tmp = lvaGrabTemp(true DEBUGARG("UNBOXing a register returnable nullable"));
14843 lvaTable[tmp].lvIsMultiRegArg = true;
14844 lvaSetStruct(tmp, resolvedToken.hClass, true /* unsafe value cls check */);
14846 op2 = gtNewLclvNode(tmp, TYP_STRUCT);
14847 op1 = impAssignStruct(op2, op1, resolvedToken.hClass, (unsigned)CHECK_SPILL_ALL);
14848 assert(op1->gtType == TYP_VOID); // We must be assigning the return struct to the temp.
14850 op2 = gtNewLclvNode(tmp, TYP_STRUCT);
14851 op2 = gtNewOperNode(GT_ADDR, TYP_BYREF, op2);
14852 op1 = gtNewOperNode(GT_COMMA, TYP_BYREF, op1, op2);
14854 // In this case the return value of the unbox helper is TYP_BYREF.
14855 // Make sure the right type is placed on the operand type stack.
14856 impPushOnStack(op1, tiRetVal);
14858 // Load the struct.
14861 assert(op1->gtType == TYP_BYREF);
14862 assert(!tiVerificationNeeded || tiRetVal.IsByRef());
14868 #endif // !FEATURE_MULTIREG_RET
14871 // If non register passable struct we have it materialized in the RetBuf.
14872 assert(op1->gtType == TYP_STRUCT);
14873 tiRetVal = verMakeTypeInfo(resolvedToken.hClass);
14874 assert(tiRetVal.IsValueClass());
14878 impPushOnStack(op1, tiRetVal);
14884 /* Get the Class index */
14885 assertImp(sz == sizeof(unsigned));
14887 _impResolveToken(CORINFO_TOKENKIND_Box);
14889 JITDUMP(" %08X", resolvedToken.token);
14891 if (tiVerificationNeeded)
14893 typeInfo tiActual = impStackTop().seTypeInfo;
14894 typeInfo tiBox = verMakeTypeInfo(resolvedToken.hClass);
14896 Verify(verIsBoxable(tiBox), "boxable type expected");
14898 // check the class constraints of the boxed type in case we are boxing an uninitialized value
14899 Verify(info.compCompHnd->satisfiesClassConstraints(resolvedToken.hClass),
14900 "boxed type has unsatisfied class constraints");
14902 Verify(tiCompatibleWith(tiActual, tiBox.NormaliseForStack(), true), "type mismatch");
14904 // Observation: the following code introduces a boxed value class on the stack, but,
14905 // according to the ECMA spec, one would simply expect: tiRetVal =
14906 // typeInfo(TI_REF,impGetObjectClass());
14908 // Push the result back on the stack,
14909 // even if clsHnd is a value class we want the TI_REF
14910 // we call back to the EE to get find out what hte type we should push (for nullable<T> we push T)
14911 tiRetVal = typeInfo(TI_REF, info.compCompHnd->getTypeForBox(resolvedToken.hClass));
14914 accessAllowedResult =
14915 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
14916 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
14918 // Note BOX can be used on things that are not value classes, in which
14919 // case we get a NOP. However the verifier's view of the type on the
14920 // stack changes (in generic code a 'T' becomes a 'boxed T')
14921 if (!eeIsValueClass(resolvedToken.hClass))
14923 JITDUMP("\n Importing BOX(refClass) as NOP\n");
14924 verCurrentState.esStack[verCurrentState.esStackDepth - 1].seTypeInfo = tiRetVal;
14928 // Look ahead for unbox.any
14929 if (codeAddr + (sz + 1 + sizeof(mdToken)) <= codeEndp && codeAddr[sz] == CEE_UNBOX_ANY)
14931 CORINFO_RESOLVED_TOKEN unboxResolvedToken;
14933 impResolveToken(codeAddr + (sz + 1), &unboxResolvedToken, CORINFO_TOKENKIND_Class);
14935 // See if the resolved tokens describe types that are equal.
14936 const TypeCompareState compare =
14937 info.compCompHnd->compareTypesForEquality(unboxResolvedToken.hClass, resolvedToken.hClass);
14939 // If so, box/unbox.any is a nop.
14940 if (compare == TypeCompareState::Must)
14942 JITDUMP("\n Importing BOX; UNBOX.ANY as NOP\n");
14943 // Skip the next unbox.any instruction
14944 sz += sizeof(mdToken) + 1;
14949 impImportAndPushBox(&resolvedToken);
14950 if (compDonotInline())
14959 /* Get the Class index */
14960 assertImp(sz == sizeof(unsigned));
14962 _impResolveToken(CORINFO_TOKENKIND_Class);
14964 JITDUMP(" %08X", resolvedToken.token);
14966 if (tiVerificationNeeded)
14968 tiRetVal = typeInfo(TI_INT);
14971 op1 = gtNewIconNode(info.compCompHnd->getClassSize(resolvedToken.hClass));
14972 impPushOnStack(op1, tiRetVal);
14975 case CEE_CASTCLASS:
14977 /* Get the Class index */
14979 assertImp(sz == sizeof(unsigned));
14981 _impResolveToken(CORINFO_TOKENKIND_Casting);
14983 JITDUMP(" %08X", resolvedToken.token);
14985 if (!opts.IsReadyToRun())
14987 op2 = impTokenToHandle(&resolvedToken, nullptr, FALSE);
14988 if (op2 == nullptr)
14989 { // compDonotInline()
14994 if (tiVerificationNeeded)
14996 Verify(impStackTop().seTypeInfo.IsObjRef(), "object ref expected");
14998 tiRetVal = typeInfo(TI_REF, resolvedToken.hClass);
15001 accessAllowedResult =
15002 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
15003 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
15005 op1 = impPopStack().val;
15007 /* Pop the address and create the 'checked cast' helper call */
15009 // At this point we expect typeRef to contain the token, op1 to contain the value being cast,
15010 // and op2 to contain code that creates the type handle corresponding to typeRef
15013 GenTree* optTree = impOptimizeCastClassOrIsInst(op1, &resolvedToken, true);
15015 if (optTree != nullptr)
15017 impPushOnStack(optTree, tiRetVal);
15022 #ifdef FEATURE_READYTORUN_COMPILER
15023 if (opts.IsReadyToRun())
15025 GenTreeCall* opLookup =
15026 impReadyToRunHelperToTree(&resolvedToken, CORINFO_HELP_READYTORUN_CHKCAST, TYP_REF,
15027 gtNewArgList(op1));
15028 usingReadyToRunHelper = (opLookup != nullptr);
15029 op1 = (usingReadyToRunHelper ? opLookup : op1);
15031 if (!usingReadyToRunHelper)
15033 // TODO: ReadyToRun: When generic dictionary lookups are necessary, replace the lookup call
15034 // and the chkcastany call with a single call to a dynamic R2R cell that will:
15035 // 1) Load the context
15036 // 2) Perform the generic dictionary lookup and caching, and generate the appropriate
15038 // 3) Check the object on the stack for the type-cast
15039 // Reason: performance (today, we'll always use the slow helper for the R2R generics case)
15041 op2 = impTokenToHandle(&resolvedToken, nullptr, FALSE);
15042 if (op2 == nullptr)
15043 { // compDonotInline()
15049 if (!usingReadyToRunHelper)
15052 op1 = impCastClassOrIsInstToTree(op1, op2, &resolvedToken, true);
15054 if (compDonotInline())
15059 /* Push the result back on the stack */
15060 impPushOnStack(op1, tiRetVal);
15067 if (compIsForInlining())
15069 // !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
15070 // TODO: Will this be too strict, given that we will inline many basic blocks?
15071 // !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
15073 /* Do we have just the exception on the stack ?*/
15075 if (verCurrentState.esStackDepth != 1)
15077 /* if not, just don't inline the method */
15079 compInlineResult->NoteFatal(InlineObservation::CALLEE_THROW_WITH_INVALID_STACK);
15084 if (tiVerificationNeeded)
15086 tiRetVal = impStackTop().seTypeInfo;
15087 Verify(tiRetVal.IsObjRef(), "object ref expected");
15088 if (verTrackObjCtorInitState && (verCurrentState.thisInitialized != TIS_Init))
15090 Verify(!tiRetVal.IsThisPtr(), "throw uninitialized this");
15094 block->bbSetRunRarely(); // any block with a throw is rare
15095 /* Pop the exception object and create the 'throw' helper call */
15097 op1 = gtNewHelperCallNode(CORINFO_HELP_THROW, TYP_VOID, gtNewArgList(impPopStack().val));
15100 if (verCurrentState.esStackDepth > 0)
15102 impEvalSideEffects();
15105 assert(verCurrentState.esStackDepth == 0);
15111 assert(!compIsForInlining());
15113 if (info.compXcptnsCount == 0)
15115 BADCODE("rethrow outside catch");
15118 if (tiVerificationNeeded)
15120 Verify(block->hasHndIndex(), "rethrow outside catch");
15121 if (block->hasHndIndex())
15123 EHblkDsc* HBtab = ehGetDsc(block->getHndIndex());
15124 Verify(!HBtab->HasFinallyOrFaultHandler(), "rethrow in finally or fault");
15125 if (HBtab->HasFilter())
15127 // we better be in the handler clause part, not the filter part
15128 Verify(jitIsBetween(compCurBB->bbCodeOffs, HBtab->ebdHndBegOffs(), HBtab->ebdHndEndOffs()),
15129 "rethrow in filter");
15134 /* Create the 'rethrow' helper call */
15136 op1 = gtNewHelperCallNode(CORINFO_HELP_RETHROW, TYP_VOID);
15142 assertImp(sz == sizeof(unsigned));
15144 _impResolveToken(CORINFO_TOKENKIND_Class);
15146 JITDUMP(" %08X", resolvedToken.token);
15148 if (tiVerificationNeeded)
15150 typeInfo tiTo = impStackTop().seTypeInfo;
15151 typeInfo tiInstr = verMakeTypeInfo(resolvedToken.hClass);
15153 Verify(tiTo.IsByRef(), "byref expected");
15154 Verify(!tiTo.IsReadonlyByRef(), "write to readonly byref");
15156 Verify(tiCompatibleWith(tiInstr, tiTo.DereferenceByRef(), false),
15157 "type operand incompatible with type of address");
15160 size = info.compCompHnd->getClassSize(resolvedToken.hClass); // Size
15161 op2 = gtNewIconNode(0); // Value
15162 op1 = impPopStack().val; // Dest
15163 op1 = gtNewBlockVal(op1, size);
15164 op1 = gtNewBlkOpNode(op1, op2, size, (prefixFlags & PREFIX_VOLATILE) != 0, false);
15169 if (tiVerificationNeeded)
15171 Verify(false, "bad opcode");
15174 op3 = impPopStack().val; // Size
15175 op2 = impPopStack().val; // Value
15176 op1 = impPopStack().val; // Dest
15178 if (op3->IsCnsIntOrI())
15180 size = (unsigned)op3->AsIntConCommon()->IconValue();
15181 op1 = new (this, GT_BLK) GenTreeBlk(GT_BLK, TYP_STRUCT, op1, size);
15185 op1 = new (this, GT_DYN_BLK) GenTreeDynBlk(op1, op3);
15188 op1 = gtNewBlkOpNode(op1, op2, size, (prefixFlags & PREFIX_VOLATILE) != 0, false);
15194 if (tiVerificationNeeded)
15196 Verify(false, "bad opcode");
15198 op3 = impPopStack().val; // Size
15199 op2 = impPopStack().val; // Src
15200 op1 = impPopStack().val; // Dest
15202 if (op3->IsCnsIntOrI())
15204 size = (unsigned)op3->AsIntConCommon()->IconValue();
15205 op1 = new (this, GT_BLK) GenTreeBlk(GT_BLK, TYP_STRUCT, op1, size);
15209 op1 = new (this, GT_DYN_BLK) GenTreeDynBlk(op1, op3);
15212 if (op2->OperGet() == GT_ADDR)
15214 op2 = op2->gtOp.gtOp1;
15218 op2 = gtNewOperNode(GT_IND, TYP_STRUCT, op2);
15221 op1 = gtNewBlkOpNode(op1, op2, size, (prefixFlags & PREFIX_VOLATILE) != 0, true);
15226 assertImp(sz == sizeof(unsigned));
15228 _impResolveToken(CORINFO_TOKENKIND_Class);
15230 JITDUMP(" %08X", resolvedToken.token);
15232 if (tiVerificationNeeded)
15234 typeInfo tiFrom = impStackTop().seTypeInfo;
15235 typeInfo tiTo = impStackTop(1).seTypeInfo;
15236 typeInfo tiInstr = verMakeTypeInfo(resolvedToken.hClass);
15238 Verify(tiFrom.IsByRef(), "expected byref source");
15239 Verify(tiTo.IsByRef(), "expected byref destination");
15241 Verify(tiCompatibleWith(tiFrom.DereferenceByRef(), tiInstr, false),
15242 "type of source address incompatible with type operand");
15243 Verify(!tiTo.IsReadonlyByRef(), "write to readonly byref");
15244 Verify(tiCompatibleWith(tiInstr, tiTo.DereferenceByRef(), false),
15245 "type operand incompatible with type of destination address");
15248 if (!eeIsValueClass(resolvedToken.hClass))
15250 op1 = impPopStack().val; // address to load from
15252 impBashVarAddrsToI(op1);
15254 assertImp(genActualType(op1->gtType) == TYP_I_IMPL || op1->gtType == TYP_BYREF);
15256 op1 = gtNewOperNode(GT_IND, TYP_REF, op1);
15257 op1->gtFlags |= GTF_EXCEPT | GTF_GLOB_REF;
15259 impPushOnStack(op1, typeInfo());
15260 opcode = CEE_STIND_REF;
15262 goto STIND_POST_VERIFY;
15265 op2 = impPopStack().val; // Src
15266 op1 = impPopStack().val; // Dest
15267 op1 = gtNewCpObjNode(op1, op2, resolvedToken.hClass, ((prefixFlags & PREFIX_VOLATILE) != 0));
15272 assertImp(sz == sizeof(unsigned));
15274 _impResolveToken(CORINFO_TOKENKIND_Class);
15276 JITDUMP(" %08X", resolvedToken.token);
15278 if (eeIsValueClass(resolvedToken.hClass))
15280 lclTyp = TYP_STRUCT;
15287 if (tiVerificationNeeded)
15290 typeInfo tiPtr = impStackTop(1).seTypeInfo;
15292 // Make sure we have a good looking byref
15293 Verify(tiPtr.IsByRef(), "pointer not byref");
15294 Verify(!tiPtr.IsReadonlyByRef(), "write to readonly byref");
15295 if (!tiPtr.IsByRef() || tiPtr.IsReadonlyByRef())
15297 compUnsafeCastUsed = true;
15300 typeInfo ptrVal = DereferenceByRef(tiPtr);
15301 typeInfo argVal = verMakeTypeInfo(resolvedToken.hClass);
15303 if (!tiCompatibleWith(impStackTop(0).seTypeInfo, NormaliseForStack(argVal), true))
15305 Verify(false, "type of value incompatible with type operand");
15306 compUnsafeCastUsed = true;
15309 if (!tiCompatibleWith(argVal, ptrVal, false))
15311 Verify(false, "type operand incompatible with type of address");
15312 compUnsafeCastUsed = true;
15317 compUnsafeCastUsed = true;
15320 if (lclTyp == TYP_REF)
15322 opcode = CEE_STIND_REF;
15323 goto STIND_POST_VERIFY;
15326 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(resolvedToken.hClass);
15327 if (impIsPrimitive(jitTyp))
15329 lclTyp = JITtype2varType(jitTyp);
15330 goto STIND_POST_VERIFY;
15333 op2 = impPopStack().val; // Value
15334 op1 = impPopStack().val; // Ptr
15336 assertImp(varTypeIsStruct(op2));
15338 op1 = impAssignStructPtr(op1, op2, resolvedToken.hClass, (unsigned)CHECK_SPILL_ALL);
15340 if (op1->OperIsBlkOp() && (prefixFlags & PREFIX_UNALIGNED))
15342 op1->gtFlags |= GTF_BLK_UNALIGNED;
15349 assert(!compIsForInlining());
15351 // Being lazy here. Refanys are tricky in terms of gc tracking.
15352 // Since it is uncommon, just don't perform struct promotion in any method that contains mkrefany.
15354 JITDUMP("disabling struct promotion because of mkrefany\n");
15355 fgNoStructPromotion = true;
15357 oper = GT_MKREFANY;
15358 assertImp(sz == sizeof(unsigned));
15360 _impResolveToken(CORINFO_TOKENKIND_Class);
15362 JITDUMP(" %08X", resolvedToken.token);
15364 op2 = impTokenToHandle(&resolvedToken, nullptr, TRUE);
15365 if (op2 == nullptr)
15366 { // compDonotInline()
15370 if (tiVerificationNeeded)
15372 typeInfo tiPtr = impStackTop().seTypeInfo;
15373 typeInfo tiInstr = verMakeTypeInfo(resolvedToken.hClass);
15375 Verify(!verIsByRefLike(tiInstr), "mkrefany of byref-like class");
15376 Verify(!tiPtr.IsReadonlyByRef(), "readonly byref used with mkrefany");
15377 Verify(typeInfo::AreEquivalent(tiPtr.DereferenceByRef(), tiInstr), "type mismatch");
15380 accessAllowedResult =
15381 info.compCompHnd->canAccessClass(&resolvedToken, info.compMethodHnd, &calloutHelper);
15382 impHandleAccessAllowed(accessAllowedResult, &calloutHelper);
15384 op1 = impPopStack().val;
15386 // @SPECVIOLATION: TYP_INT should not be allowed here by a strict reading of the spec.
15387 // But JIT32 allowed it, so we continue to allow it.
15388 assertImp(op1->TypeGet() == TYP_BYREF || op1->TypeGet() == TYP_I_IMPL || op1->TypeGet() == TYP_INT);
15390 // MKREFANY returns a struct. op2 is the class token.
15391 op1 = gtNewOperNode(oper, TYP_STRUCT, op1, op2);
15393 impPushOnStack(op1, verMakeTypeInfo(impGetRefAnyClass()));
15399 assertImp(sz == sizeof(unsigned));
15401 _impResolveToken(CORINFO_TOKENKIND_Class);
15403 JITDUMP(" %08X", resolvedToken.token);
15407 tiRetVal = verMakeTypeInfo(resolvedToken.hClass);
15409 if (tiVerificationNeeded)
15411 typeInfo tiPtr = impStackTop().seTypeInfo;
15413 // Make sure we have a byref
15414 if (!tiPtr.IsByRef())
15416 Verify(false, "pointer not byref");
15417 compUnsafeCastUsed = true;
15419 typeInfo tiPtrVal = DereferenceByRef(tiPtr);
15421 if (!tiCompatibleWith(tiPtrVal, tiRetVal, false))
15423 Verify(false, "type of address incompatible with type operand");
15424 compUnsafeCastUsed = true;
15426 tiRetVal.NormaliseForStack();
15430 compUnsafeCastUsed = true;
15433 if (eeIsValueClass(resolvedToken.hClass))
15435 lclTyp = TYP_STRUCT;
15440 opcode = CEE_LDIND_REF;
15441 goto LDIND_POST_VERIFY;
15444 op1 = impPopStack().val;
15446 assertImp(op1->TypeGet() == TYP_BYREF || op1->TypeGet() == TYP_I_IMPL);
15448 CorInfoType jitTyp = info.compCompHnd->asCorInfoType(resolvedToken.hClass);
15449 if (impIsPrimitive(jitTyp))
15451 op1 = gtNewOperNode(GT_IND, JITtype2varType(jitTyp), op1);
15453 // Could point anywhere, example a boxed class static int
15454 op1->gtFlags |= GTF_IND_TGTANYWHERE | GTF_GLOB_REF;
15455 assertImp(varTypeIsArithmetic(op1->gtType));
15459 // OBJ returns a struct
15460 // and an inline argument which is the class token of the loaded obj
15461 op1 = gtNewObjNode(resolvedToken.hClass, op1);
15463 op1->gtFlags |= GTF_EXCEPT;
15465 if (prefixFlags & PREFIX_UNALIGNED)
15467 op1->gtFlags |= GTF_IND_UNALIGNED;
15470 impPushOnStack(op1, tiRetVal);
15475 if (tiVerificationNeeded)
15477 typeInfo tiArray = impStackTop().seTypeInfo;
15478 Verify(verIsSDArray(tiArray), "bad array");
15479 tiRetVal = typeInfo(TI_INT);
15482 op1 = impPopStack().val;
15483 if (!opts.MinOpts() && !opts.compDbgCode)
15485 /* Use GT_ARR_LENGTH operator so rng check opts see this */
15486 GenTreeArrLen* arrLen = gtNewArrLen(TYP_INT, op1, offsetof(CORINFO_Array, length));
15488 /* Mark the block as containing a length expression */
15490 if (op1->gtOper == GT_LCL_VAR)
15492 block->bbFlags |= BBF_HAS_IDX_LEN;
15499 /* Create the expression "*(array_addr + ArrLenOffs)" */
15500 op1 = gtNewOperNode(GT_ADD, TYP_BYREF, op1,
15501 gtNewIconNode(offsetof(CORINFO_Array, length), TYP_I_IMPL));
15502 op1 = gtNewIndir(TYP_INT, op1);
15503 op1->gtFlags |= GTF_IND_ARR_LEN;
15506 /* Push the result back on the stack */
15507 impPushOnStack(op1, tiRetVal);
15511 op1 = gtNewHelperCallNode(CORINFO_HELP_USER_BREAKPOINT, TYP_VOID);
15515 if (opts.compDbgCode)
15517 op1 = new (this, GT_NO_OP) GenTree(GT_NO_OP, TYP_VOID);
15522 /******************************** NYI *******************************/
15525 OutputDebugStringA("CLR: Invalid x86 breakpoint in IL stream\n");
15528 case CEE_MACRO_END:
15531 BADCODE3("unknown opcode", ": %02X", (int)opcode);
15535 prevOpcode = opcode;
15541 #undef _impResolveToken
15544 #pragma warning(pop)
15547 // Push a local/argument treeon the operand stack
15548 void Compiler::impPushVar(GenTree* op, typeInfo tiRetVal)
15550 tiRetVal.NormaliseForStack();
15552 if (verTrackObjCtorInitState && (verCurrentState.thisInitialized != TIS_Init) && tiRetVal.IsThisPtr())
15554 tiRetVal.SetUninitialisedObjRef();
15557 impPushOnStack(op, tiRetVal);
15560 // Load a local/argument on the operand stack
15561 // lclNum is an index into lvaTable *NOT* the arg/lcl index in the IL
15562 void Compiler::impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal)
15566 if (lvaTable[lclNum].lvNormalizeOnLoad())
15568 lclTyp = lvaGetRealType(lclNum);
15572 lclTyp = lvaGetActualType(lclNum);
15575 impPushVar(gtNewLclvNode(lclNum, lclTyp, offset), tiRetVal);
15578 // Load an argument on the operand stack
15579 // Shared by the various CEE_LDARG opcodes
15580 // ilArgNum is the argument index as specified in IL.
15581 // It will be mapped to the correct lvaTable index
15582 void Compiler::impLoadArg(unsigned ilArgNum, IL_OFFSET offset)
15584 Verify(ilArgNum < info.compILargsCount, "bad arg num");
15586 if (compIsForInlining())
15588 if (ilArgNum >= info.compArgsCount)
15590 compInlineResult->NoteFatal(InlineObservation::CALLEE_BAD_ARGUMENT_NUMBER);
15594 impPushVar(impInlineFetchArg(ilArgNum, impInlineInfo->inlArgInfo, impInlineInfo->lclVarInfo),
15595 impInlineInfo->lclVarInfo[ilArgNum].lclVerTypeInfo);
15599 if (ilArgNum >= info.compArgsCount)
15604 unsigned lclNum = compMapILargNum(ilArgNum); // account for possible hidden param
15606 if (lclNum == info.compThisArg)
15608 lclNum = lvaArg0Var;
15611 impLoadVar(lclNum, offset);
15615 // Load a local on the operand stack
15616 // Shared by the various CEE_LDLOC opcodes
15617 // ilLclNum is the local index as specified in IL.
15618 // It will be mapped to the correct lvaTable index
15619 void Compiler::impLoadLoc(unsigned ilLclNum, IL_OFFSET offset)
15621 if (tiVerificationNeeded)
15623 Verify(ilLclNum < info.compMethodInfo->locals.numArgs, "bad loc num");
15624 Verify(info.compInitMem, "initLocals not set");
15627 if (compIsForInlining())
15629 if (ilLclNum >= info.compMethodInfo->locals.numArgs)
15631 compInlineResult->NoteFatal(InlineObservation::CALLEE_BAD_LOCAL_NUMBER);
15635 // Get the local type
15636 var_types lclTyp = impInlineInfo->lclVarInfo[ilLclNum + impInlineInfo->argCnt].lclTypeInfo;
15638 typeInfo tiRetVal = impInlineInfo->lclVarInfo[ilLclNum + impInlineInfo->argCnt].lclVerTypeInfo;
15640 /* Have we allocated a temp for this local? */
15642 unsigned lclNum = impInlineFetchLocal(ilLclNum DEBUGARG("Inline ldloc first use temp"));
15644 // All vars of inlined methods should be !lvNormalizeOnLoad()
15646 assert(!lvaTable[lclNum].lvNormalizeOnLoad());
15647 lclTyp = genActualType(lclTyp);
15649 impPushVar(gtNewLclvNode(lclNum, lclTyp), tiRetVal);
15653 if (ilLclNum >= info.compMethodInfo->locals.numArgs)
15658 unsigned lclNum = info.compArgsCount + ilLclNum;
15660 impLoadVar(lclNum, offset);
15664 #ifdef _TARGET_ARM_
15665 /**************************************************************************************
15667 * When assigning a vararg call src to a HFA lcl dest, mark that we cannot promote the
15668 * dst struct, because struct promotion will turn it into a float/double variable while
15669 * the rhs will be an int/long variable. We don't code generate assignment of int into
15670 * a float, but there is nothing that might prevent us from doing so. The tree however
15671 * would like: (=, (typ_float, typ_int)) or (GT_TRANSFER, (typ_float, typ_int))
15673 * tmpNum - the lcl dst variable num that is a struct.
15674 * src - the src tree assigned to the dest that is a struct/int (when varargs call.)
15675 * hClass - the type handle for the struct variable.
15677 * TODO-ARM-CQ: [301608] This is a rare scenario with varargs and struct promotion coming into play,
15678 * however, we could do a codegen of transferring from int to float registers
15679 * (transfer, not a cast.)
15682 void Compiler::impMarkLclDstNotPromotable(unsigned tmpNum, GenTree* src, CORINFO_CLASS_HANDLE hClass)
15684 if (src->gtOper == GT_CALL && src->gtCall.IsVarargs() && IsHfa(hClass))
15686 int hfaSlots = GetHfaCount(hClass);
15687 var_types hfaType = GetHfaType(hClass);
15689 // If we have varargs we morph the method's return type to be "int" irrespective of its original
15690 // type: struct/float at importer because the ABI calls out return in integer registers.
15691 // We don't want struct promotion to replace an expression like this:
15692 // lclFld_int = callvar_int() into lclFld_float = callvar_int();
15693 // This means an int is getting assigned to a float without a cast. Prevent the promotion.
15694 if ((hfaType == TYP_DOUBLE && hfaSlots == sizeof(double) / REGSIZE_BYTES) ||
15695 (hfaType == TYP_FLOAT && hfaSlots == sizeof(float) / REGSIZE_BYTES))
15697 // Make sure this struct type stays as struct so we can receive the call in a struct.
15698 lvaTable[tmpNum].lvIsMultiRegRet = true;
15702 #endif // _TARGET_ARM_
15704 #if FEATURE_MULTIREG_RET
15705 GenTree* Compiler::impAssignMultiRegTypeToVar(GenTree* op, CORINFO_CLASS_HANDLE hClass)
15707 unsigned tmpNum = lvaGrabTemp(true DEBUGARG("Return value temp for multireg return."));
15708 impAssignTempGen(tmpNum, op, hClass, (unsigned)CHECK_SPILL_ALL);
15709 GenTree* ret = gtNewLclvNode(tmpNum, lvaTable[tmpNum].lvType);
15711 // TODO-1stClassStructs: Handle constant propagation and CSE-ing of multireg returns.
15712 ret->gtFlags |= GTF_DONT_CSE;
15714 assert(IsMultiRegReturnedType(hClass));
15716 // Mark the var so that fields are not promoted and stay together.
15717 lvaTable[tmpNum].lvIsMultiRegRet = true;
15721 #endif // FEATURE_MULTIREG_RET
15723 // do import for a return
15724 // returns false if inlining was aborted
15725 // opcode can be ret or call in the case of a tail.call
15726 bool Compiler::impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode)
15728 if (tiVerificationNeeded)
15730 verVerifyThisPtrInitialised();
15732 unsigned expectedStack = 0;
15733 if (info.compRetType != TYP_VOID)
15735 typeInfo tiVal = impStackTop().seTypeInfo;
15736 typeInfo tiDeclared =
15737 verMakeTypeInfo(info.compMethodInfo->args.retType, info.compMethodInfo->args.retTypeClass);
15739 Verify(!verIsByRefLike(tiDeclared) || verIsSafeToReturnByRef(tiVal), "byref return");
15741 Verify(tiCompatibleWith(tiVal, tiDeclared.NormaliseForStack(), true), "type mismatch");
15744 Verify(verCurrentState.esStackDepth == expectedStack, "stack non-empty on return");
15748 // If we are importing an inlinee and have GC ref locals we always
15749 // need to have a spill temp for the return value. This temp
15750 // should have been set up in advance, over in fgFindBasicBlocks.
15751 if (compIsForInlining() && impInlineInfo->HasGcRefLocals() && (info.compRetType != TYP_VOID))
15753 assert(lvaInlineeReturnSpillTemp != BAD_VAR_NUM);
15757 GenTree* op2 = nullptr;
15758 GenTree* op1 = nullptr;
15759 CORINFO_CLASS_HANDLE retClsHnd = nullptr;
15761 if (info.compRetType != TYP_VOID)
15763 StackEntry se = impPopStack();
15764 retClsHnd = se.seTypeInfo.GetClassHandle();
15767 if (!compIsForInlining())
15769 impBashVarAddrsToI(op2);
15770 op2 = impImplicitIorI4Cast(op2, info.compRetType);
15771 op2 = impImplicitR4orR8Cast(op2, info.compRetType);
15772 assertImp((genActualType(op2->TypeGet()) == genActualType(info.compRetType)) ||
15773 ((op2->TypeGet() == TYP_I_IMPL) && (info.compRetType == TYP_BYREF)) ||
15774 ((op2->TypeGet() == TYP_BYREF) && (info.compRetType == TYP_I_IMPL)) ||
15775 (varTypeIsFloating(op2->gtType) && varTypeIsFloating(info.compRetType)) ||
15776 (varTypeIsStruct(op2) && varTypeIsStruct(info.compRetType)));
15779 if (opts.compGcChecks && info.compRetType == TYP_REF)
15781 // DDB 3483 : JIT Stress: early termination of GC ref's life time in exception code path
15782 // VSW 440513: Incorrect gcinfo on the return value under COMPlus_JitGCChecks=1 for methods with
15785 assert(op2->gtType == TYP_REF);
15787 // confirm that the argument is a GC pointer (for debugging (GC stress))
15788 GenTreeArgList* args = gtNewArgList(op2);
15789 op2 = gtNewHelperCallNode(CORINFO_HELP_CHECK_OBJ, TYP_REF, args);
15793 printf("\ncompGcChecks tree:\n");
15801 // inlinee's stack should be empty now.
15802 assert(verCurrentState.esStackDepth == 0);
15807 printf("\n\n Inlinee Return expression (before normalization) =>\n");
15812 // Make sure the type matches the original call.
15814 var_types returnType = genActualType(op2->gtType);
15815 var_types originalCallType = impInlineInfo->inlineCandidateInfo->fncRetType;
15816 if ((returnType != originalCallType) && (originalCallType == TYP_STRUCT))
15818 originalCallType = impNormStructType(impInlineInfo->inlineCandidateInfo->methInfo.args.retTypeClass);
15821 if (returnType != originalCallType)
15823 JITDUMP("Return type mismatch, have %s, needed %s\n", varTypeName(returnType),
15824 varTypeName(originalCallType));
15825 compInlineResult->NoteFatal(InlineObservation::CALLSITE_RETURN_TYPE_MISMATCH);
15829 // Below, we are going to set impInlineInfo->retExpr to the tree with the return
15830 // expression. At this point, retExpr could already be set if there are multiple
15831 // return blocks (meaning fgNeedReturnSpillTemp() == true) and one of
15832 // the other blocks already set it. If there is only a single return block,
15833 // retExpr shouldn't be set. However, this is not true if we reimport a block
15834 // with a return. In that case, retExpr will be set, then the block will be
15835 // reimported, but retExpr won't get cleared as part of setting the block to
15836 // be reimported. The reimported retExpr value should be the same, so even if
15837 // we don't unconditionally overwrite it, it shouldn't matter.
15838 if (info.compRetNativeType != TYP_STRUCT)
15840 // compRetNativeType is not TYP_STRUCT.
15841 // This implies it could be either a scalar type or SIMD vector type or
15842 // a struct type that can be normalized to a scalar type.
15844 if (varTypeIsStruct(info.compRetType))
15846 noway_assert(info.compRetBuffArg == BAD_VAR_NUM);
15847 // adjust the type away from struct to integral
15848 // and no normalizing
15849 op2 = impFixupStructReturnType(op2, retClsHnd);
15853 // Do we have to normalize?
15854 var_types fncRealRetType = JITtype2varType(info.compMethodInfo->args.retType);
15855 if ((varTypeIsSmall(op2->TypeGet()) || varTypeIsSmall(fncRealRetType)) &&
15856 fgCastNeeded(op2, fncRealRetType))
15858 // Small-typed return values are normalized by the callee
15859 op2 = gtNewCastNode(TYP_INT, op2, false, fncRealRetType);
15863 if (fgNeedReturnSpillTemp())
15865 assert(info.compRetNativeType != TYP_VOID &&
15866 (fgMoreThanOneReturnBlock() || impInlineInfo->HasGcRefLocals()));
15868 // If this method returns a ref type, track the actual types seen
15870 if (info.compRetType == TYP_REF)
15872 bool isExact = false;
15873 bool isNonNull = false;
15874 CORINFO_CLASS_HANDLE returnClsHnd = gtGetClassHandle(op2, &isExact, &isNonNull);
15876 if (impInlineInfo->retExpr == nullptr)
15878 // This is the first return, so best known type is the type
15879 // of this return value.
15880 impInlineInfo->retExprClassHnd = returnClsHnd;
15881 impInlineInfo->retExprClassHndIsExact = isExact;
15883 else if (impInlineInfo->retExprClassHnd != returnClsHnd)
15885 // This return site type differs from earlier seen sites,
15886 // so reset the info and we'll fall back to using the method's
15887 // declared return type for the return spill temp.
15888 impInlineInfo->retExprClassHnd = nullptr;
15889 impInlineInfo->retExprClassHndIsExact = false;
15893 // This is a bit of a workaround...
15894 // If we are inlining a call that returns a struct, where the actual "native" return type is
15895 // not a struct (for example, the struct is composed of exactly one int, and the native
15896 // return type is thus an int), and the inlinee has multiple return blocks (thus,
15897 // fgNeedReturnSpillTemp() == true, and is the index of a local var that is set
15898 // to the *native* return type), and at least one of the return blocks is the result of
15899 // a call, then we have a problem. The situation is like this (from a failed test case):
15902 // // Note: valuetype plinq_devtests.LazyTests/LIX is a struct with only a single int
15903 // call !!0 [mscorlib]System.Threading.LazyInitializer::EnsureInitialized<valuetype
15904 // plinq_devtests.LazyTests/LIX>(!!0&, bool&, object&, class [mscorlib]System.Func`1<!!0>)
15908 // ldobj !!T // this gets bashed to a GT_LCL_FLD, type TYP_INT
15911 // call !!0 System.Threading.LazyInitializer::EnsureInitializedCore<!!0>(!!0&, bool&,
15912 // object&, class System.Func`1<!!0>)
15915 // In the code above, when we call impFixupStructReturnType(), we will change the op2 return type
15916 // of the inlinee return node, but we don't do that for GT_CALL nodes, which we delay until
15917 // morphing when we call fgFixupStructReturn(). We do this, apparently, to handle nested
15918 // inlining properly by leaving the correct type on the GT_CALL node through importing.
15920 // To fix this, for this case, we temporarily change the GT_CALL node type to the
15921 // native return type, which is what it will be set to eventually. We generate the
15922 // assignment to the return temp, using the correct type, and then restore the GT_CALL
15923 // node type. During morphing, the GT_CALL will get the correct, final, native return type.
15925 bool restoreType = false;
15926 if ((op2->OperGet() == GT_CALL) && (info.compRetType == TYP_STRUCT))
15928 noway_assert(op2->TypeGet() == TYP_STRUCT);
15929 op2->gtType = info.compRetNativeType;
15930 restoreType = true;
15933 impAssignTempGen(lvaInlineeReturnSpillTemp, op2, se.seTypeInfo.GetClassHandle(),
15934 (unsigned)CHECK_SPILL_ALL);
15936 GenTree* tmpOp2 = gtNewLclvNode(lvaInlineeReturnSpillTemp, op2->TypeGet());
15940 op2->gtType = TYP_STRUCT; // restore it to what it was
15946 if (impInlineInfo->retExpr)
15948 // Some other block(s) have seen the CEE_RET first.
15949 // Better they spilled to the same temp.
15950 assert(impInlineInfo->retExpr->gtOper == GT_LCL_VAR);
15951 assert(impInlineInfo->retExpr->gtLclVarCommon.gtLclNum == op2->gtLclVarCommon.gtLclNum);
15959 printf("\n\n Inlinee Return expression (after normalization) =>\n");
15964 // Report the return expression
15965 impInlineInfo->retExpr = op2;
15969 // compRetNativeType is TYP_STRUCT.
15970 // This implies that struct return via RetBuf arg or multi-reg struct return
15972 GenTreeCall* iciCall = impInlineInfo->iciCall->AsCall();
15974 // Assign the inlinee return into a spill temp.
15975 // spill temp only exists if there are multiple return points
15976 if (lvaInlineeReturnSpillTemp != BAD_VAR_NUM)
15978 // in this case we have to insert multiple struct copies to the temp
15979 // and the retexpr is just the temp.
15980 assert(info.compRetNativeType != TYP_VOID);
15981 assert(fgMoreThanOneReturnBlock() || impInlineInfo->HasGcRefLocals());
15983 impAssignTempGen(lvaInlineeReturnSpillTemp, op2, se.seTypeInfo.GetClassHandle(),
15984 (unsigned)CHECK_SPILL_ALL);
15987 #if defined(_TARGET_ARM_) || defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
15988 #if defined(_TARGET_ARM_)
15989 // TODO-ARM64-NYI: HFA
15990 // TODO-AMD64-Unix and TODO-ARM once the ARM64 functionality is implemented the
15991 // next ifdefs could be refactored in a single method with the ifdef inside.
15992 if (IsHfa(retClsHnd))
15994 // Same as !IsHfa but just don't bother with impAssignStructPtr.
15995 #else // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
15996 ReturnTypeDesc retTypeDesc;
15997 retTypeDesc.InitializeStructReturnType(this, retClsHnd);
15998 unsigned retRegCount = retTypeDesc.GetReturnRegCount();
16000 if (retRegCount != 0)
16002 // If single eightbyte, the return type would have been normalized and there won't be a temp var.
16003 // This code will be called only if the struct return has not been normalized (i.e. 2 eightbytes -
16005 assert(retRegCount == MAX_RET_REG_COUNT);
16006 // Same as !structDesc.passedInRegisters but just don't bother with impAssignStructPtr.
16007 CLANG_FORMAT_COMMENT_ANCHOR;
16008 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
16010 if (fgNeedReturnSpillTemp())
16012 if (!impInlineInfo->retExpr)
16014 #if defined(_TARGET_ARM_)
16015 impInlineInfo->retExpr = gtNewLclvNode(lvaInlineeReturnSpillTemp, info.compRetType);
16016 #else // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
16017 // The inlinee compiler has figured out the type of the temp already. Use it here.
16018 impInlineInfo->retExpr =
16019 gtNewLclvNode(lvaInlineeReturnSpillTemp, lvaTable[lvaInlineeReturnSpillTemp].lvType);
16020 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
16025 impInlineInfo->retExpr = op2;
16029 #elif defined(_TARGET_ARM64_)
16030 ReturnTypeDesc retTypeDesc;
16031 retTypeDesc.InitializeStructReturnType(this, retClsHnd);
16032 unsigned retRegCount = retTypeDesc.GetReturnRegCount();
16034 if (retRegCount != 0)
16036 assert(!iciCall->HasRetBufArg());
16037 assert(retRegCount >= 2);
16038 if (fgNeedReturnSpillTemp())
16040 if (!impInlineInfo->retExpr)
16042 // The inlinee compiler has figured out the type of the temp already. Use it here.
16043 impInlineInfo->retExpr =
16044 gtNewLclvNode(lvaInlineeReturnSpillTemp, lvaTable[lvaInlineeReturnSpillTemp].lvType);
16049 impInlineInfo->retExpr = op2;
16053 #endif // defined(_TARGET_ARM64_)
16055 assert(iciCall->HasRetBufArg());
16056 GenTree* dest = gtCloneExpr(iciCall->gtCallArgs->gtOp.gtOp1);
16057 // spill temp only exists if there are multiple return points
16058 if (fgNeedReturnSpillTemp())
16060 // if this is the first return we have seen set the retExpr
16061 if (!impInlineInfo->retExpr)
16063 impInlineInfo->retExpr =
16064 impAssignStructPtr(dest, gtNewLclvNode(lvaInlineeReturnSpillTemp, info.compRetType),
16065 retClsHnd, (unsigned)CHECK_SPILL_ALL);
16070 impInlineInfo->retExpr = impAssignStructPtr(dest, op2, retClsHnd, (unsigned)CHECK_SPILL_ALL);
16077 if (compIsForInlining())
16082 if (info.compRetType == TYP_VOID)
16085 op1 = new (this, GT_RETURN) GenTreeOp(GT_RETURN, TYP_VOID);
16087 else if (info.compRetBuffArg != BAD_VAR_NUM)
16089 // Assign value to return buff (first param)
16090 GenTree* retBuffAddr = gtNewLclvNode(info.compRetBuffArg, TYP_BYREF, impCurStmtOffs);
16092 op2 = impAssignStructPtr(retBuffAddr, op2, retClsHnd, (unsigned)CHECK_SPILL_ALL);
16093 impAppendTree(op2, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
16095 // There are cases where the address of the implicit RetBuf should be returned explicitly (in RAX).
16096 CLANG_FORMAT_COMMENT_ANCHOR;
16098 #if defined(_TARGET_AMD64_)
16100 // x64 (System V and Win64) calling convention requires to
16101 // return the implicit return buffer explicitly (in RAX).
16102 // Change the return type to be BYREF.
16103 op1 = gtNewOperNode(GT_RETURN, TYP_BYREF, gtNewLclvNode(info.compRetBuffArg, TYP_BYREF));
16104 #else // !defined(_TARGET_AMD64_)
16105 // In case of non-AMD64 targets the profiler hook requires to return the implicit RetBuf explicitly (in RAX).
16106 // In such case the return value of the function is changed to BYREF.
16107 // If profiler hook is not needed the return type of the function is TYP_VOID.
16108 if (compIsProfilerHookNeeded())
16110 op1 = gtNewOperNode(GT_RETURN, TYP_BYREF, gtNewLclvNode(info.compRetBuffArg, TYP_BYREF));
16115 op1 = new (this, GT_RETURN) GenTreeOp(GT_RETURN, TYP_VOID);
16117 #endif // !defined(_TARGET_AMD64_)
16119 else if (varTypeIsStruct(info.compRetType))
16121 #if !FEATURE_MULTIREG_RET
16122 // For both ARM architectures the HFA native types are maintained as structs.
16123 // Also on System V AMD64 the multireg structs returns are also left as structs.
16124 noway_assert(info.compRetNativeType != TYP_STRUCT);
16126 op2 = impFixupStructReturnType(op2, retClsHnd);
16128 op1 = gtNewOperNode(GT_RETURN, genActualType(info.compRetNativeType), op2);
16133 op1 = gtNewOperNode(GT_RETURN, genActualType(info.compRetType), op2);
16136 // We must have imported a tailcall and jumped to RET
16137 if (prefixFlags & PREFIX_TAILCALL)
16139 #if defined(FEATURE_CORECLR) || !defined(_TARGET_AMD64_)
16141 // This cannot be asserted on Amd64 since we permit the following IL pattern:
16145 assert(verCurrentState.esStackDepth == 0 && impOpcodeIsCallOpcode(opcode));
16146 #endif // FEATURE_CORECLR || !_TARGET_AMD64_
16148 opcode = CEE_RET; // To prevent trying to spill if CALL_SITE_BOUNDARIES
16150 // impImportCall() would have already appended TYP_VOID calls
16151 if (info.compRetType == TYP_VOID)
16157 impAppendTree(op1, (unsigned)CHECK_SPILL_NONE, impCurStmtOffs);
16159 // Remember at which BC offset the tree was finished
16160 impNoteLastILoffs();
16165 /*****************************************************************************
16166 * Mark the block as unimported.
16167 * Note that the caller is responsible for calling impImportBlockPending(),
16168 * with the appropriate stack-state
16171 inline void Compiler::impReimportMarkBlock(BasicBlock* block)
16174 if (verbose && (block->bbFlags & BBF_IMPORTED))
16176 printf("\nBB%02u will be reimported\n", block->bbNum);
16180 block->bbFlags &= ~BBF_IMPORTED;
16183 /*****************************************************************************
16184 * Mark the successors of the given block as unimported.
16185 * Note that the caller is responsible for calling impImportBlockPending()
16186 * for all the successors, with the appropriate stack-state.
16189 void Compiler::impReimportMarkSuccessors(BasicBlock* block)
16191 const unsigned numSuccs = block->NumSucc();
16192 for (unsigned i = 0; i < numSuccs; i++)
16194 impReimportMarkBlock(block->GetSucc(i));
16198 /*****************************************************************************
16200 * Filter wrapper to handle only passed in exception code
16204 LONG FilterVerificationExceptions(PEXCEPTION_POINTERS pExceptionPointers, LPVOID lpvParam)
16206 if (pExceptionPointers->ExceptionRecord->ExceptionCode == SEH_VERIFICATION_EXCEPTION)
16208 return EXCEPTION_EXECUTE_HANDLER;
16211 return EXCEPTION_CONTINUE_SEARCH;
16214 void Compiler::impVerifyEHBlock(BasicBlock* block, bool isTryStart)
16216 assert(block->hasTryIndex());
16217 assert(!compIsForInlining());
16219 unsigned tryIndex = block->getTryIndex();
16220 EHblkDsc* HBtab = ehGetDsc(tryIndex);
16224 assert(block->bbFlags & BBF_TRY_BEG);
16226 // The Stack must be empty
16228 if (block->bbStkDepth != 0)
16230 BADCODE("Evaluation stack must be empty on entry into a try block");
16234 // Save the stack contents, we'll need to restore it later
16236 SavedStack blockState;
16237 impSaveStackState(&blockState, false);
16239 while (HBtab != nullptr)
16243 // Are we verifying that an instance constructor properly initializes it's 'this' pointer once?
16244 // We do not allow the 'this' pointer to be uninitialized when entering most kinds try regions
16246 if (verTrackObjCtorInitState && (verCurrentState.thisInitialized != TIS_Init))
16248 // We trigger an invalid program exception here unless we have a try/fault region.
16250 if (HBtab->HasCatchHandler() || HBtab->HasFinallyHandler() || HBtab->HasFilter())
16253 "The 'this' pointer of an instance constructor is not intialized upon entry to a try region");
16257 // Allow a try/fault region to proceed.
16258 assert(HBtab->HasFaultHandler());
16262 /* Recursively process the handler block */
16263 BasicBlock* hndBegBB = HBtab->ebdHndBeg;
16265 // Construct the proper verification stack state
16266 // either empty or one that contains just
16267 // the Exception Object that we are dealing with
16269 verCurrentState.esStackDepth = 0;
16271 if (handlerGetsXcptnObj(hndBegBB->bbCatchTyp))
16273 CORINFO_CLASS_HANDLE clsHnd;
16275 if (HBtab->HasFilter())
16277 clsHnd = impGetObjectClass();
16281 CORINFO_RESOLVED_TOKEN resolvedToken;
16283 resolvedToken.tokenContext = impTokenLookupContextHandle;
16284 resolvedToken.tokenScope = info.compScopeHnd;
16285 resolvedToken.token = HBtab->ebdTyp;
16286 resolvedToken.tokenType = CORINFO_TOKENKIND_Class;
16287 info.compCompHnd->resolveToken(&resolvedToken);
16289 clsHnd = resolvedToken.hClass;
16292 // push catch arg the stack, spill to a temp if necessary
16293 // Note: can update HBtab->ebdHndBeg!
16294 hndBegBB = impPushCatchArgOnStack(hndBegBB, clsHnd, false);
16297 // Queue up the handler for importing
16299 impImportBlockPending(hndBegBB);
16301 if (HBtab->HasFilter())
16303 /* @VERIFICATION : Ideally the end of filter state should get
16304 propagated to the catch handler, this is an incompleteness,
16305 but is not a security/compliance issue, since the only
16306 interesting state is the 'thisInit' state.
16309 verCurrentState.esStackDepth = 0;
16311 BasicBlock* filterBB = HBtab->ebdFilter;
16313 // push catch arg the stack, spill to a temp if necessary
16314 // Note: can update HBtab->ebdFilter!
16315 const bool isSingleBlockFilter = (filterBB->bbNext == hndBegBB);
16316 filterBB = impPushCatchArgOnStack(filterBB, impGetObjectClass(), isSingleBlockFilter);
16318 impImportBlockPending(filterBB);
16321 else if (verTrackObjCtorInitState && HBtab->HasFaultHandler())
16323 /* Recursively process the handler block */
16325 verCurrentState.esStackDepth = 0;
16327 // Queue up the fault handler for importing
16329 impImportBlockPending(HBtab->ebdHndBeg);
16332 // Now process our enclosing try index (if any)
16334 tryIndex = HBtab->ebdEnclosingTryIndex;
16335 if (tryIndex == EHblkDsc::NO_ENCLOSING_INDEX)
16341 HBtab = ehGetDsc(tryIndex);
16345 // Restore the stack contents
16346 impRestoreStackState(&blockState);
16349 //***************************************************************
16350 // Import the instructions for the given basic block. Perform
16351 // verification, throwing an exception on failure. Push any successor blocks that are enabled for the first
16352 // time, or whose verification pre-state is changed.
16355 #pragma warning(push)
16356 #pragma warning(disable : 21000) // Suppress PREFast warning about overly large function
16358 void Compiler::impImportBlock(BasicBlock* block)
16360 // BBF_INTERNAL blocks only exist during importation due to EH canonicalization. We need to
16361 // handle them specially. In particular, there is no IL to import for them, but we do need
16362 // to mark them as imported and put their successors on the pending import list.
16363 if (block->bbFlags & BBF_INTERNAL)
16365 JITDUMP("Marking BBF_INTERNAL block BB%02u as BBF_IMPORTED\n", block->bbNum);
16366 block->bbFlags |= BBF_IMPORTED;
16368 const unsigned numSuccs = block->NumSucc();
16369 for (unsigned i = 0; i < numSuccs; i++)
16371 impImportBlockPending(block->GetSucc(i));
16381 /* Make the block globaly available */
16386 /* Initialize the debug variables */
16387 impCurOpcName = "unknown";
16388 impCurOpcOffs = block->bbCodeOffs;
16391 /* Set the current stack state to the merged result */
16392 verResetCurrentState(block, &verCurrentState);
16394 /* Now walk the code and import the IL into GenTrees */
16396 struct FilterVerificationExceptionsParam
16401 FilterVerificationExceptionsParam param;
16403 param.pThis = this;
16404 param.block = block;
16406 PAL_TRY(FilterVerificationExceptionsParam*, pParam, ¶m)
16408 /* @VERIFICATION : For now, the only state propagation from try
16409 to it's handler is "thisInit" state (stack is empty at start of try).
16410 In general, for state that we track in verification, we need to
16411 model the possibility that an exception might happen at any IL
16412 instruction, so we really need to merge all states that obtain
16413 between IL instructions in a try block into the start states of
16416 However we do not allow the 'this' pointer to be uninitialized when
16417 entering most kinds try regions (only try/fault are allowed to have
16418 an uninitialized this pointer on entry to the try)
16420 Fortunately, the stack is thrown away when an exception
16421 leads to a handler, so we don't have to worry about that.
16422 We DO, however, have to worry about the "thisInit" state.
16423 But only for the try/fault case.
16425 The only allowed transition is from TIS_Uninit to TIS_Init.
16427 So for a try/fault region for the fault handler block
16428 we will merge the start state of the try begin
16429 and the post-state of each block that is part of this try region
16432 // merge the start state of the try begin
16434 if (pParam->block->bbFlags & BBF_TRY_BEG)
16436 pParam->pThis->impVerifyEHBlock(pParam->block, true);
16439 pParam->pThis->impImportBlockCode(pParam->block);
16441 // As discussed above:
16442 // merge the post-state of each block that is part of this try region
16444 if (pParam->block->hasTryIndex())
16446 pParam->pThis->impVerifyEHBlock(pParam->block, false);
16449 PAL_EXCEPT_FILTER(FilterVerificationExceptions)
16451 verHandleVerificationFailure(block DEBUGARG(false));
16455 if (compDonotInline())
16460 assert(!compDonotInline());
16462 markImport = false;
16466 unsigned baseTmp = NO_BASE_TMP; // input temps assigned to successor blocks
16467 bool reimportSpillClique = false;
16468 BasicBlock* tgtBlock = nullptr;
16470 /* If the stack is non-empty, we might have to spill its contents */
16472 if (verCurrentState.esStackDepth != 0)
16474 impBoxTemp = BAD_VAR_NUM; // if a box temp is used in a block that leaves something
16475 // on the stack, its lifetime is hard to determine, simply
16476 // don't reuse such temps.
16478 GenTree* addStmt = nullptr;
16480 /* Do the successors of 'block' have any other predecessors ?
16481 We do not want to do some of the optimizations related to multiRef
16482 if we can reimport blocks */
16484 unsigned multRef = impCanReimport ? unsigned(~0) : 0;
16486 switch (block->bbJumpKind)
16490 /* Temporarily remove the 'jtrue' from the end of the tree list */
16492 assert(impTreeLast);
16493 assert(impTreeLast->gtOper == GT_STMT);
16494 assert(impTreeLast->gtStmt.gtStmtExpr->gtOper == GT_JTRUE);
16496 addStmt = impTreeLast;
16497 impTreeLast = impTreeLast->gtPrev;
16499 /* Note if the next block has more than one ancestor */
16501 multRef |= block->bbNext->bbRefs;
16503 /* Does the next block have temps assigned? */
16505 baseTmp = block->bbNext->bbStkTempsIn;
16506 tgtBlock = block->bbNext;
16508 if (baseTmp != NO_BASE_TMP)
16513 /* Try the target of the jump then */
16515 multRef |= block->bbJumpDest->bbRefs;
16516 baseTmp = block->bbJumpDest->bbStkTempsIn;
16517 tgtBlock = block->bbJumpDest;
16521 multRef |= block->bbJumpDest->bbRefs;
16522 baseTmp = block->bbJumpDest->bbStkTempsIn;
16523 tgtBlock = block->bbJumpDest;
16527 multRef |= block->bbNext->bbRefs;
16528 baseTmp = block->bbNext->bbStkTempsIn;
16529 tgtBlock = block->bbNext;
16534 BasicBlock** jmpTab;
16537 /* Temporarily remove the GT_SWITCH from the end of the tree list */
16539 assert(impTreeLast);
16540 assert(impTreeLast->gtOper == GT_STMT);
16541 assert(impTreeLast->gtStmt.gtStmtExpr->gtOper == GT_SWITCH);
16543 addStmt = impTreeLast;
16544 impTreeLast = impTreeLast->gtPrev;
16546 jmpCnt = block->bbJumpSwt->bbsCount;
16547 jmpTab = block->bbJumpSwt->bbsDstTab;
16551 tgtBlock = (*jmpTab);
16553 multRef |= tgtBlock->bbRefs;
16555 // Thanks to spill cliques, we should have assigned all or none
16556 assert((baseTmp == NO_BASE_TMP) || (baseTmp == tgtBlock->bbStkTempsIn));
16557 baseTmp = tgtBlock->bbStkTempsIn;
16562 } while (++jmpTab, --jmpCnt);
16566 case BBJ_CALLFINALLY:
16567 case BBJ_EHCATCHRET:
16569 case BBJ_EHFINALLYRET:
16570 case BBJ_EHFILTERRET:
16572 NO_WAY("can't have 'unreached' end of BB with non-empty stack");
16576 noway_assert(!"Unexpected bbJumpKind");
16580 assert(multRef >= 1);
16582 /* Do we have a base temp number? */
16584 bool newTemps = (baseTmp == NO_BASE_TMP);
16588 /* Grab enough temps for the whole stack */
16589 baseTmp = impGetSpillTmpBase(block);
16592 /* Spill all stack entries into temps */
16593 unsigned level, tempNum;
16595 JITDUMP("\nSpilling stack entries into temps\n");
16596 for (level = 0, tempNum = baseTmp; level < verCurrentState.esStackDepth; level++, tempNum++)
16598 GenTree* tree = verCurrentState.esStack[level].val;
16600 /* VC generates code where it pushes a byref from one branch, and an int (ldc.i4 0) from
16601 the other. This should merge to a byref in unverifiable code.
16602 However, if the branch which leaves the TYP_I_IMPL on the stack is imported first, the
16603 successor would be imported assuming there was a TYP_I_IMPL on
16604 the stack. Thus the value would not get GC-tracked. Hence,
16605 change the temp to TYP_BYREF and reimport the successors.
16606 Note: We should only allow this in unverifiable code.
16608 if (tree->gtType == TYP_BYREF && lvaTable[tempNum].lvType == TYP_I_IMPL && !verNeedsVerification())
16610 lvaTable[tempNum].lvType = TYP_BYREF;
16611 impReimportMarkSuccessors(block);
16615 #ifdef _TARGET_64BIT_
16616 if (genActualType(tree->gtType) == TYP_I_IMPL && lvaTable[tempNum].lvType == TYP_INT)
16618 if (tiVerificationNeeded && tgtBlock->bbEntryState != nullptr &&
16619 (tgtBlock->bbFlags & BBF_FAILED_VERIFICATION) == 0)
16621 // Merge the current state into the entry state of block;
16622 // the call to verMergeEntryStates must have changed
16623 // the entry state of the block by merging the int local var
16624 // and the native-int stack entry.
16625 bool changed = false;
16626 if (verMergeEntryStates(tgtBlock, &changed))
16628 impRetypeEntryStateTemps(tgtBlock);
16629 impReimportBlockPending(tgtBlock);
16634 tgtBlock->bbFlags |= BBF_FAILED_VERIFICATION;
16639 // Some other block in the spill clique set this to "int", but now we have "native int".
16640 // Change the type and go back to re-import any blocks that used the wrong type.
16641 lvaTable[tempNum].lvType = TYP_I_IMPL;
16642 reimportSpillClique = true;
16644 else if (genActualType(tree->gtType) == TYP_INT && lvaTable[tempNum].lvType == TYP_I_IMPL)
16646 // Spill clique has decided this should be "native int", but this block only pushes an "int".
16647 // Insert a sign-extension to "native int" so we match the clique.
16648 verCurrentState.esStack[level].val = gtNewCastNode(TYP_I_IMPL, tree, false, TYP_I_IMPL);
16651 // Consider the case where one branch left a 'byref' on the stack and the other leaves
16652 // an 'int'. On 32-bit, this is allowed (in non-verifiable code) since they are the same
16653 // size. JIT64 managed to make this work on 64-bit. For compatibility, we support JIT64
16654 // behavior instead of asserting and then generating bad code (where we save/restore the
16655 // low 32 bits of a byref pointer to an 'int' sized local). If the 'int' side has been
16656 // imported already, we need to change the type of the local and reimport the spill clique.
16657 // If the 'byref' side has imported, we insert a cast from int to 'native int' to match
16658 // the 'byref' size.
16659 if (!tiVerificationNeeded)
16661 if (genActualType(tree->gtType) == TYP_BYREF && lvaTable[tempNum].lvType == TYP_INT)
16663 // Some other block in the spill clique set this to "int", but now we have "byref".
16664 // Change the type and go back to re-import any blocks that used the wrong type.
16665 lvaTable[tempNum].lvType = TYP_BYREF;
16666 reimportSpillClique = true;
16668 else if (genActualType(tree->gtType) == TYP_INT && lvaTable[tempNum].lvType == TYP_BYREF)
16670 // Spill clique has decided this should be "byref", but this block only pushes an "int".
16671 // Insert a sign-extension to "native int" so we match the clique size.
16672 verCurrentState.esStack[level].val = gtNewCastNode(TYP_I_IMPL, tree, false, TYP_I_IMPL);
16675 #endif // _TARGET_64BIT_
16677 #if FEATURE_X87_DOUBLES
16678 // X87 stack doesn't differentiate between float/double
16679 // so promoting is no big deal.
16680 // For everybody else keep it as float until we have a collision and then promote
16681 // Just like for x64's TYP_INT<->TYP_I_IMPL
16683 if (multRef > 1 && tree->gtType == TYP_FLOAT)
16685 verCurrentState.esStack[level].val = gtNewCastNode(TYP_DOUBLE, tree, TYP_DOUBLE);
16688 #else // !FEATURE_X87_DOUBLES
16690 if (tree->gtType == TYP_DOUBLE && lvaTable[tempNum].lvType == TYP_FLOAT)
16692 // Some other block in the spill clique set this to "float", but now we have "double".
16693 // Change the type and go back to re-import any blocks that used the wrong type.
16694 lvaTable[tempNum].lvType = TYP_DOUBLE;
16695 reimportSpillClique = true;
16697 else if (tree->gtType == TYP_FLOAT && lvaTable[tempNum].lvType == TYP_DOUBLE)
16699 // Spill clique has decided this should be "double", but this block only pushes a "float".
16700 // Insert a cast to "double" so we match the clique.
16701 verCurrentState.esStack[level].val = gtNewCastNode(TYP_DOUBLE, tree, false, TYP_DOUBLE);
16704 #endif // FEATURE_X87_DOUBLES
16706 /* If addStmt has a reference to tempNum (can only happen if we
16707 are spilling to the temps already used by a previous block),
16708 we need to spill addStmt */
16710 if (addStmt && !newTemps && gtHasRef(addStmt->gtStmt.gtStmtExpr, tempNum, false))
16712 GenTree* addTree = addStmt->gtStmt.gtStmtExpr;
16714 if (addTree->gtOper == GT_JTRUE)
16716 GenTree* relOp = addTree->gtOp.gtOp1;
16717 assert(relOp->OperIsCompare());
16719 var_types type = genActualType(relOp->gtOp.gtOp1->TypeGet());
16721 if (gtHasRef(relOp->gtOp.gtOp1, tempNum, false))
16723 unsigned temp = lvaGrabTemp(true DEBUGARG("spill addStmt JTRUE ref Op1"));
16724 impAssignTempGen(temp, relOp->gtOp.gtOp1, level);
16725 type = genActualType(lvaTable[temp].TypeGet());
16726 relOp->gtOp.gtOp1 = gtNewLclvNode(temp, type);
16729 if (gtHasRef(relOp->gtOp.gtOp2, tempNum, false))
16731 unsigned temp = lvaGrabTemp(true DEBUGARG("spill addStmt JTRUE ref Op2"));
16732 impAssignTempGen(temp, relOp->gtOp.gtOp2, level);
16733 type = genActualType(lvaTable[temp].TypeGet());
16734 relOp->gtOp.gtOp2 = gtNewLclvNode(temp, type);
16739 assert(addTree->gtOper == GT_SWITCH && genActualTypeIsIntOrI(addTree->gtOp.gtOp1->TypeGet()));
16741 unsigned temp = lvaGrabTemp(true DEBUGARG("spill addStmt SWITCH"));
16742 impAssignTempGen(temp, addTree->gtOp.gtOp1, level);
16743 addTree->gtOp.gtOp1 = gtNewLclvNode(temp, genActualType(addTree->gtOp.gtOp1->TypeGet()));
16747 /* Spill the stack entry, and replace with the temp */
16749 if (!impSpillStackEntry(level, tempNum
16752 true, "Spill Stack Entry"
16758 BADCODE("bad stack state");
16761 // Oops. Something went wrong when spilling. Bad code.
16762 verHandleVerificationFailure(block DEBUGARG(true));
16768 /* Put back the 'jtrue'/'switch' if we removed it earlier */
16772 impAppendStmt(addStmt, (unsigned)CHECK_SPILL_NONE);
16776 // Some of the append/spill logic works on compCurBB
16778 assert(compCurBB == block);
16780 /* Save the tree list in the block */
16781 impEndTreeList(block);
16783 // impEndTreeList sets BBF_IMPORTED on the block
16784 // We do *NOT* want to set it later than this because
16785 // impReimportSpillClique might clear it if this block is both a
16786 // predecessor and successor in the current spill clique
16787 assert(block->bbFlags & BBF_IMPORTED);
16789 // If we had a int/native int, or float/double collision, we need to re-import
16790 if (reimportSpillClique)
16792 // This will re-import all the successors of block (as well as each of their predecessors)
16793 impReimportSpillClique(block);
16795 // For blocks that haven't been imported yet, we still need to mark them as pending import.
16796 const unsigned numSuccs = block->NumSucc();
16797 for (unsigned i = 0; i < numSuccs; i++)
16799 BasicBlock* succ = block->GetSucc(i);
16800 if ((succ->bbFlags & BBF_IMPORTED) == 0)
16802 impImportBlockPending(succ);
16806 else // the normal case
16808 // otherwise just import the successors of block
16810 /* Does this block jump to any other blocks? */
16811 const unsigned numSuccs = block->NumSucc();
16812 for (unsigned i = 0; i < numSuccs; i++)
16814 impImportBlockPending(block->GetSucc(i));
16819 #pragma warning(pop)
16822 /*****************************************************************************/
16824 // Ensures that "block" is a member of the list of BBs waiting to be imported, pushing it on the list if
16825 // necessary (and ensures that it is a member of the set of BB's on the list, by setting its byte in
16826 // impPendingBlockMembers). Merges the current verification state into the verification state of "block"
16827 // (its "pre-state").
16829 void Compiler::impImportBlockPending(BasicBlock* block)
16834 printf("\nimpImportBlockPending for BB%02u\n", block->bbNum);
16838 // We will add a block to the pending set if it has not already been imported (or needs to be re-imported),
16839 // or if it has, but merging in a predecessor's post-state changes the block's pre-state.
16840 // (When we're doing verification, we always attempt the merge to detect verification errors.)
16842 // If the block has not been imported, add to pending set.
16843 bool addToPending = ((block->bbFlags & BBF_IMPORTED) == 0);
16845 // Initialize bbEntryState just the first time we try to add this block to the pending list
16846 // Just because bbEntryState is NULL, doesn't mean the pre-state wasn't previously set
16847 // We use NULL to indicate the 'common' state to avoid memory allocation
16848 if ((block->bbEntryState == nullptr) && ((block->bbFlags & (BBF_IMPORTED | BBF_FAILED_VERIFICATION)) == 0) &&
16849 (impGetPendingBlockMember(block) == 0))
16851 verInitBBEntryState(block, &verCurrentState);
16852 assert(block->bbStkDepth == 0);
16853 block->bbStkDepth = static_cast<unsigned short>(verCurrentState.esStackDepth);
16854 assert(addToPending);
16855 assert(impGetPendingBlockMember(block) == 0);
16859 // The stack should have the same height on entry to the block from all its predecessors.
16860 if (block->bbStkDepth != verCurrentState.esStackDepth)
16864 sprintf_s(buffer, sizeof(buffer),
16865 "Block at offset %4.4x to %4.4x in %s entered with different stack depths.\n"
16866 "Previous depth was %d, current depth is %d",
16867 block->bbCodeOffs, block->bbCodeOffsEnd, info.compFullName, block->bbStkDepth,
16868 verCurrentState.esStackDepth);
16869 buffer[400 - 1] = 0;
16872 NO_WAY("Block entered with different stack depths");
16876 // Additionally, if we need to verify, merge the verification state.
16877 if (tiVerificationNeeded)
16879 // Merge the current state into the entry state of block; if this does not change the entry state
16880 // by merging, do not add the block to the pending-list.
16881 bool changed = false;
16882 if (!verMergeEntryStates(block, &changed))
16884 block->bbFlags |= BBF_FAILED_VERIFICATION;
16885 addToPending = true; // We will pop it off, and check the flag set above.
16889 addToPending = true;
16891 JITDUMP("Adding BB%02u to pending set due to new merge result\n", block->bbNum);
16900 if (block->bbStkDepth > 0)
16902 // We need to fix the types of any spill temps that might have changed:
16903 // int->native int, float->double, int->byref, etc.
16904 impRetypeEntryStateTemps(block);
16907 // OK, we must add to the pending list, if it's not already in it.
16908 if (impGetPendingBlockMember(block) != 0)
16914 // Get an entry to add to the pending list
16918 if (impPendingFree)
16920 // We can reuse one of the freed up dscs.
16921 dsc = impPendingFree;
16922 impPendingFree = dsc->pdNext;
16926 // We have to create a new dsc
16927 dsc = new (this, CMK_Unknown) PendingDsc;
16931 dsc->pdSavedStack.ssDepth = verCurrentState.esStackDepth;
16932 dsc->pdThisPtrInit = verCurrentState.thisInitialized;
16934 // Save the stack trees for later
16936 if (verCurrentState.esStackDepth)
16938 impSaveStackState(&dsc->pdSavedStack, false);
16941 // Add the entry to the pending list
16943 dsc->pdNext = impPendingList;
16944 impPendingList = dsc;
16945 impSetPendingBlockMember(block, 1); // And indicate that it's now a member of the set.
16947 // Various assertions require us to now to consider the block as not imported (at least for
16948 // the final time...)
16949 block->bbFlags &= ~BBF_IMPORTED;
16954 printf("Added PendingDsc - %08p for BB%02u\n", dspPtr(dsc), block->bbNum);
16959 /*****************************************************************************/
16961 // Ensures that "block" is a member of the list of BBs waiting to be imported, pushing it on the list if
16962 // necessary (and ensures that it is a member of the set of BB's on the list, by setting its byte in
16963 // impPendingBlockMembers). Does *NOT* change the existing "pre-state" of the block.
16965 void Compiler::impReimportBlockPending(BasicBlock* block)
16967 JITDUMP("\nimpReimportBlockPending for BB%02u", block->bbNum);
16969 assert(block->bbFlags & BBF_IMPORTED);
16971 // OK, we must add to the pending list, if it's not already in it.
16972 if (impGetPendingBlockMember(block) != 0)
16977 // Get an entry to add to the pending list
16981 if (impPendingFree)
16983 // We can reuse one of the freed up dscs.
16984 dsc = impPendingFree;
16985 impPendingFree = dsc->pdNext;
16989 // We have to create a new dsc
16990 dsc = new (this, CMK_ImpStack) PendingDsc;
16995 if (block->bbEntryState)
16997 dsc->pdThisPtrInit = block->bbEntryState->thisInitialized;
16998 dsc->pdSavedStack.ssDepth = block->bbEntryState->esStackDepth;
16999 dsc->pdSavedStack.ssTrees = block->bbEntryState->esStack;
17003 dsc->pdThisPtrInit = TIS_Bottom;
17004 dsc->pdSavedStack.ssDepth = 0;
17005 dsc->pdSavedStack.ssTrees = nullptr;
17008 // Add the entry to the pending list
17010 dsc->pdNext = impPendingList;
17011 impPendingList = dsc;
17012 impSetPendingBlockMember(block, 1); // And indicate that it's now a member of the set.
17014 // Various assertions require us to now to consider the block as not imported (at least for
17015 // the final time...)
17016 block->bbFlags &= ~BBF_IMPORTED;
17021 printf("Added PendingDsc - %08p for BB%02u\n", dspPtr(dsc), block->bbNum);
17026 void* Compiler::BlockListNode::operator new(size_t sz, Compiler* comp)
17028 if (comp->impBlockListNodeFreeList == nullptr)
17030 return (BlockListNode*)comp->compGetMem(sizeof(BlockListNode), CMK_BasicBlock);
17034 BlockListNode* res = comp->impBlockListNodeFreeList;
17035 comp->impBlockListNodeFreeList = res->m_next;
17040 void Compiler::FreeBlockListNode(Compiler::BlockListNode* node)
17042 node->m_next = impBlockListNodeFreeList;
17043 impBlockListNodeFreeList = node;
17046 void Compiler::impWalkSpillCliqueFromPred(BasicBlock* block, SpillCliqueWalker* callback)
17050 noway_assert(!fgComputePredsDone);
17051 if (!fgCheapPredsValid)
17053 fgComputeCheapPreds();
17056 BlockListNode* succCliqueToDo = nullptr;
17057 BlockListNode* predCliqueToDo = new (this) BlockListNode(block);
17061 // Look at the successors of every member of the predecessor to-do list.
17062 while (predCliqueToDo != nullptr)
17064 BlockListNode* node = predCliqueToDo;
17065 predCliqueToDo = node->m_next;
17066 BasicBlock* blk = node->m_blk;
17067 FreeBlockListNode(node);
17069 const unsigned numSuccs = blk->NumSucc();
17070 for (unsigned succNum = 0; succNum < numSuccs; succNum++)
17072 BasicBlock* succ = blk->GetSucc(succNum);
17073 // If it's not already in the clique, add it, and also add it
17074 // as a member of the successor "toDo" set.
17075 if (impSpillCliqueGetMember(SpillCliqueSucc, succ) == 0)
17077 callback->Visit(SpillCliqueSucc, succ);
17078 impSpillCliqueSetMember(SpillCliqueSucc, succ, 1);
17079 succCliqueToDo = new (this) BlockListNode(succ, succCliqueToDo);
17084 // Look at the predecessors of every member of the successor to-do list.
17085 while (succCliqueToDo != nullptr)
17087 BlockListNode* node = succCliqueToDo;
17088 succCliqueToDo = node->m_next;
17089 BasicBlock* blk = node->m_blk;
17090 FreeBlockListNode(node);
17092 for (BasicBlockList* pred = blk->bbCheapPreds; pred != nullptr; pred = pred->next)
17094 BasicBlock* predBlock = pred->block;
17095 // If it's not already in the clique, add it, and also add it
17096 // as a member of the predecessor "toDo" set.
17097 if (impSpillCliqueGetMember(SpillCliquePred, predBlock) == 0)
17099 callback->Visit(SpillCliquePred, predBlock);
17100 impSpillCliqueSetMember(SpillCliquePred, predBlock, 1);
17101 predCliqueToDo = new (this) BlockListNode(predBlock, predCliqueToDo);
17108 // If this fails, it means we didn't walk the spill clique properly and somehow managed
17109 // miss walking back to include the predecessor we started from.
17110 // This most likely cause: missing or out of date bbPreds
17111 assert(impSpillCliqueGetMember(SpillCliquePred, block) != 0);
17114 void Compiler::SetSpillTempsBase::Visit(SpillCliqueDir predOrSucc, BasicBlock* blk)
17116 if (predOrSucc == SpillCliqueSucc)
17118 assert(blk->bbStkTempsIn == NO_BASE_TMP); // Should not already be a member of a clique as a successor.
17119 blk->bbStkTempsIn = m_baseTmp;
17123 assert(predOrSucc == SpillCliquePred);
17124 assert(blk->bbStkTempsOut == NO_BASE_TMP); // Should not already be a member of a clique as a predecessor.
17125 blk->bbStkTempsOut = m_baseTmp;
17129 void Compiler::ReimportSpillClique::Visit(SpillCliqueDir predOrSucc, BasicBlock* blk)
17131 // For Preds we could be a little smarter and just find the existing store
17132 // and re-type it/add a cast, but that is complicated and hopefully very rare, so
17133 // just re-import the whole block (just like we do for successors)
17135 if (((blk->bbFlags & BBF_IMPORTED) == 0) && (m_pComp->impGetPendingBlockMember(blk) == 0))
17137 // If we haven't imported this block and we're not going to (because it isn't on
17138 // the pending list) then just ignore it for now.
17140 // This block has either never been imported (EntryState == NULL) or it failed
17141 // verification. Neither state requires us to force it to be imported now.
17142 assert((blk->bbEntryState == nullptr) || (blk->bbFlags & BBF_FAILED_VERIFICATION));
17146 // For successors we have a valid verCurrentState, so just mark them for reimport
17147 // the 'normal' way
17148 // Unlike predecessors, we *DO* need to reimport the current block because the
17149 // initial import had the wrong entry state types.
17150 // Similarly, blocks that are currently on the pending list, still need to call
17151 // impImportBlockPending to fixup their entry state.
17152 if (predOrSucc == SpillCliqueSucc)
17154 m_pComp->impReimportMarkBlock(blk);
17156 // Set the current stack state to that of the blk->bbEntryState
17157 m_pComp->verResetCurrentState(blk, &m_pComp->verCurrentState);
17158 assert(m_pComp->verCurrentState.thisInitialized == blk->bbThisOnEntry());
17160 m_pComp->impImportBlockPending(blk);
17162 else if ((blk != m_pComp->compCurBB) && ((blk->bbFlags & BBF_IMPORTED) != 0))
17164 // As described above, we are only visiting predecessors so they can
17165 // add the appropriate casts, since we have already done that for the current
17166 // block, it does not need to be reimported.
17167 // Nor do we need to reimport blocks that are still pending, but not yet
17170 // For predecessors, we have no state to seed the EntryState, so we just have
17171 // to assume the existing one is correct.
17172 // If the block is also a successor, it will get the EntryState properly
17173 // updated when it is visited as a successor in the above "if" block.
17174 assert(predOrSucc == SpillCliquePred);
17175 m_pComp->impReimportBlockPending(blk);
17179 // Re-type the incoming lclVar nodes to match the varDsc.
17180 void Compiler::impRetypeEntryStateTemps(BasicBlock* blk)
17182 if (blk->bbEntryState != nullptr)
17184 EntryState* es = blk->bbEntryState;
17185 for (unsigned level = 0; level < es->esStackDepth; level++)
17187 GenTree* tree = es->esStack[level].val;
17188 if ((tree->gtOper == GT_LCL_VAR) || (tree->gtOper == GT_LCL_FLD))
17190 unsigned lclNum = tree->gtLclVarCommon.gtLclNum;
17191 noway_assert(lclNum < lvaCount);
17192 LclVarDsc* varDsc = lvaTable + lclNum;
17193 es->esStack[level].val->gtType = varDsc->TypeGet();
17199 unsigned Compiler::impGetSpillTmpBase(BasicBlock* block)
17201 if (block->bbStkTempsOut != NO_BASE_TMP)
17203 return block->bbStkTempsOut;
17209 printf("\n*************** In impGetSpillTmpBase(BB%02u)\n", block->bbNum);
17213 // Otherwise, choose one, and propagate to all members of the spill clique.
17214 // Grab enough temps for the whole stack.
17215 unsigned baseTmp = lvaGrabTemps(verCurrentState.esStackDepth DEBUGARG("IL Stack Entries"));
17216 SetSpillTempsBase callback(baseTmp);
17218 // We do *NOT* need to reset the SpillClique*Members because a block can only be the predecessor
17219 // to one spill clique, and similarly can only be the sucessor to one spill clique
17220 impWalkSpillCliqueFromPred(block, &callback);
17225 void Compiler::impReimportSpillClique(BasicBlock* block)
17230 printf("\n*************** In impReimportSpillClique(BB%02u)\n", block->bbNum);
17234 // If we get here, it is because this block is already part of a spill clique
17235 // and one predecessor had an outgoing live stack slot of type int, and this
17236 // block has an outgoing live stack slot of type native int.
17237 // We need to reset these before traversal because they have already been set
17238 // by the previous walk to determine all the members of the spill clique.
17239 impInlineRoot()->impSpillCliquePredMembers.Reset();
17240 impInlineRoot()->impSpillCliqueSuccMembers.Reset();
17242 ReimportSpillClique callback(this);
17244 impWalkSpillCliqueFromPred(block, &callback);
17247 // Set the pre-state of "block" (which should not have a pre-state allocated) to
17248 // a copy of "srcState", cloning tree pointers as required.
17249 void Compiler::verInitBBEntryState(BasicBlock* block, EntryState* srcState)
17251 if (srcState->esStackDepth == 0 && srcState->thisInitialized == TIS_Bottom)
17253 block->bbEntryState = nullptr;
17257 block->bbEntryState = (EntryState*)compGetMem(sizeof(EntryState));
17259 // block->bbEntryState.esRefcount = 1;
17261 block->bbEntryState->esStackDepth = srcState->esStackDepth;
17262 block->bbEntryState->thisInitialized = TIS_Bottom;
17264 if (srcState->esStackDepth > 0)
17266 block->bbSetStack(new (this, CMK_Unknown) StackEntry[srcState->esStackDepth]);
17267 unsigned stackSize = srcState->esStackDepth * sizeof(StackEntry);
17269 memcpy(block->bbEntryState->esStack, srcState->esStack, stackSize);
17270 for (unsigned level = 0; level < srcState->esStackDepth; level++)
17272 GenTree* tree = srcState->esStack[level].val;
17273 block->bbEntryState->esStack[level].val = gtCloneExpr(tree);
17277 if (verTrackObjCtorInitState)
17279 verSetThisInit(block, srcState->thisInitialized);
17285 void Compiler::verSetThisInit(BasicBlock* block, ThisInitState tis)
17287 assert(tis != TIS_Bottom); // Precondition.
17288 if (block->bbEntryState == nullptr)
17290 block->bbEntryState = new (this, CMK_Unknown) EntryState();
17293 block->bbEntryState->thisInitialized = tis;
17297 * Resets the current state to the state at the start of the basic block
17299 void Compiler::verResetCurrentState(BasicBlock* block, EntryState* destState)
17302 if (block->bbEntryState == nullptr)
17304 destState->esStackDepth = 0;
17305 destState->thisInitialized = TIS_Bottom;
17309 destState->esStackDepth = block->bbEntryState->esStackDepth;
17311 if (destState->esStackDepth > 0)
17313 unsigned stackSize = destState->esStackDepth * sizeof(StackEntry);
17315 memcpy(destState->esStack, block->bbStackOnEntry(), stackSize);
17318 destState->thisInitialized = block->bbThisOnEntry();
17323 ThisInitState BasicBlock::bbThisOnEntry()
17325 return bbEntryState ? bbEntryState->thisInitialized : TIS_Bottom;
17328 unsigned BasicBlock::bbStackDepthOnEntry()
17330 return (bbEntryState ? bbEntryState->esStackDepth : 0);
17333 void BasicBlock::bbSetStack(void* stackBuffer)
17335 assert(bbEntryState);
17336 assert(stackBuffer);
17337 bbEntryState->esStack = (StackEntry*)stackBuffer;
17340 StackEntry* BasicBlock::bbStackOnEntry()
17342 assert(bbEntryState);
17343 return bbEntryState->esStack;
17346 void Compiler::verInitCurrentState()
17348 verTrackObjCtorInitState = FALSE;
17349 verCurrentState.thisInitialized = TIS_Bottom;
17351 if (tiVerificationNeeded)
17353 // Track this ptr initialization
17354 if (!info.compIsStatic && (info.compFlags & CORINFO_FLG_CONSTRUCTOR) && lvaTable[0].lvVerTypeInfo.IsObjRef())
17356 verTrackObjCtorInitState = TRUE;
17357 verCurrentState.thisInitialized = TIS_Uninit;
17361 // initialize stack info
17363 verCurrentState.esStackDepth = 0;
17364 assert(verCurrentState.esStack != nullptr);
17366 // copy current state to entry state of first BB
17367 verInitBBEntryState(fgFirstBB, &verCurrentState);
17370 Compiler* Compiler::impInlineRoot()
17372 if (impInlineInfo == nullptr)
17378 return impInlineInfo->InlineRoot;
17382 BYTE Compiler::impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk)
17384 if (predOrSucc == SpillCliquePred)
17386 return impInlineRoot()->impSpillCliquePredMembers.Get(blk->bbInd());
17390 assert(predOrSucc == SpillCliqueSucc);
17391 return impInlineRoot()->impSpillCliqueSuccMembers.Get(blk->bbInd());
17395 void Compiler::impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val)
17397 if (predOrSucc == SpillCliquePred)
17399 impInlineRoot()->impSpillCliquePredMembers.Set(blk->bbInd(), val);
17403 assert(predOrSucc == SpillCliqueSucc);
17404 impInlineRoot()->impSpillCliqueSuccMembers.Set(blk->bbInd(), val);
17408 /*****************************************************************************
17410 * Convert the instrs ("import") into our internal format (trees). The
17411 * basic flowgraph has already been constructed and is passed in.
17414 void Compiler::impImport(BasicBlock* method)
17419 printf("*************** In impImport() for %s\n", info.compFullName);
17423 /* Allocate the stack contents */
17425 if (info.compMaxStack <= _countof(impSmallStack))
17427 /* Use local variable, don't waste time allocating on the heap */
17429 impStkSize = _countof(impSmallStack);
17430 verCurrentState.esStack = impSmallStack;
17434 impStkSize = info.compMaxStack;
17435 verCurrentState.esStack = new (this, CMK_ImpStack) StackEntry[impStkSize];
17438 // initialize the entry state at start of method
17439 verInitCurrentState();
17441 // Initialize stuff related to figuring "spill cliques" (see spec comment for impGetSpillTmpBase).
17442 Compiler* inlineRoot = impInlineRoot();
17443 if (this == inlineRoot) // These are only used on the root of the inlining tree.
17445 // We have initialized these previously, but to size 0. Make them larger.
17446 impPendingBlockMembers.Init(getAllocator(), fgBBNumMax * 2);
17447 impSpillCliquePredMembers.Init(getAllocator(), fgBBNumMax * 2);
17448 impSpillCliqueSuccMembers.Init(getAllocator(), fgBBNumMax * 2);
17450 inlineRoot->impPendingBlockMembers.Reset(fgBBNumMax * 2);
17451 inlineRoot->impSpillCliquePredMembers.Reset(fgBBNumMax * 2);
17452 inlineRoot->impSpillCliqueSuccMembers.Reset(fgBBNumMax * 2);
17453 impBlockListNodeFreeList = nullptr;
17456 impLastILoffsStmt = nullptr;
17457 impNestedStackSpill = false;
17459 impBoxTemp = BAD_VAR_NUM;
17461 impPendingList = impPendingFree = nullptr;
17463 /* Add the entry-point to the worker-list */
17465 // Skip leading internal blocks. There can be one as a leading scratch BB, and more
17466 // from EH normalization.
17467 // NOTE: It might be possible to always just put fgFirstBB on the pending list, and let everything else just fall
17469 for (; method->bbFlags & BBF_INTERNAL; method = method->bbNext)
17471 // Treat these as imported.
17472 assert(method->bbJumpKind == BBJ_NONE); // We assume all the leading ones are fallthrough.
17473 JITDUMP("Marking leading BBF_INTERNAL block BB%02u as BBF_IMPORTED\n", method->bbNum);
17474 method->bbFlags |= BBF_IMPORTED;
17477 impImportBlockPending(method);
17479 /* Import blocks in the worker-list until there are no more */
17481 while (impPendingList)
17483 /* Remove the entry at the front of the list */
17485 PendingDsc* dsc = impPendingList;
17486 impPendingList = impPendingList->pdNext;
17487 impSetPendingBlockMember(dsc->pdBB, 0);
17489 /* Restore the stack state */
17491 verCurrentState.thisInitialized = dsc->pdThisPtrInit;
17492 verCurrentState.esStackDepth = dsc->pdSavedStack.ssDepth;
17493 if (verCurrentState.esStackDepth)
17495 impRestoreStackState(&dsc->pdSavedStack);
17498 /* Add the entry to the free list for reuse */
17500 dsc->pdNext = impPendingFree;
17501 impPendingFree = dsc;
17503 /* Now import the block */
17505 if (dsc->pdBB->bbFlags & BBF_FAILED_VERIFICATION)
17508 #ifdef _TARGET_64BIT_
17509 // On AMD64, during verification we have to match JIT64 behavior since the VM is very tighly
17510 // coupled with the JIT64 IL Verification logic. Look inside verHandleVerificationFailure
17511 // method for further explanation on why we raise this exception instead of making the jitted
17512 // code throw the verification exception during execution.
17513 if (tiVerificationNeeded && opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IMPORT_ONLY))
17515 BADCODE("Basic block marked as not verifiable");
17518 #endif // _TARGET_64BIT_
17520 verConvertBBToThrowVerificationException(dsc->pdBB DEBUGARG(true));
17521 impEndTreeList(dsc->pdBB);
17526 impImportBlock(dsc->pdBB);
17528 if (compDonotInline())
17532 if (compIsForImportOnly() && !tiVerificationNeeded)
17540 if (verbose && info.compXcptnsCount)
17542 printf("\nAfter impImport() added block for try,catch,finally");
17543 fgDispBasicBlocks();
17547 // Used in impImportBlockPending() for STRESS_CHK_REIMPORT
17548 for (BasicBlock* block = fgFirstBB; block; block = block->bbNext)
17550 block->bbFlags &= ~BBF_VISITED;
17554 assert(!compIsForInlining() || !tiVerificationNeeded);
17557 // Checks if a typeinfo (usually stored in the type stack) is a struct.
17558 // The invariant here is that if it's not a ref or a method and has a class handle
17559 // it's a valuetype
17560 bool Compiler::impIsValueType(typeInfo* pTypeInfo)
17562 if (pTypeInfo && pTypeInfo->IsValueClassWithClsHnd())
17572 /*****************************************************************************
17573 * Check to see if the tree is the address of a local or
17574 the address of a field in a local.
17576 *lclVarTreeOut will contain the GT_LCL_VAR tree when it returns TRUE.
17580 BOOL Compiler::impIsAddressInLocal(GenTree* tree, GenTree** lclVarTreeOut)
17582 if (tree->gtOper != GT_ADDR)
17587 GenTree* op = tree->gtOp.gtOp1;
17588 while (op->gtOper == GT_FIELD)
17590 op = op->gtField.gtFldObj;
17591 if (op && op->gtOper == GT_ADDR) // Skip static fields where op will be NULL.
17593 op = op->gtOp.gtOp1;
17601 if (op->gtOper == GT_LCL_VAR)
17603 *lclVarTreeOut = op;
17612 //------------------------------------------------------------------------
17613 // impMakeDiscretionaryInlineObservations: make observations that help
17614 // determine the profitability of a discretionary inline
17617 // pInlineInfo -- InlineInfo for the inline, or null for the prejit root
17618 // inlineResult -- InlineResult accumulating information about this inline
17621 // If inlining or prejitting the root, this method also makes
17622 // various observations about the method that factor into inline
17623 // decisions. It sets `compNativeSizeEstimate` as a side effect.
17625 void Compiler::impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult)
17627 assert(pInlineInfo != nullptr && compIsForInlining() || // Perform the actual inlining.
17628 pInlineInfo == nullptr && !compIsForInlining() // Calculate the static inlining hint for ngen.
17631 // If we're really inlining, we should just have one result in play.
17632 assert((pInlineInfo == nullptr) || (inlineResult == pInlineInfo->inlineResult));
17634 // If this is a "forceinline" method, the JIT probably shouldn't have gone
17635 // to the trouble of estimating the native code size. Even if it did, it
17636 // shouldn't be relying on the result of this method.
17637 assert(inlineResult->GetObservation() == InlineObservation::CALLEE_IS_DISCRETIONARY_INLINE);
17639 // Note if the caller contains NEWOBJ or NEWARR.
17640 Compiler* rootCompiler = impInlineRoot();
17642 if ((rootCompiler->optMethodFlags & OMF_HAS_NEWARRAY) != 0)
17644 inlineResult->Note(InlineObservation::CALLER_HAS_NEWARRAY);
17647 if ((rootCompiler->optMethodFlags & OMF_HAS_NEWOBJ) != 0)
17649 inlineResult->Note(InlineObservation::CALLER_HAS_NEWOBJ);
17652 bool calleeIsStatic = (info.compFlags & CORINFO_FLG_STATIC) != 0;
17653 bool isSpecialMethod = (info.compFlags & CORINFO_FLG_CONSTRUCTOR) != 0;
17655 if (isSpecialMethod)
17657 if (calleeIsStatic)
17659 inlineResult->Note(InlineObservation::CALLEE_IS_CLASS_CTOR);
17663 inlineResult->Note(InlineObservation::CALLEE_IS_INSTANCE_CTOR);
17666 else if (!calleeIsStatic)
17668 // Callee is an instance method.
17670 // Check if the callee has the same 'this' as the root.
17671 if (pInlineInfo != nullptr)
17673 GenTree* thisArg = pInlineInfo->iciCall->gtCall.gtCallObjp;
17675 bool isSameThis = impIsThis(thisArg);
17676 inlineResult->NoteBool(InlineObservation::CALLSITE_IS_SAME_THIS, isSameThis);
17680 // Note if the callee's class is a promotable struct
17681 if ((info.compClassAttr & CORINFO_FLG_VALUECLASS) != 0)
17683 lvaStructPromotionInfo structPromotionInfo;
17684 lvaCanPromoteStructType(info.compClassHnd, &structPromotionInfo, false);
17685 if (structPromotionInfo.canPromote)
17687 inlineResult->Note(InlineObservation::CALLEE_CLASS_PROMOTABLE);
17691 #ifdef FEATURE_SIMD
17693 // Note if this method is has SIMD args or return value
17694 if (pInlineInfo != nullptr && pInlineInfo->hasSIMDTypeArgLocalOrReturn)
17696 inlineResult->Note(InlineObservation::CALLEE_HAS_SIMD);
17699 #endif // FEATURE_SIMD
17701 // Roughly classify callsite frequency.
17702 InlineCallsiteFrequency frequency = InlineCallsiteFrequency::UNUSED;
17704 // If this is a prejit root, or a maximally hot block...
17705 if ((pInlineInfo == nullptr) || (pInlineInfo->iciBlock->bbWeight >= BB_MAX_WEIGHT))
17707 frequency = InlineCallsiteFrequency::HOT;
17709 // No training data. Look for loop-like things.
17710 // We consider a recursive call loop-like. Do not give the inlining boost to the method itself.
17711 // However, give it to things nearby.
17712 else if ((pInlineInfo->iciBlock->bbFlags & BBF_BACKWARD_JUMP) &&
17713 (pInlineInfo->fncHandle != pInlineInfo->inlineCandidateInfo->ilCallerHandle))
17715 frequency = InlineCallsiteFrequency::LOOP;
17717 else if (pInlineInfo->iciBlock->hasProfileWeight() && (pInlineInfo->iciBlock->bbWeight > BB_ZERO_WEIGHT))
17719 frequency = InlineCallsiteFrequency::WARM;
17721 // Now modify the multiplier based on where we're called from.
17722 else if (pInlineInfo->iciBlock->isRunRarely() || ((info.compFlags & FLG_CCTOR) == FLG_CCTOR))
17724 frequency = InlineCallsiteFrequency::RARE;
17728 frequency = InlineCallsiteFrequency::BORING;
17731 // Also capture the block weight of the call site. In the prejit
17732 // root case, assume there's some hot call site for this method.
17733 unsigned weight = 0;
17735 if (pInlineInfo != nullptr)
17737 weight = pInlineInfo->iciBlock->bbWeight;
17741 weight = BB_MAX_WEIGHT;
17744 inlineResult->NoteInt(InlineObservation::CALLSITE_FREQUENCY, static_cast<int>(frequency));
17745 inlineResult->NoteInt(InlineObservation::CALLSITE_WEIGHT, static_cast<int>(weight));
17748 /*****************************************************************************
17749 This method makes STATIC inlining decision based on the IL code.
17750 It should not make any inlining decision based on the context.
17751 If forceInline is true, then the inlining decision should not depend on
17752 performance heuristics (code size, etc.).
17755 void Compiler::impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
17756 CORINFO_METHOD_INFO* methInfo,
17758 InlineResult* inlineResult)
17760 unsigned codeSize = methInfo->ILCodeSize;
17762 // We shouldn't have made up our minds yet...
17763 assert(!inlineResult->IsDecided());
17765 if (methInfo->EHcount)
17767 inlineResult->NoteFatal(InlineObservation::CALLEE_HAS_EH);
17771 if ((methInfo->ILCode == nullptr) || (codeSize == 0))
17773 inlineResult->NoteFatal(InlineObservation::CALLEE_HAS_NO_BODY);
17777 // For now we don't inline varargs (import code can't handle it)
17779 if (methInfo->args.isVarArg())
17781 inlineResult->NoteFatal(InlineObservation::CALLEE_HAS_MANAGED_VARARGS);
17785 // Reject if it has too many locals.
17786 // This is currently an implementation limit due to fixed-size arrays in the
17787 // inline info, rather than a performance heuristic.
17789 inlineResult->NoteInt(InlineObservation::CALLEE_NUMBER_OF_LOCALS, methInfo->locals.numArgs);
17791 if (methInfo->locals.numArgs > MAX_INL_LCLS)
17793 inlineResult->NoteFatal(InlineObservation::CALLEE_TOO_MANY_LOCALS);
17797 // Make sure there aren't too many arguments.
17798 // This is currently an implementation limit due to fixed-size arrays in the
17799 // inline info, rather than a performance heuristic.
17801 inlineResult->NoteInt(InlineObservation::CALLEE_NUMBER_OF_ARGUMENTS, methInfo->args.numArgs);
17803 if (methInfo->args.numArgs > MAX_INL_ARGS)
17805 inlineResult->NoteFatal(InlineObservation::CALLEE_TOO_MANY_ARGUMENTS);
17809 // Note force inline state
17811 inlineResult->NoteBool(InlineObservation::CALLEE_IS_FORCE_INLINE, forceInline);
17813 // Note IL code size
17815 inlineResult->NoteInt(InlineObservation::CALLEE_IL_CODE_SIZE, codeSize);
17817 if (inlineResult->IsFailure())
17822 // Make sure maxstack is not too big
17824 inlineResult->NoteInt(InlineObservation::CALLEE_MAXSTACK, methInfo->maxStack);
17826 if (inlineResult->IsFailure())
17832 /*****************************************************************************
17835 void Compiler::impCheckCanInline(GenTree* call,
17836 CORINFO_METHOD_HANDLE fncHandle,
17838 CORINFO_CONTEXT_HANDLE exactContextHnd,
17839 InlineCandidateInfo** ppInlineCandidateInfo,
17840 InlineResult* inlineResult)
17842 // Either EE or JIT might throw exceptions below.
17843 // If that happens, just don't inline the method.
17849 CORINFO_METHOD_HANDLE fncHandle;
17851 CORINFO_CONTEXT_HANDLE exactContextHnd;
17852 InlineResult* result;
17853 InlineCandidateInfo** ppInlineCandidateInfo;
17855 memset(¶m, 0, sizeof(param));
17857 param.pThis = this;
17859 param.fncHandle = fncHandle;
17860 param.methAttr = methAttr;
17861 param.exactContextHnd = (exactContextHnd != nullptr) ? exactContextHnd : MAKE_METHODCONTEXT(fncHandle);
17862 param.result = inlineResult;
17863 param.ppInlineCandidateInfo = ppInlineCandidateInfo;
17865 bool success = eeRunWithErrorTrap<Param>(
17866 [](Param* pParam) {
17867 DWORD dwRestrictions = 0;
17868 CorInfoInitClassResult initClassResult;
17871 const char* methodName;
17872 const char* className;
17873 methodName = pParam->pThis->eeGetMethodName(pParam->fncHandle, &className);
17875 if (JitConfig.JitNoInline())
17877 pParam->result->NoteFatal(InlineObservation::CALLEE_IS_JIT_NOINLINE);
17882 /* Try to get the code address/size for the method */
17884 CORINFO_METHOD_INFO methInfo;
17885 if (!pParam->pThis->info.compCompHnd->getMethodInfo(pParam->fncHandle, &methInfo))
17887 pParam->result->NoteFatal(InlineObservation::CALLEE_NO_METHOD_INFO);
17892 forceInline = !!(pParam->methAttr & CORINFO_FLG_FORCEINLINE);
17894 pParam->pThis->impCanInlineIL(pParam->fncHandle, &methInfo, forceInline, pParam->result);
17896 if (pParam->result->IsFailure())
17898 assert(pParam->result->IsNever());
17902 // Speculatively check if initClass() can be done.
17903 // If it can be done, we will try to inline the method. If inlining
17904 // succeeds, then we will do the non-speculative initClass() and commit it.
17905 // If this speculative call to initClass() fails, there is no point
17906 // trying to inline this method.
17908 pParam->pThis->info.compCompHnd->initClass(nullptr /* field */, pParam->fncHandle /* method */,
17909 pParam->exactContextHnd /* context */,
17910 TRUE /* speculative */);
17912 if (initClassResult & CORINFO_INITCLASS_DONT_INLINE)
17914 pParam->result->NoteFatal(InlineObservation::CALLSITE_CLASS_INIT_FAILURE_SPEC);
17918 // Given the EE the final say in whether to inline or not.
17919 // This should be last since for verifiable code, this can be expensive
17921 /* VM Inline check also ensures that the method is verifiable if needed */
17922 CorInfoInline vmResult;
17923 vmResult = pParam->pThis->info.compCompHnd->canInline(pParam->pThis->info.compMethodHnd, pParam->fncHandle,
17926 if (vmResult == INLINE_FAIL)
17928 pParam->result->NoteFatal(InlineObservation::CALLSITE_IS_VM_NOINLINE);
17930 else if (vmResult == INLINE_NEVER)
17932 pParam->result->NoteFatal(InlineObservation::CALLEE_IS_VM_NOINLINE);
17935 if (pParam->result->IsFailure())
17937 // Make sure not to report this one. It was already reported by the VM.
17938 pParam->result->SetReported();
17942 // check for unsupported inlining restrictions
17943 assert((dwRestrictions & ~(INLINE_RESPECT_BOUNDARY | INLINE_NO_CALLEE_LDSTR | INLINE_SAME_THIS)) == 0);
17945 if (dwRestrictions & INLINE_SAME_THIS)
17947 GenTree* thisArg = pParam->call->gtCall.gtCallObjp;
17950 if (!pParam->pThis->impIsThis(thisArg))
17952 pParam->result->NoteFatal(InlineObservation::CALLSITE_REQUIRES_SAME_THIS);
17957 /* Get the method properties */
17959 CORINFO_CLASS_HANDLE clsHandle;
17960 clsHandle = pParam->pThis->info.compCompHnd->getMethodClass(pParam->fncHandle);
17962 clsAttr = pParam->pThis->info.compCompHnd->getClassAttribs(clsHandle);
17964 /* Get the return type */
17966 var_types fncRetType;
17967 fncRetType = pParam->call->TypeGet();
17970 var_types fncRealRetType;
17971 fncRealRetType = JITtype2varType(methInfo.args.retType);
17973 assert((genActualType(fncRealRetType) == genActualType(fncRetType)) ||
17974 // <BUGNUM> VSW 288602 </BUGNUM>
17975 // In case of IJW, we allow to assign a native pointer to a BYREF.
17976 (fncRetType == TYP_BYREF && methInfo.args.retType == CORINFO_TYPE_PTR) ||
17977 (varTypeIsStruct(fncRetType) && (fncRealRetType == TYP_STRUCT)));
17981 // Allocate an InlineCandidateInfo structure
17983 InlineCandidateInfo* pInfo;
17984 pInfo = new (pParam->pThis, CMK_Inlining) InlineCandidateInfo;
17986 pInfo->dwRestrictions = dwRestrictions;
17987 pInfo->methInfo = methInfo;
17988 pInfo->methAttr = pParam->methAttr;
17989 pInfo->clsHandle = clsHandle;
17990 pInfo->clsAttr = clsAttr;
17991 pInfo->fncRetType = fncRetType;
17992 pInfo->exactContextHnd = pParam->exactContextHnd;
17993 pInfo->ilCallerHandle = pParam->pThis->info.compMethodHnd;
17994 pInfo->initClassResult = initClassResult;
17996 *(pParam->ppInlineCandidateInfo) = pInfo;
18003 param.result->NoteFatal(InlineObservation::CALLSITE_COMPILATION_ERROR);
18007 //------------------------------------------------------------------------
18008 // impInlineRecordArgInfo: record information about an inline candidate argument
18011 // pInlineInfo - inline info for the inline candidate
18012 // curArgVal - tree for the caller actual argument value
18013 // argNum - logical index of this argument
18014 // inlineResult - result of ongoing inline evaluation
18018 // Checks for various inline blocking conditions and makes notes in
18019 // the inline info arg table about the properties of the actual. These
18020 // properties are used later by impFetchArg to determine how best to
18021 // pass the argument into the inlinee.
18023 void Compiler::impInlineRecordArgInfo(InlineInfo* pInlineInfo,
18024 GenTree* curArgVal,
18026 InlineResult* inlineResult)
18028 InlArgInfo* inlCurArgInfo = &pInlineInfo->inlArgInfo[argNum];
18030 if (curArgVal->gtOper == GT_MKREFANY)
18032 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_IS_MKREFANY);
18036 inlCurArgInfo->argNode = curArgVal;
18038 GenTree* lclVarTree;
18039 if (impIsAddressInLocal(curArgVal, &lclVarTree) && varTypeIsStruct(lclVarTree))
18041 inlCurArgInfo->argIsByRefToStructLocal = true;
18042 #ifdef FEATURE_SIMD
18043 if (lvaTable[lclVarTree->AsLclVarCommon()->gtLclNum].lvSIMDType)
18045 pInlineInfo->hasSIMDTypeArgLocalOrReturn = true;
18047 #endif // FEATURE_SIMD
18050 if (curArgVal->gtFlags & GTF_ALL_EFFECT)
18052 inlCurArgInfo->argHasGlobRef = (curArgVal->gtFlags & GTF_GLOB_REF) != 0;
18053 inlCurArgInfo->argHasSideEff = (curArgVal->gtFlags & (GTF_ALL_EFFECT & ~GTF_GLOB_REF)) != 0;
18056 if (curArgVal->gtOper == GT_LCL_VAR)
18058 inlCurArgInfo->argIsLclVar = true;
18060 /* Remember the "original" argument number */
18061 curArgVal->gtLclVar.gtLclILoffs = argNum;
18064 if ((curArgVal->OperKind() & GTK_CONST) ||
18065 ((curArgVal->gtOper == GT_ADDR) && (curArgVal->gtOp.gtOp1->gtOper == GT_LCL_VAR)))
18067 inlCurArgInfo->argIsInvariant = true;
18068 if (inlCurArgInfo->argIsThis && (curArgVal->gtOper == GT_CNS_INT) && (curArgVal->gtIntCon.gtIconVal == 0))
18070 // Abort inlining at this call site
18071 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_HAS_NULL_THIS);
18076 // If the arg is a local that is address-taken, we can't safely
18077 // directly substitute it into the inlinee.
18079 // Previously we'd accomplish this by setting "argHasLdargaOp" but
18080 // that has a stronger meaning: that the arg value can change in
18081 // the method body. Using that flag prevents type propagation,
18082 // which is safe in this case.
18084 // Instead mark the arg as having a caller local ref.
18085 if (!inlCurArgInfo->argIsInvariant && gtHasLocalsWithAddrOp(curArgVal))
18087 inlCurArgInfo->argHasCallerLocalRef = true;
18093 if (inlCurArgInfo->argIsThis)
18095 printf("thisArg:");
18099 printf("\nArgument #%u:", argNum);
18101 if (inlCurArgInfo->argIsLclVar)
18103 printf(" is a local var");
18105 if (inlCurArgInfo->argIsInvariant)
18107 printf(" is a constant");
18109 if (inlCurArgInfo->argHasGlobRef)
18111 printf(" has global refs");
18113 if (inlCurArgInfo->argHasCallerLocalRef)
18115 printf(" has caller local ref");
18117 if (inlCurArgInfo->argHasSideEff)
18119 printf(" has side effects");
18121 if (inlCurArgInfo->argHasLdargaOp)
18123 printf(" has ldarga effect");
18125 if (inlCurArgInfo->argHasStargOp)
18127 printf(" has starg effect");
18129 if (inlCurArgInfo->argIsByRefToStructLocal)
18131 printf(" is byref to a struct local");
18135 gtDispTree(curArgVal);
18141 //------------------------------------------------------------------------
18142 // impInlineInitVars: setup inline information for inlinee args and locals
18145 // pInlineInfo - inline info for the inline candidate
18148 // This method primarily adds caller-supplied info to the inlArgInfo
18149 // and sets up the lclVarInfo table.
18151 // For args, the inlArgInfo records properties of the actual argument
18152 // including the tree node that produces the arg value. This node is
18153 // usually the tree node present at the call, but may also differ in
18155 // - when the call arg is a GT_RET_EXPR, we search back through the ret
18156 // expr chain for the actual node. Note this will either be the original
18157 // call (which will be a failed inline by this point), or the return
18158 // expression from some set of inlines.
18159 // - when argument type casting is needed the necessary casts are added
18160 // around the argument node.
18161 // - if an argment can be simplified by folding then the node here is the
18164 // The method may make observations that lead to marking this candidate as
18165 // a failed inline. If this happens the initialization is abandoned immediately
18166 // to try and reduce the jit time cost for a failed inline.
18168 void Compiler::impInlineInitVars(InlineInfo* pInlineInfo)
18170 assert(!compIsForInlining());
18172 GenTree* call = pInlineInfo->iciCall;
18173 CORINFO_METHOD_INFO* methInfo = &pInlineInfo->inlineCandidateInfo->methInfo;
18174 unsigned clsAttr = pInlineInfo->inlineCandidateInfo->clsAttr;
18175 InlArgInfo* inlArgInfo = pInlineInfo->inlArgInfo;
18176 InlLclVarInfo* lclVarInfo = pInlineInfo->lclVarInfo;
18177 InlineResult* inlineResult = pInlineInfo->inlineResult;
18179 const bool hasRetBuffArg = impMethodInfo_hasRetBuffArg(methInfo);
18181 /* init the argument stuct */
18183 memset(inlArgInfo, 0, (MAX_INL_ARGS + 1) * sizeof(inlArgInfo[0]));
18185 /* Get hold of the 'this' pointer and the argument list proper */
18187 GenTree* thisArg = call->gtCall.gtCallObjp;
18188 GenTree* argList = call->gtCall.gtCallArgs;
18189 unsigned argCnt = 0; // Count of the arguments
18191 assert((methInfo->args.hasThis()) == (thisArg != nullptr));
18195 inlArgInfo[0].argIsThis = true;
18196 GenTree* actualThisArg = thisArg->gtRetExprVal();
18197 impInlineRecordArgInfo(pInlineInfo, actualThisArg, argCnt, inlineResult);
18199 if (inlineResult->IsFailure())
18204 /* Increment the argument count */
18208 /* Record some information about each of the arguments */
18209 bool hasTypeCtxtArg = (methInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) != 0;
18211 #if USER_ARGS_COME_LAST
18212 unsigned typeCtxtArg = thisArg ? 1 : 0;
18213 #else // USER_ARGS_COME_LAST
18214 unsigned typeCtxtArg = methInfo->args.totalILArgs();
18215 #endif // USER_ARGS_COME_LAST
18217 for (GenTree* argTmp = argList; argTmp; argTmp = argTmp->gtOp.gtOp2)
18219 if (argTmp == argList && hasRetBuffArg)
18224 // Ignore the type context argument
18225 if (hasTypeCtxtArg && (argCnt == typeCtxtArg))
18227 pInlineInfo->typeContextArg = typeCtxtArg;
18228 typeCtxtArg = 0xFFFFFFFF;
18232 assert(argTmp->gtOper == GT_LIST);
18233 GenTree* arg = argTmp->gtOp.gtOp1;
18234 GenTree* actualArg = arg->gtRetExprVal();
18235 impInlineRecordArgInfo(pInlineInfo, actualArg, argCnt, inlineResult);
18237 if (inlineResult->IsFailure())
18242 /* Increment the argument count */
18246 /* Make sure we got the arg number right */
18247 assert(argCnt == methInfo->args.totalILArgs());
18249 #ifdef FEATURE_SIMD
18250 bool foundSIMDType = pInlineInfo->hasSIMDTypeArgLocalOrReturn;
18251 #endif // FEATURE_SIMD
18253 /* We have typeless opcodes, get type information from the signature */
18259 if (clsAttr & CORINFO_FLG_VALUECLASS)
18261 sigType = TYP_BYREF;
18268 lclVarInfo[0].lclVerTypeInfo = verMakeTypeInfo(pInlineInfo->inlineCandidateInfo->clsHandle);
18269 lclVarInfo[0].lclHasLdlocaOp = false;
18271 #ifdef FEATURE_SIMD
18272 // We always want to check isSIMDClass, since we want to set foundSIMDType (to increase
18273 // the inlining multiplier) for anything in that assembly.
18274 // But we only need to normalize it if it is a TYP_STRUCT
18275 // (which we need to do even if we have already set foundSIMDType).
18276 if ((!foundSIMDType || (sigType == TYP_STRUCT)) && isSIMDorHWSIMDClass(&(lclVarInfo[0].lclVerTypeInfo)))
18278 if (sigType == TYP_STRUCT)
18280 sigType = impNormStructType(lclVarInfo[0].lclVerTypeInfo.GetClassHandle());
18282 foundSIMDType = true;
18284 #endif // FEATURE_SIMD
18285 lclVarInfo[0].lclTypeInfo = sigType;
18287 assert(varTypeIsGC(thisArg->gtType) || // "this" is managed
18288 (thisArg->gtType == TYP_I_IMPL && // "this" is unmgd but the method's class doesnt care
18289 (clsAttr & CORINFO_FLG_VALUECLASS)));
18291 if (genActualType(thisArg->gtType) != genActualType(sigType))
18293 if (sigType == TYP_REF)
18295 /* The argument cannot be bashed into a ref (see bug 750871) */
18296 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_NO_BASH_TO_REF);
18300 /* This can only happen with byrefs <-> ints/shorts */
18302 assert(genActualType(sigType) == TYP_I_IMPL || sigType == TYP_BYREF);
18303 assert(genActualType(thisArg->gtType) == TYP_I_IMPL || thisArg->gtType == TYP_BYREF);
18305 if (sigType == TYP_BYREF)
18307 lclVarInfo[0].lclVerTypeInfo = typeInfo(varType2tiType(TYP_I_IMPL));
18309 else if (thisArg->gtType == TYP_BYREF)
18311 assert(sigType == TYP_I_IMPL);
18313 /* If possible change the BYREF to an int */
18314 if (thisArg->IsVarAddr())
18316 thisArg->gtType = TYP_I_IMPL;
18317 lclVarInfo[0].lclVerTypeInfo = typeInfo(varType2tiType(TYP_I_IMPL));
18321 /* Arguments 'int <- byref' cannot be bashed */
18322 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_NO_BASH_TO_INT);
18329 /* Init the types of the arguments and make sure the types
18330 * from the trees match the types in the signature */
18332 CORINFO_ARG_LIST_HANDLE argLst;
18333 argLst = methInfo->args.args;
18336 for (i = (thisArg ? 1 : 0); i < argCnt; i++, argLst = info.compCompHnd->getArgNext(argLst))
18338 var_types sigType = (var_types)eeGetArgType(argLst, &methInfo->args);
18340 lclVarInfo[i].lclVerTypeInfo = verParseArgSigToTypeInfo(&methInfo->args, argLst);
18342 #ifdef FEATURE_SIMD
18343 if ((!foundSIMDType || (sigType == TYP_STRUCT)) && isSIMDorHWSIMDClass(&(lclVarInfo[i].lclVerTypeInfo)))
18345 // If this is a SIMD class (i.e. in the SIMD assembly), then we will consider that we've
18346 // found a SIMD type, even if this may not be a type we recognize (the assumption is that
18347 // it is likely to use a SIMD type, and therefore we want to increase the inlining multiplier).
18348 foundSIMDType = true;
18349 if (sigType == TYP_STRUCT)
18351 var_types structType = impNormStructType(lclVarInfo[i].lclVerTypeInfo.GetClassHandle());
18352 sigType = structType;
18355 #endif // FEATURE_SIMD
18357 lclVarInfo[i].lclTypeInfo = sigType;
18358 lclVarInfo[i].lclHasLdlocaOp = false;
18360 /* Does the tree type match the signature type? */
18362 GenTree* inlArgNode = inlArgInfo[i].argNode;
18364 if (sigType != inlArgNode->gtType)
18366 /* In valid IL, this can only happen for short integer types or byrefs <-> [native] ints,
18367 but in bad IL cases with caller-callee signature mismatches we can see other types.
18368 Intentionally reject cases with mismatches so the jit is more flexible when
18369 encountering bad IL. */
18371 bool isPlausibleTypeMatch = (genActualType(sigType) == genActualType(inlArgNode->gtType)) ||
18372 (genActualTypeIsIntOrI(sigType) && inlArgNode->gtType == TYP_BYREF) ||
18373 (sigType == TYP_BYREF && genActualTypeIsIntOrI(inlArgNode->gtType));
18375 if (!isPlausibleTypeMatch)
18377 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_TYPES_INCOMPATIBLE);
18381 /* Is it a narrowing or widening cast?
18382 * Widening casts are ok since the value computed is already
18383 * normalized to an int (on the IL stack) */
18385 if (genTypeSize(inlArgNode->gtType) >= genTypeSize(sigType))
18387 if (sigType == TYP_BYREF)
18389 lclVarInfo[i].lclVerTypeInfo = typeInfo(varType2tiType(TYP_I_IMPL));
18391 else if (inlArgNode->gtType == TYP_BYREF)
18393 assert(varTypeIsIntOrI(sigType));
18395 /* If possible bash the BYREF to an int */
18396 if (inlArgNode->IsVarAddr())
18398 inlArgNode->gtType = TYP_I_IMPL;
18399 lclVarInfo[i].lclVerTypeInfo = typeInfo(varType2tiType(TYP_I_IMPL));
18403 /* Arguments 'int <- byref' cannot be changed */
18404 inlineResult->NoteFatal(InlineObservation::CALLSITE_ARG_NO_BASH_TO_INT);
18408 else if (genTypeSize(sigType) < EA_PTRSIZE)
18410 /* Narrowing cast */
18412 if (inlArgNode->gtOper == GT_LCL_VAR &&
18413 !lvaTable[inlArgNode->gtLclVarCommon.gtLclNum].lvNormalizeOnLoad() &&
18414 sigType == lvaGetRealType(inlArgNode->gtLclVarCommon.gtLclNum))
18416 /* We don't need to insert a cast here as the variable
18417 was assigned a normalized value of the right type */
18422 inlArgNode = inlArgInfo[i].argNode = gtNewCastNode(TYP_INT, inlArgNode, false, sigType);
18424 inlArgInfo[i].argIsLclVar = false;
18426 /* Try to fold the node in case we have constant arguments */
18428 if (inlArgInfo[i].argIsInvariant)
18430 inlArgNode = gtFoldExprConst(inlArgNode);
18431 inlArgInfo[i].argNode = inlArgNode;
18432 assert(inlArgNode->OperIsConst());
18435 #ifdef _TARGET_64BIT_
18436 else if (genTypeSize(genActualType(inlArgNode->gtType)) < genTypeSize(sigType))
18438 // This should only happen for int -> native int widening
18439 inlArgNode = inlArgInfo[i].argNode =
18440 gtNewCastNode(genActualType(sigType), inlArgNode, false, sigType);
18442 inlArgInfo[i].argIsLclVar = false;
18444 /* Try to fold the node in case we have constant arguments */
18446 if (inlArgInfo[i].argIsInvariant)
18448 inlArgNode = gtFoldExprConst(inlArgNode);
18449 inlArgInfo[i].argNode = inlArgNode;
18450 assert(inlArgNode->OperIsConst());
18453 #endif // _TARGET_64BIT_
18458 /* Init the types of the local variables */
18460 CORINFO_ARG_LIST_HANDLE localsSig;
18461 localsSig = methInfo->locals.args;
18463 for (i = 0; i < methInfo->locals.numArgs; i++)
18466 var_types type = (var_types)eeGetArgType(localsSig, &methInfo->locals, &isPinned);
18468 lclVarInfo[i + argCnt].lclHasLdlocaOp = false;
18469 lclVarInfo[i + argCnt].lclIsPinned = isPinned;
18470 lclVarInfo[i + argCnt].lclTypeInfo = type;
18472 if (varTypeIsGC(type))
18474 pInlineInfo->numberOfGcRefLocals++;
18479 // Pinned locals may cause inlines to fail.
18480 inlineResult->Note(InlineObservation::CALLEE_HAS_PINNED_LOCALS);
18481 if (inlineResult->IsFailure())
18487 lclVarInfo[i + argCnt].lclVerTypeInfo = verParseArgSigToTypeInfo(&methInfo->locals, localsSig);
18489 // If this local is a struct type with GC fields, inform the inliner. It may choose to bail
18490 // out on the inline.
18491 if (type == TYP_STRUCT)
18493 CORINFO_CLASS_HANDLE lclHandle = lclVarInfo[i + argCnt].lclVerTypeInfo.GetClassHandle();
18494 DWORD typeFlags = info.compCompHnd->getClassAttribs(lclHandle);
18495 if ((typeFlags & CORINFO_FLG_CONTAINS_GC_PTR) != 0)
18497 inlineResult->Note(InlineObservation::CALLEE_HAS_GC_STRUCT);
18498 if (inlineResult->IsFailure())
18503 // Do further notification in the case where the call site is rare; some policies do
18504 // not track the relative hotness of call sites for "always" inline cases.
18505 if (pInlineInfo->iciBlock->isRunRarely())
18507 inlineResult->Note(InlineObservation::CALLSITE_RARE_GC_STRUCT);
18508 if (inlineResult->IsFailure())
18517 localsSig = info.compCompHnd->getArgNext(localsSig);
18519 #ifdef FEATURE_SIMD
18520 if ((!foundSIMDType || (type == TYP_STRUCT)) && isSIMDorHWSIMDClass(&(lclVarInfo[i + argCnt].lclVerTypeInfo)))
18522 foundSIMDType = true;
18523 if (featureSIMD && type == TYP_STRUCT)
18525 var_types structType = impNormStructType(lclVarInfo[i + argCnt].lclVerTypeInfo.GetClassHandle());
18526 lclVarInfo[i + argCnt].lclTypeInfo = structType;
18529 #endif // FEATURE_SIMD
18532 #ifdef FEATURE_SIMD
18533 if (!foundSIMDType && (call->AsCall()->gtRetClsHnd != nullptr) && isSIMDorHWSIMDClass(call->AsCall()->gtRetClsHnd))
18535 foundSIMDType = true;
18537 pInlineInfo->hasSIMDTypeArgLocalOrReturn = foundSIMDType;
18538 #endif // FEATURE_SIMD
18541 //------------------------------------------------------------------------
18542 // impInlineFetchLocal: get a local var that represents an inlinee local
18545 // lclNum -- number of the inlinee local
18546 // reason -- debug string describing purpose of the local var
18549 // Number of the local to use
18552 // This method is invoked only for locals actually used in the
18555 // Allocates a new temp if necessary, and copies key properties
18556 // over from the inlinee local var info.
18558 unsigned Compiler::impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason))
18560 assert(compIsForInlining());
18562 unsigned tmpNum = impInlineInfo->lclTmpNum[lclNum];
18564 if (tmpNum == BAD_VAR_NUM)
18566 const InlLclVarInfo& inlineeLocal = impInlineInfo->lclVarInfo[lclNum + impInlineInfo->argCnt];
18567 const var_types lclTyp = inlineeLocal.lclTypeInfo;
18569 // The lifetime of this local might span multiple BBs.
18570 // So it is a long lifetime local.
18571 impInlineInfo->lclTmpNum[lclNum] = tmpNum = lvaGrabTemp(false DEBUGARG(reason));
18573 // Copy over key info
18574 lvaTable[tmpNum].lvType = lclTyp;
18575 lvaTable[tmpNum].lvHasLdAddrOp = inlineeLocal.lclHasLdlocaOp;
18576 lvaTable[tmpNum].lvPinned = inlineeLocal.lclIsPinned;
18577 lvaTable[tmpNum].lvHasILStoreOp = inlineeLocal.lclHasStlocOp;
18578 lvaTable[tmpNum].lvHasMultipleILStoreOp = inlineeLocal.lclHasMultipleStlocOp;
18580 // Copy over class handle for ref types. Note this may be a
18581 // shared type -- someday perhaps we can get the exact
18582 // signature and pass in a more precise type.
18583 if (lclTyp == TYP_REF)
18585 lvaSetClass(tmpNum, inlineeLocal.lclVerTypeInfo.GetClassHandleForObjRef());
18588 if (inlineeLocal.lclVerTypeInfo.IsStruct())
18590 if (varTypeIsStruct(lclTyp))
18592 lvaSetStruct(tmpNum, inlineeLocal.lclVerTypeInfo.GetClassHandle(), true /* unsafe value cls check */);
18596 // This is a wrapped primitive. Make sure the verstate knows that
18597 lvaTable[tmpNum].lvVerTypeInfo = inlineeLocal.lclVerTypeInfo;
18602 // Sanity check that we're properly prepared for gc ref locals.
18603 if (varTypeIsGC(lclTyp))
18605 // Since there are gc locals we should have seen them earlier
18606 // and if there was a return value, set up the spill temp.
18607 assert(impInlineInfo->HasGcRefLocals());
18608 assert((info.compRetNativeType == TYP_VOID) || fgNeedReturnSpillTemp());
18612 // Make sure all pinned locals count as gc refs.
18613 assert(!inlineeLocal.lclIsPinned);
18621 //------------------------------------------------------------------------
18622 // impInlineFetchArg: return tree node for argument value in an inlinee
18625 // lclNum -- argument number in inlinee IL
18626 // inlArgInfo -- argument info for inlinee
18627 // lclVarInfo -- var info for inlinee
18630 // Tree for the argument's value. Often an inlinee-scoped temp
18631 // GT_LCL_VAR but can be other tree kinds, if the argument
18632 // expression from the caller can be directly substituted into the
18636 // Must be used only for arguments -- use impInlineFetchLocal for
18639 // Direct substitution is performed when the formal argument cannot
18640 // change value in the inlinee body (no starg or ldarga), and the
18641 // actual argument expression's value cannot be changed if it is
18642 // substituted it into the inlinee body.
18644 // Even if an inlinee-scoped temp is returned here, it may later be
18645 // "bashed" to a caller-supplied tree when arguments are actually
18646 // passed (see fgInlinePrependStatements). Bashing can happen if
18647 // the argument ends up being single use and other conditions are
18648 // met. So the contents of the tree returned here may not end up
18649 // being the ones ultimately used for the argument.
18651 // This method will side effect inlArgInfo. It should only be called
18652 // for actual uses of the argument in the inlinee.
18654 GenTree* Compiler::impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclVarInfo)
18656 // Cache the relevant arg and lcl info for this argument.
18657 // We will modify argInfo but not lclVarInfo.
18658 InlArgInfo& argInfo = inlArgInfo[lclNum];
18659 const InlLclVarInfo& lclInfo = lclVarInfo[lclNum];
18660 const bool argCanBeModified = argInfo.argHasLdargaOp || argInfo.argHasStargOp;
18661 const var_types lclTyp = lclInfo.lclTypeInfo;
18662 GenTree* op1 = nullptr;
18664 if (argInfo.argIsInvariant && !argCanBeModified)
18666 // Directly substitute constants or addresses of locals
18668 // Clone the constant. Note that we cannot directly use
18669 // argNode in the trees even if !argInfo.argIsUsed as this
18670 // would introduce aliasing between inlArgInfo[].argNode and
18671 // impInlineExpr. Then gtFoldExpr() could change it, causing
18672 // further references to the argument working off of the
18674 op1 = gtCloneExpr(argInfo.argNode);
18675 PREFIX_ASSUME(op1 != nullptr);
18676 argInfo.argTmpNum = BAD_VAR_NUM;
18678 // We may need to retype to ensure we match the callee's view of the type.
18679 // Otherwise callee-pass throughs of arguments can create return type
18680 // mismatches that block inlining.
18682 // Note argument type mismatches that prevent inlining should
18683 // have been caught in impInlineInitVars.
18684 if (op1->TypeGet() != lclTyp)
18686 op1->gtType = genActualType(lclTyp);
18689 else if (argInfo.argIsLclVar && !argCanBeModified && !argInfo.argHasCallerLocalRef)
18691 // Directly substitute unaliased caller locals for args that cannot be modified
18693 // Use the caller-supplied node if this is the first use.
18694 op1 = argInfo.argNode;
18695 argInfo.argTmpNum = op1->gtLclVarCommon.gtLclNum;
18697 // Use an equivalent copy if this is the second or subsequent
18698 // use, or if we need to retype.
18700 // Note argument type mismatches that prevent inlining should
18701 // have been caught in impInlineInitVars.
18702 if (argInfo.argIsUsed || (op1->TypeGet() != lclTyp))
18704 assert(op1->gtOper == GT_LCL_VAR);
18705 assert(lclNum == op1->gtLclVar.gtLclILoffs);
18707 var_types newTyp = lclTyp;
18709 if (!lvaTable[op1->gtLclVarCommon.gtLclNum].lvNormalizeOnLoad())
18711 newTyp = genActualType(lclTyp);
18714 // Create a new lcl var node - remember the argument lclNum
18715 op1 = gtNewLclvNode(op1->gtLclVarCommon.gtLclNum, newTyp, op1->gtLclVar.gtLclILoffs);
18718 else if (argInfo.argIsByRefToStructLocal && !argInfo.argHasStargOp)
18720 /* Argument is a by-ref address to a struct, a normed struct, or its field.
18721 In these cases, don't spill the byref to a local, simply clone the tree and use it.
18722 This way we will increase the chance for this byref to be optimized away by
18723 a subsequent "dereference" operation.
18725 From Dev11 bug #139955: Argument node can also be TYP_I_IMPL if we've bashed the tree
18726 (in impInlineInitVars()), if the arg has argHasLdargaOp as well as argIsByRefToStructLocal.
18727 For example, if the caller is:
18728 ldloca.s V_1 // V_1 is a local struct
18729 call void Test.ILPart::RunLdargaOnPointerArg(int32*)
18730 and the callee being inlined has:
18731 .method public static void RunLdargaOnPointerArg(int32* ptrToInts) cil managed
18733 call void Test.FourInts::NotInlined_SetExpectedValuesThroughPointerToPointer(int32**)
18734 then we change the argument tree (of "ldloca.s V_1") to TYP_I_IMPL to match the callee signature. We'll
18735 soon afterwards reject the inlining anyway, since the tree we return isn't a GT_LCL_VAR.
18737 assert(argInfo.argNode->TypeGet() == TYP_BYREF || argInfo.argNode->TypeGet() == TYP_I_IMPL);
18738 op1 = gtCloneExpr(argInfo.argNode);
18742 /* Argument is a complex expression - it must be evaluated into a temp */
18744 if (argInfo.argHasTmp)
18746 assert(argInfo.argIsUsed);
18747 assert(argInfo.argTmpNum < lvaCount);
18749 /* Create a new lcl var node - remember the argument lclNum */
18750 op1 = gtNewLclvNode(argInfo.argTmpNum, genActualType(lclTyp));
18752 /* This is the second or later use of the this argument,
18753 so we have to use the temp (instead of the actual arg) */
18754 argInfo.argBashTmpNode = nullptr;
18758 /* First time use */
18759 assert(!argInfo.argIsUsed);
18761 /* Reserve a temp for the expression.
18762 * Use a large size node as we may change it later */
18764 const unsigned tmpNum = lvaGrabTemp(true DEBUGARG("Inlining Arg"));
18766 lvaTable[tmpNum].lvType = lclTyp;
18768 // For ref types, determine the type of the temp.
18769 if (lclTyp == TYP_REF)
18771 if (!argCanBeModified)
18773 // If the arg can't be modified in the method
18774 // body, use the type of the value, if
18775 // known. Otherwise, use the declared type.
18776 lvaSetClass(tmpNum, argInfo.argNode, lclInfo.lclVerTypeInfo.GetClassHandleForObjRef());
18780 // Arg might be modified, use the declared type of
18782 lvaSetClass(tmpNum, lclInfo.lclVerTypeInfo.GetClassHandleForObjRef());
18786 assert(lvaTable[tmpNum].lvAddrExposed == 0);
18787 if (argInfo.argHasLdargaOp)
18789 lvaTable[tmpNum].lvHasLdAddrOp = 1;
18792 if (lclInfo.lclVerTypeInfo.IsStruct())
18794 if (varTypeIsStruct(lclTyp))
18796 lvaSetStruct(tmpNum, lclInfo.lclVerTypeInfo.GetClassHandle(), true /* unsafe value cls check */);
18800 // This is a wrapped primitive. Make sure the verstate knows that
18801 lvaTable[tmpNum].lvVerTypeInfo = lclInfo.lclVerTypeInfo;
18805 argInfo.argHasTmp = true;
18806 argInfo.argTmpNum = tmpNum;
18808 // If we require strict exception order, then arguments must
18809 // be evaluated in sequence before the body of the inlined method.
18810 // So we need to evaluate them to a temp.
18811 // Also, if arguments have global or local references, we need to
18812 // evaluate them to a temp before the inlined body as the
18813 // inlined body may be modifying the global ref.
18814 // TODO-1stClassStructs: We currently do not reuse an existing lclVar
18815 // if it is a struct, because it requires some additional handling.
18817 if (!varTypeIsStruct(lclTyp) && !argInfo.argHasSideEff && !argInfo.argHasGlobRef &&
18818 !argInfo.argHasCallerLocalRef)
18820 /* Get a *LARGE* LCL_VAR node */
18821 op1 = gtNewLclLNode(tmpNum, genActualType(lclTyp), lclNum);
18823 /* Record op1 as the very first use of this argument.
18824 If there are no further uses of the arg, we may be
18825 able to use the actual arg node instead of the temp.
18826 If we do see any further uses, we will clear this. */
18827 argInfo.argBashTmpNode = op1;
18831 /* Get a small LCL_VAR node */
18832 op1 = gtNewLclvNode(tmpNum, genActualType(lclTyp));
18833 /* No bashing of this argument */
18834 argInfo.argBashTmpNode = nullptr;
18839 // Mark this argument as used.
18840 argInfo.argIsUsed = true;
18845 /******************************************************************************
18846 Is this the original "this" argument to the call being inlined?
18848 Note that we do not inline methods with "starg 0", and so we do not need to
18852 BOOL Compiler::impInlineIsThis(GenTree* tree, InlArgInfo* inlArgInfo)
18854 assert(compIsForInlining());
18855 return (tree->gtOper == GT_LCL_VAR && tree->gtLclVarCommon.gtLclNum == inlArgInfo[0].argTmpNum);
18858 //-----------------------------------------------------------------------------
18859 // This function checks if a dereference in the inlinee can guarantee that
18860 // the "this" is non-NULL.
18861 // If we haven't hit a branch or a side effect, and we are dereferencing
18862 // from 'this' to access a field or make GTF_CALL_NULLCHECK call,
18863 // then we can avoid a separate null pointer check.
18865 // "additionalTreesToBeEvaluatedBefore"
18866 // is the set of pending trees that have not yet been added to the statement list,
18867 // and which have been removed from verCurrentState.esStack[]
18869 BOOL Compiler::impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTree* additionalTreesToBeEvaluatedBefore,
18870 GenTree* variableBeingDereferenced,
18871 InlArgInfo* inlArgInfo)
18873 assert(compIsForInlining());
18874 assert(opts.OptEnabled(CLFLG_INLINING));
18876 BasicBlock* block = compCurBB;
18881 if (block != fgFirstBB)
18886 if (!impInlineIsThis(variableBeingDereferenced, inlArgInfo))
18891 if (additionalTreesToBeEvaluatedBefore &&
18892 GTF_GLOBALLY_VISIBLE_SIDE_EFFECTS(additionalTreesToBeEvaluatedBefore->gtFlags))
18897 for (stmt = impTreeList->gtNext; stmt; stmt = stmt->gtNext)
18899 expr = stmt->gtStmt.gtStmtExpr;
18901 if (GTF_GLOBALLY_VISIBLE_SIDE_EFFECTS(expr->gtFlags))
18907 for (unsigned level = 0; level < verCurrentState.esStackDepth; level++)
18909 unsigned stackTreeFlags = verCurrentState.esStack[level].val->gtFlags;
18910 if (GTF_GLOBALLY_VISIBLE_SIDE_EFFECTS(stackTreeFlags))
18919 //------------------------------------------------------------------------
18920 // impMarkInlineCandidate: determine if this call can be subsequently inlined
18923 // callNode -- call under scrutiny
18924 // exactContextHnd -- context handle for inlining
18925 // exactContextNeedsRuntimeLookup -- true if context required runtime lookup
18926 // callInfo -- call info from VM
18929 // If callNode is an inline candidate, this method sets the flag
18930 // GTF_CALL_INLINE_CANDIDATE, and ensures that helper methods have
18931 // filled in the associated InlineCandidateInfo.
18933 // If callNode is not an inline candidate, and the reason is
18934 // something that is inherent to the method being called, the
18935 // method may be marked as "noinline" to short-circuit any
18936 // future assessments of calls to this method.
18938 void Compiler::impMarkInlineCandidate(GenTree* callNode,
18939 CORINFO_CONTEXT_HANDLE exactContextHnd,
18940 bool exactContextNeedsRuntimeLookup,
18941 CORINFO_CALL_INFO* callInfo)
18943 // Let the strategy know there's another call
18944 impInlineRoot()->m_inlineStrategy->NoteCall();
18946 if (!opts.OptEnabled(CLFLG_INLINING))
18948 /* XXX Mon 8/18/2008
18949 * This assert is misleading. The caller does not ensure that we have CLFLG_INLINING set before
18950 * calling impMarkInlineCandidate. However, if this assert trips it means that we're an inlinee and
18951 * CLFLG_MINOPT is set. That doesn't make a lot of sense. If you hit this assert, work back and
18952 * figure out why we did not set MAXOPT for this compile.
18954 assert(!compIsForInlining());
18958 if (compIsForImportOnly())
18960 // Don't bother creating the inline candidate during verification.
18961 // Otherwise the call to info.compCompHnd->canInline will trigger a recursive verification
18962 // that leads to the creation of multiple instances of Compiler.
18966 GenTreeCall* call = callNode->AsCall();
18967 InlineResult inlineResult(this, call, nullptr, "impMarkInlineCandidate");
18969 // Don't inline if not optimizing root method
18970 if (opts.compDbgCode)
18972 inlineResult.NoteFatal(InlineObservation::CALLER_DEBUG_CODEGEN);
18976 // Don't inline if inlining into root method is disabled.
18977 if (InlineStrategy::IsNoInline(info.compCompHnd, info.compMethodHnd))
18979 inlineResult.NoteFatal(InlineObservation::CALLER_IS_JIT_NOINLINE);
18983 // Inlining candidate determination needs to honor only IL tail prefix.
18984 // Inlining takes precedence over implicit tail call optimization (if the call is not directly recursive).
18985 if (call->IsTailPrefixedCall())
18987 inlineResult.NoteFatal(InlineObservation::CALLSITE_EXPLICIT_TAIL_PREFIX);
18991 // Tail recursion elimination takes precedence over inlining.
18992 // TODO: We may want to do some of the additional checks from fgMorphCall
18993 // here to reduce the chance we don't inline a call that won't be optimized
18994 // as a fast tail call or turned into a loop.
18995 if (gtIsRecursiveCall(call) && call->IsImplicitTailCall())
18997 inlineResult.NoteFatal(InlineObservation::CALLSITE_IMPLICIT_REC_TAIL_CALL);
19001 if (call->IsVirtual())
19003 inlineResult.NoteFatal(InlineObservation::CALLSITE_IS_NOT_DIRECT);
19007 /* Ignore helper calls */
19009 if (call->gtCallType == CT_HELPER)
19011 inlineResult.NoteFatal(InlineObservation::CALLSITE_IS_CALL_TO_HELPER);
19015 /* Ignore indirect calls */
19016 if (call->gtCallType == CT_INDIRECT)
19018 inlineResult.NoteFatal(InlineObservation::CALLSITE_IS_NOT_DIRECT_MANAGED);
19022 /* I removed the check for BBJ_THROW. BBJ_THROW is usually marked as rarely run. This more or less
19023 * restricts the inliner to non-expanding inlines. I removed the check to allow for non-expanding
19024 * inlining in throw blocks. I should consider the same thing for catch and filter regions. */
19026 CORINFO_METHOD_HANDLE fncHandle = call->gtCallMethHnd;
19029 // Reuse method flags from the original callInfo if possible
19030 if (fncHandle == callInfo->hMethod)
19032 methAttr = callInfo->methodFlags;
19036 methAttr = info.compCompHnd->getMethodAttribs(fncHandle);
19040 if (compStressCompile(STRESS_FORCE_INLINE, 0))
19042 methAttr |= CORINFO_FLG_FORCEINLINE;
19046 // Check for COMPlus_AggressiveInlining
19047 if (compDoAggressiveInlining)
19049 methAttr |= CORINFO_FLG_FORCEINLINE;
19052 if (!(methAttr & CORINFO_FLG_FORCEINLINE))
19054 /* Don't bother inline blocks that are in the filter region */
19055 if (bbInCatchHandlerILRange(compCurBB))
19060 printf("\nWill not inline blocks that are in the catch handler region\n");
19065 inlineResult.NoteFatal(InlineObservation::CALLSITE_IS_WITHIN_CATCH);
19069 if (bbInFilterILRange(compCurBB))
19074 printf("\nWill not inline blocks that are in the filter region\n");
19078 inlineResult.NoteFatal(InlineObservation::CALLSITE_IS_WITHIN_FILTER);
19083 /* If the caller's stack frame is marked, then we can't do any inlining. Period. */
19085 if (opts.compNeedSecurityCheck)
19087 inlineResult.NoteFatal(InlineObservation::CALLER_NEEDS_SECURITY_CHECK);
19091 /* Check if we tried to inline this method before */
19093 if (methAttr & CORINFO_FLG_DONT_INLINE)
19095 inlineResult.NoteFatal(InlineObservation::CALLEE_IS_NOINLINE);
19099 /* Cannot inline synchronized methods */
19101 if (methAttr & CORINFO_FLG_SYNCH)
19103 inlineResult.NoteFatal(InlineObservation::CALLEE_IS_SYNCHRONIZED);
19107 /* Do not inline if callee needs security checks (since they would then mark the wrong frame) */
19109 if (methAttr & CORINFO_FLG_SECURITYCHECK)
19111 inlineResult.NoteFatal(InlineObservation::CALLEE_NEEDS_SECURITY_CHECK);
19115 InlineCandidateInfo* inlineCandidateInfo = nullptr;
19116 impCheckCanInline(call, fncHandle, methAttr, exactContextHnd, &inlineCandidateInfo, &inlineResult);
19118 if (inlineResult.IsFailure())
19123 // The old value should be NULL
19124 assert(call->gtInlineCandidateInfo == nullptr);
19126 // The new value should not be NULL.
19127 assert(inlineCandidateInfo != nullptr);
19128 inlineCandidateInfo->exactContextNeedsRuntimeLookup = exactContextNeedsRuntimeLookup;
19130 call->gtInlineCandidateInfo = inlineCandidateInfo;
19132 // Mark the call node as inline candidate.
19133 call->gtFlags |= GTF_CALL_INLINE_CANDIDATE;
19135 // Let the strategy know there's another candidate.
19136 impInlineRoot()->m_inlineStrategy->NoteCandidate();
19138 // Since we're not actually inlining yet, and this call site is
19139 // still just an inline candidate, there's nothing to report.
19140 inlineResult.SetReported();
19143 /******************************************************************************/
19144 // Returns true if the given intrinsic will be implemented by target-specific
19147 bool Compiler::IsTargetIntrinsic(CorInfoIntrinsics intrinsicId)
19149 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && !defined(LEGACY_BACKEND))
19150 switch (intrinsicId)
19152 // AMD64/x86 has SSE2 instructions to directly compute sqrt/abs and SSE4.1
19153 // instructions to directly compute round/ceiling/floor.
19155 // TODO: Because the x86 backend only targets SSE for floating-point code,
19156 // it does not treat Sine, Cosine, or Round as intrinsics (JIT32
19157 // implemented those intrinsics as x87 instructions). If this poses
19158 // a CQ problem, it may be necessary to change the implementation of
19159 // the helper calls to decrease call overhead or switch back to the
19160 // x87 instructions. This is tracked by #7097.
19161 case CORINFO_INTRINSIC_Sqrt:
19162 case CORINFO_INTRINSIC_Abs:
19165 case CORINFO_INTRINSIC_Round:
19166 case CORINFO_INTRINSIC_Ceiling:
19167 case CORINFO_INTRINSIC_Floor:
19168 return compSupports(InstructionSet_SSE41);
19173 #elif defined(_TARGET_ARM64_)
19174 switch (intrinsicId)
19176 case CORINFO_INTRINSIC_Sqrt:
19177 case CORINFO_INTRINSIC_Abs:
19178 case CORINFO_INTRINSIC_Round:
19179 case CORINFO_INTRINSIC_Floor:
19180 case CORINFO_INTRINSIC_Ceiling:
19186 #elif defined(_TARGET_ARM_)
19187 switch (intrinsicId)
19189 case CORINFO_INTRINSIC_Sqrt:
19190 case CORINFO_INTRINSIC_Abs:
19191 case CORINFO_INTRINSIC_Round:
19197 #elif defined(_TARGET_X86_)
19198 switch (intrinsicId)
19200 case CORINFO_INTRINSIC_Sin:
19201 case CORINFO_INTRINSIC_Cos:
19202 case CORINFO_INTRINSIC_Sqrt:
19203 case CORINFO_INTRINSIC_Abs:
19204 case CORINFO_INTRINSIC_Round:
19211 // TODO: This portion of logic is not implemented for other arch.
19212 // The reason for returning true is that on all other arch the only intrinsic
19213 // enabled are target intrinsics.
19215 #endif //_TARGET_AMD64_
19218 /******************************************************************************/
19219 // Returns true if the given intrinsic will be implemented by calling System.Math
19222 bool Compiler::IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId)
19224 // Currently, if a math intrinsic is not implemented by target-specific
19225 // instructions, it will be implemented by a System.Math call. In the
19226 // future, if we turn to implementing some of them with helper calls,
19227 // this predicate needs to be revisited.
19228 return !IsTargetIntrinsic(intrinsicId);
19231 bool Compiler::IsMathIntrinsic(CorInfoIntrinsics intrinsicId)
19233 switch (intrinsicId)
19235 case CORINFO_INTRINSIC_Sin:
19236 case CORINFO_INTRINSIC_Cbrt:
19237 case CORINFO_INTRINSIC_Sqrt:
19238 case CORINFO_INTRINSIC_Abs:
19239 case CORINFO_INTRINSIC_Cos:
19240 case CORINFO_INTRINSIC_Round:
19241 case CORINFO_INTRINSIC_Cosh:
19242 case CORINFO_INTRINSIC_Sinh:
19243 case CORINFO_INTRINSIC_Tan:
19244 case CORINFO_INTRINSIC_Tanh:
19245 case CORINFO_INTRINSIC_Asin:
19246 case CORINFO_INTRINSIC_Asinh:
19247 case CORINFO_INTRINSIC_Acos:
19248 case CORINFO_INTRINSIC_Acosh:
19249 case CORINFO_INTRINSIC_Atan:
19250 case CORINFO_INTRINSIC_Atan2:
19251 case CORINFO_INTRINSIC_Atanh:
19252 case CORINFO_INTRINSIC_Log10:
19253 case CORINFO_INTRINSIC_Pow:
19254 case CORINFO_INTRINSIC_Exp:
19255 case CORINFO_INTRINSIC_Ceiling:
19256 case CORINFO_INTRINSIC_Floor:
19263 bool Compiler::IsMathIntrinsic(GenTree* tree)
19265 return (tree->OperGet() == GT_INTRINSIC) && IsMathIntrinsic(tree->gtIntrinsic.gtIntrinsicId);
19268 //------------------------------------------------------------------------
19269 // impDevirtualizeCall: Attempt to change a virtual vtable call into a
19273 // call -- the call node to examine/modify
19274 // method -- [IN/OUT] the method handle for call. Updated iff call devirtualized.
19275 // methodFlags -- [IN/OUT] flags for the method to call. Updated iff call devirtualized.
19276 // contextHandle -- [IN/OUT] context handle for the call. Updated iff call devirtualized.
19277 // exactContextHnd -- [OUT] updated context handle iff call devirtualized
19280 // Virtual calls in IL will always "invoke" the base class method.
19282 // This transformation looks for evidence that the type of 'this'
19283 // in the call is exactly known, is a final class or would invoke
19284 // a final method, and if that and other safety checks pan out,
19285 // modifies the call and the call info to create a direct call.
19287 // This transformation is initially done in the importer and not
19288 // in some subsequent optimization pass because we want it to be
19289 // upstream of inline candidate identification.
19291 // However, later phases may supply improved type information that
19292 // can enable further devirtualization. We currently reinvoke this
19293 // code after inlining, if the return value of the inlined call is
19294 // the 'this obj' of a subsequent virtual call.
19296 // If devirtualization succeeds and the call's this object is the
19297 // result of a box, the jit will ask the EE for the unboxed entry
19298 // point. If this exists, the jit will see if it can rework the box
19299 // to instead make a local copy. If that is doable, the call is
19300 // updated to invoke the unboxed entry on the local copy.
19302 void Compiler::impDevirtualizeCall(GenTreeCall* call,
19303 CORINFO_METHOD_HANDLE* method,
19304 unsigned* methodFlags,
19305 CORINFO_CONTEXT_HANDLE* contextHandle,
19306 CORINFO_CONTEXT_HANDLE* exactContextHandle)
19308 assert(call != nullptr);
19309 assert(method != nullptr);
19310 assert(methodFlags != nullptr);
19311 assert(contextHandle != nullptr);
19313 // This should be a virtual vtable or virtual stub call.
19314 assert(call->IsVirtual());
19316 // Bail if not optimizing
19317 if (opts.MinOpts())
19322 // Bail if debuggable codegen
19323 if (opts.compDbgCode)
19329 // Bail if devirt is disabled.
19330 if (JitConfig.JitEnableDevirtualization() == 0)
19335 const bool doPrint = JitConfig.JitPrintDevirtualizedMethods() == 1;
19338 // Fetch information about the virtual method we're calling.
19339 CORINFO_METHOD_HANDLE baseMethod = *method;
19340 unsigned baseMethodAttribs = *methodFlags;
19342 if (baseMethodAttribs == 0)
19344 // For late devirt we may not have method attributes, so fetch them.
19345 baseMethodAttribs = info.compCompHnd->getMethodAttribs(baseMethod);
19350 // Validate that callInfo has up to date method flags
19351 const DWORD freshBaseMethodAttribs = info.compCompHnd->getMethodAttribs(baseMethod);
19353 // All the base method attributes should agree, save that
19354 // CORINFO_FLG_DONT_INLINE may have changed from 0 to 1
19355 // because of concurrent jitting activity.
19357 // Note we don't look at this particular flag bit below, and
19358 // later on (if we do try and inline) we will rediscover why
19359 // the method can't be inlined, so there's no danger here in
19360 // seeing this particular flag bit in different states between
19361 // the cached and fresh values.
19362 if ((freshBaseMethodAttribs & ~CORINFO_FLG_DONT_INLINE) != (baseMethodAttribs & ~CORINFO_FLG_DONT_INLINE))
19364 assert(!"mismatched method attributes");
19369 // In R2R mode, we might see virtual stub calls to
19370 // non-virtuals. For instance cases where the non-virtual method
19371 // is in a different assembly but is called via CALLVIRT. For
19372 // verison resilience we must allow for the fact that the method
19373 // might become virtual in some update.
19375 // In non-R2R modes CALLVIRT <nonvirtual> will be turned into a
19376 // regular call+nullcheck upstream, so we won't reach this
19378 if ((baseMethodAttribs & CORINFO_FLG_VIRTUAL) == 0)
19380 assert(call->IsVirtualStub());
19381 assert(opts.IsReadyToRun());
19382 JITDUMP("\nimpDevirtualizeCall: [R2R] base method not virtual, sorry\n");
19386 // See what we know about the type of 'this' in the call.
19387 GenTree* thisObj = call->gtCallObjp->gtEffectiveVal(false);
19388 GenTree* actualThisObj = nullptr;
19389 bool isExact = false;
19390 bool objIsNonNull = false;
19391 CORINFO_CLASS_HANDLE objClass = gtGetClassHandle(thisObj, &isExact, &objIsNonNull);
19393 // See if we have special knowlege that can get us a type or a better type.
19394 if ((objClass == nullptr) || !isExact)
19396 actualThisObj = thisObj;
19398 // Walk back through any return expression placeholders
19399 while (actualThisObj->OperGet() == GT_RET_EXPR)
19401 actualThisObj = actualThisObj->gtRetExpr.gtInlineCandidate;
19404 // See if we landed on a call to a special intrinsic method
19405 if (actualThisObj->IsCall())
19407 GenTreeCall* thisObjCall = actualThisObj->AsCall();
19408 if ((thisObjCall->gtCallMoreFlags & GTF_CALL_M_SPECIAL_INTRINSIC) != 0)
19410 assert(thisObjCall->gtCallType == CT_USER_FUNC);
19411 CORINFO_METHOD_HANDLE specialIntrinsicHandle = thisObjCall->gtCallMethHnd;
19412 CORINFO_CLASS_HANDLE specialObjClass = impGetSpecialIntrinsicExactReturnType(specialIntrinsicHandle);
19413 if (specialObjClass != nullptr)
19415 objClass = specialObjClass;
19417 objIsNonNull = true;
19423 // Bail if we know nothing.
19424 if (objClass == nullptr)
19426 JITDUMP("\nimpDevirtualizeCall: no type available (op=%s)\n", GenTree::OpName(thisObj->OperGet()));
19430 // Fetch information about the class that introduced the virtual method.
19431 CORINFO_CLASS_HANDLE baseClass = info.compCompHnd->getMethodClass(baseMethod);
19432 const DWORD baseClassAttribs = info.compCompHnd->getClassAttribs(baseClass);
19434 #if !defined(FEATURE_CORECLR)
19435 // If base class is not beforefieldinit then devirtualizing may
19436 // cause us to miss a base class init trigger. Spec says we don't
19437 // need a trigger for ref class callvirts but desktop seems to
19438 // have one anyways. So defer.
19439 if ((baseClassAttribs & CORINFO_FLG_BEFOREFIELDINIT) == 0)
19441 JITDUMP("\nimpDevirtualizeCall: base class has precise initialization, sorry\n");
19444 #endif // FEATURE_CORECLR
19446 // Is the call an interface call?
19447 const bool isInterface = (baseClassAttribs & CORINFO_FLG_INTERFACE) != 0;
19449 // If the objClass is sealed (final), then we may be able to devirtualize.
19450 const DWORD objClassAttribs = info.compCompHnd->getClassAttribs(objClass);
19451 const bool objClassIsFinal = (objClassAttribs & CORINFO_FLG_FINAL) != 0;
19454 const char* callKind = isInterface ? "interface" : "virtual";
19455 const char* objClassNote = "[?]";
19456 const char* objClassName = "?objClass";
19457 const char* baseClassName = "?baseClass";
19458 const char* baseMethodName = "?baseMethod";
19460 if (verbose || doPrint)
19462 objClassNote = isExact ? " [exact]" : objClassIsFinal ? " [final]" : "";
19463 objClassName = info.compCompHnd->getClassName(objClass);
19464 baseClassName = info.compCompHnd->getClassName(baseClass);
19465 baseMethodName = eeGetMethodName(baseMethod, nullptr);
19469 printf("\nimpDevirtualizeCall: Trying to devirtualize %s call:\n"
19470 " class for 'this' is %s%s (attrib %08x)\n"
19471 " base method is %s::%s\n",
19472 callKind, objClassName, objClassNote, objClassAttribs, baseClassName, baseMethodName);
19475 #endif // defined(DEBUG)
19477 // Bail if obj class is an interface.
19478 // See for instance System.ValueTuple`8::GetHashCode, where lcl 0 is System.IValueTupleInternal
19479 // IL_021d: ldloc.0
19480 // IL_021e: callvirt instance int32 System.Object::GetHashCode()
19481 if ((objClassAttribs & CORINFO_FLG_INTERFACE) != 0)
19483 JITDUMP("--- obj class is interface, sorry\n");
19489 assert(call->IsVirtualStub());
19490 JITDUMP("--- base class is interface\n");
19493 // Fetch the method that would be called based on the declared type of 'this'
19494 CORINFO_CONTEXT_HANDLE ownerType = *contextHandle;
19495 CORINFO_METHOD_HANDLE derivedMethod = info.compCompHnd->resolveVirtualMethod(baseMethod, objClass, ownerType);
19497 // If we failed to get a handle, we can't devirtualize. This can
19498 // happen when prejitting, if the devirtualization crosses
19499 // servicing bubble boundaries.
19500 if (derivedMethod == nullptr)
19502 JITDUMP("--- no derived method, sorry\n");
19506 // Fetch method attributes to see if method is marked final.
19507 DWORD derivedMethodAttribs = info.compCompHnd->getMethodAttribs(derivedMethod);
19508 const bool derivedMethodIsFinal = ((derivedMethodAttribs & CORINFO_FLG_FINAL) != 0);
19511 const char* derivedClassName = "?derivedClass";
19512 const char* derivedMethodName = "?derivedMethod";
19514 const char* note = "speculative";
19519 else if (objClassIsFinal)
19521 note = "final class";
19523 else if (derivedMethodIsFinal)
19525 note = "final method";
19528 if (verbose || doPrint)
19530 derivedMethodName = eeGetMethodName(derivedMethod, &derivedClassName);
19533 printf(" devirt to %s::%s -- %s\n", derivedClassName, derivedMethodName, note);
19537 #endif // defined(DEBUG)
19539 if (!isExact && !objClassIsFinal && !derivedMethodIsFinal)
19541 // Type is not exact, and neither class or method is final.
19543 // We could speculatively devirtualize, but there's no
19544 // reason to believe the derived method is the one that
19545 // is likely to be invoked.
19547 // If there's currently no further overriding (that is, at
19548 // the time of jitting, objClass has no subclasses that
19549 // override this method), then perhaps we'd be willing to
19551 JITDUMP(" Class not final or exact, method not final, no devirtualization\n");
19555 // For interface calls we must have an exact type or final class.
19556 if (isInterface && !isExact && !objClassIsFinal)
19558 JITDUMP(" Class not final or exact for interface, no devirtualization\n");
19562 JITDUMP(" %s; can devirtualize\n", note);
19564 // Make the updates.
19565 call->gtFlags &= ~GTF_CALL_VIRT_VTABLE;
19566 call->gtFlags &= ~GTF_CALL_VIRT_STUB;
19567 call->gtCallMethHnd = derivedMethod;
19568 call->gtCallType = CT_USER_FUNC;
19570 // Virtual calls include an implicit null check, which we may
19571 // now need to make explicit.
19574 call->gtFlags |= GTF_CALL_NULLCHECK;
19577 // Clear the inline candidate info (may be non-null since
19578 // it's a union field used for other things by virtual
19580 call->gtInlineCandidateInfo = nullptr;
19585 printf("... after devirt...\n");
19591 printf("Devirtualized %s call to %s:%s; now direct call to %s:%s [%s]\n", callKind, baseClassName,
19592 baseMethodName, derivedClassName, derivedMethodName, note);
19594 #endif // defined(DEBUG)
19596 // If the 'this' object is a box, see if we can find the unboxed entry point for the call.
19597 if (thisObj->IsBoxedValue())
19599 JITDUMP("Now have direct call to boxed entry point, looking for unboxed entry point\n");
19601 // Note for some shared methods the unboxed entry point requires an extra parameter.
19602 bool requiresInstMethodTableArg = false;
19603 CORINFO_METHOD_HANDLE unboxedEntryMethod =
19604 info.compCompHnd->getUnboxedEntry(derivedMethod, &requiresInstMethodTableArg);
19606 if (unboxedEntryMethod != nullptr)
19608 // Since the call is the only consumer of the box, we know the box can't escape
19609 // since it is being passed an interior pointer.
19611 // So, revise the box to simply create a local copy, use the address of that copy
19612 // as the this pointer, and update the entry point to the unboxed entry.
19614 // Ideally, we then inline the boxed method and and if it turns out not to modify
19615 // the copy, we can undo the copy too.
19616 if (requiresInstMethodTableArg)
19618 // Perform a trial box removal and ask for the type handle tree.
19619 JITDUMP("Unboxed entry needs method table arg...\n");
19620 GenTree* methodTableArg = gtTryRemoveBoxUpstreamEffects(thisObj, BR_DONT_REMOVE_WANT_TYPE_HANDLE);
19622 if (methodTableArg != nullptr)
19624 // If that worked, turn the box into a copy to a local var
19625 JITDUMP("Found suitable method table arg tree [%06u]\n", dspTreeID(methodTableArg));
19626 GenTree* localCopyThis = gtTryRemoveBoxUpstreamEffects(thisObj, BR_MAKE_LOCAL_COPY);
19628 if (localCopyThis != nullptr)
19630 // Pass the local var as this and the type handle as a new arg
19631 JITDUMP("Success! invoking unboxed entry point on local copy, and passing method table arg\n");
19632 call->gtCallObjp = localCopyThis;
19634 // Prepend for R2L arg passing or empty L2R passing
19635 if ((Target::g_tgtArgOrder == Target::ARG_ORDER_R2L) || (call->gtCallArgs == nullptr))
19637 call->gtCallArgs = gtNewListNode(methodTableArg, call->gtCallArgs);
19639 // Append for non-empty L2R
19642 GenTreeArgList* beforeArg = call->gtCallArgs;
19643 while (beforeArg->Rest() != nullptr)
19645 beforeArg = beforeArg->Rest();
19648 beforeArg->Rest() = gtNewListNode(methodTableArg, nullptr);
19651 call->gtCallMethHnd = unboxedEntryMethod;
19652 derivedMethod = unboxedEntryMethod;
19654 // Method attributes will differ because unboxed entry point is shared
19655 const DWORD unboxedMethodAttribs = info.compCompHnd->getMethodAttribs(unboxedEntryMethod);
19656 JITDUMP("Updating method attribs from 0x%08x to 0x%08x\n", derivedMethodAttribs,
19657 unboxedMethodAttribs);
19658 derivedMethodAttribs = unboxedMethodAttribs;
19662 JITDUMP("Sorry, failed to undo the box -- can't convert to local copy\n");
19667 JITDUMP("Sorry, failed to undo the box -- can't find method table arg\n");
19672 JITDUMP("Found unboxed entry point, trying to simplify box to a local copy\n");
19673 GenTree* localCopyThis = gtTryRemoveBoxUpstreamEffects(thisObj, BR_MAKE_LOCAL_COPY);
19675 if (localCopyThis != nullptr)
19677 JITDUMP("Success! invoking unboxed entry point on local copy\n");
19678 call->gtCallObjp = localCopyThis;
19679 call->gtCallMethHnd = unboxedEntryMethod;
19680 derivedMethod = unboxedEntryMethod;
19684 JITDUMP("Sorry, failed to undo the box\n");
19690 // Many of the low-level methods on value classes won't have unboxed entries,
19691 // as they need access to the type of the object.
19693 // Note this may be a cue for us to stack allocate the boxed object, since
19694 // we probably know that these objects don't escape.
19695 JITDUMP("Sorry, failed to find unboxed entry point\n");
19699 // Fetch the class that introduced the derived method.
19701 // Note this may not equal objClass, if there is a
19702 // final method that objClass inherits.
19703 CORINFO_CLASS_HANDLE derivedClass = info.compCompHnd->getMethodClass(derivedMethod);
19705 // Need to update call info too. This is fragile
19706 // but hopefully the derived method conforms to
19707 // the base in most other ways.
19708 *method = derivedMethod;
19709 *methodFlags = derivedMethodAttribs;
19710 *contextHandle = MAKE_METHODCONTEXT(derivedMethod);
19712 // Update context handle.
19713 if ((exactContextHandle != nullptr) && (*exactContextHandle != nullptr))
19715 *exactContextHandle = MAKE_METHODCONTEXT(derivedMethod);
19718 #ifdef FEATURE_READYTORUN_COMPILER
19719 if (opts.IsReadyToRun())
19721 // For R2R, getCallInfo triggers bookkeeping on the zap
19722 // side so we need to call it here.
19724 // First, cons up a suitable resolved token.
19725 CORINFO_RESOLVED_TOKEN derivedResolvedToken = {};
19727 derivedResolvedToken.tokenScope = info.compScopeHnd;
19728 derivedResolvedToken.tokenContext = *contextHandle;
19729 derivedResolvedToken.token = info.compCompHnd->getMethodDefFromMethod(derivedMethod);
19730 derivedResolvedToken.tokenType = CORINFO_TOKENKIND_Method;
19731 derivedResolvedToken.hClass = derivedClass;
19732 derivedResolvedToken.hMethod = derivedMethod;
19734 // Look up the new call info.
19735 CORINFO_CALL_INFO derivedCallInfo;
19736 eeGetCallInfo(&derivedResolvedToken, nullptr, addVerifyFlag(CORINFO_CALLINFO_ALLOWINSTPARAM), &derivedCallInfo);
19738 // Update the call.
19739 call->gtCallMoreFlags &= ~GTF_CALL_M_VIRTSTUB_REL_INDIRECT;
19740 call->gtCallMoreFlags &= ~GTF_CALL_M_R2R_REL_INDIRECT;
19741 call->setEntryPoint(derivedCallInfo.codePointerLookup.constLookup);
19743 #endif // FEATURE_READYTORUN_COMPILER
19746 //------------------------------------------------------------------------
19747 // impGetSpecialIntrinsicExactReturnType: Look for special cases where a call
19748 // to an intrinsic returns an exact type
19751 // methodHnd -- handle for the special intrinsic method
19754 // Exact class handle returned by the intrinsic call, if known.
19755 // Nullptr if not known, or not likely to lead to beneficial optimization.
19757 CORINFO_CLASS_HANDLE Compiler::impGetSpecialIntrinsicExactReturnType(CORINFO_METHOD_HANDLE methodHnd)
19759 JITDUMP("Special intrinsic: looking for exact type returned by %s\n", eeGetMethodFullName(methodHnd));
19761 CORINFO_CLASS_HANDLE result = nullptr;
19763 // See what intrinisc we have...
19764 const NamedIntrinsic ni = lookupNamedIntrinsic(methodHnd);
19767 case NI_System_Collections_Generic_EqualityComparer_get_Default:
19769 // Expect one class generic parameter; figure out which it is.
19770 CORINFO_SIG_INFO sig;
19771 info.compCompHnd->getMethodSig(methodHnd, &sig);
19772 assert(sig.sigInst.classInstCount == 1);
19773 CORINFO_CLASS_HANDLE typeHnd = sig.sigInst.classInst[0];
19774 assert(typeHnd != nullptr);
19776 // Lookup can incorrect when we have __Canon as it won't appear
19777 // to implement any interface types.
19779 // And if we do not have a final type, devirt & inlining is
19780 // unlikely to result in much simplification.
19782 // We can use CORINFO_FLG_FINAL to screen out both of these cases.
19783 const DWORD typeAttribs = info.compCompHnd->getClassAttribs(typeHnd);
19784 const bool isFinalType = ((typeAttribs & CORINFO_FLG_FINAL) != 0);
19788 result = info.compCompHnd->getDefaultEqualityComparerClass(typeHnd);
19789 JITDUMP("Special intrinsic for type %s: return type is %s\n", eeGetClassName(typeHnd),
19790 result != nullptr ? eeGetClassName(result) : "unknown");
19794 JITDUMP("Special intrinsic for type %s: type not final, so deferring opt\n", eeGetClassName(typeHnd));
19802 JITDUMP("This special intrinsic not handled, sorry...\n");
19810 //------------------------------------------------------------------------
19811 // impAllocateToken: create CORINFO_RESOLVED_TOKEN into jit-allocated memory and init it.
19814 // token - init value for the allocated token.
19817 // pointer to token into jit-allocated memory.
19818 CORINFO_RESOLVED_TOKEN* Compiler::impAllocateToken(CORINFO_RESOLVED_TOKEN token)
19820 CORINFO_RESOLVED_TOKEN* memory = (CORINFO_RESOLVED_TOKEN*)compGetMem(sizeof(token));
19825 //------------------------------------------------------------------------
19826 // SpillRetExprHelper: iterate through arguments tree and spill ret_expr to local varibales.
19828 class SpillRetExprHelper
19831 SpillRetExprHelper(Compiler* comp) : comp(comp)
19835 void StoreRetExprResultsInArgs(GenTreeCall* call)
19837 GenTree* args = call->gtCallArgs;
19838 if (args != nullptr)
19840 comp->fgWalkTreePre(&args, SpillRetExprVisitor, this);
19842 GenTree* thisArg = call->gtCallObjp;
19843 if (thisArg != nullptr)
19845 comp->fgWalkTreePre(&thisArg, SpillRetExprVisitor, this);
19850 static Compiler::fgWalkResult SpillRetExprVisitor(GenTree** pTree, Compiler::fgWalkData* fgWalkPre)
19852 assert((pTree != nullptr) && (*pTree != nullptr));
19853 GenTree* tree = *pTree;
19854 if ((tree->gtFlags & GTF_CALL) == 0)
19856 // Trees with ret_expr are marked as GTF_CALL.
19857 return Compiler::WALK_SKIP_SUBTREES;
19859 if (tree->OperGet() == GT_RET_EXPR)
19861 SpillRetExprHelper* walker = static_cast<SpillRetExprHelper*>(fgWalkPre->pCallbackData);
19862 walker->StoreRetExprAsLocalVar(pTree);
19864 return Compiler::WALK_CONTINUE;
19867 void StoreRetExprAsLocalVar(GenTree** pRetExpr)
19869 GenTree* retExpr = *pRetExpr;
19870 assert(retExpr->OperGet() == GT_RET_EXPR);
19871 JITDUMP("Store return expression %u as a local var.\n", retExpr->gtTreeID);
19872 unsigned tmp = comp->lvaGrabTemp(true DEBUGARG("spilling ret_expr"));
19873 comp->impAssignTempGen(tmp, retExpr, (unsigned)Compiler::CHECK_SPILL_NONE);
19874 *pRetExpr = comp->gtNewLclvNode(tmp, retExpr->TypeGet());
19881 //------------------------------------------------------------------------
19882 // addFatPointerCandidate: mark the call and the method, that they have a fat pointer candidate.
19883 // Spill ret_expr in the call node, because they can't be cloned.
19886 // call - fat calli candidate
19888 void Compiler::addFatPointerCandidate(GenTreeCall* call)
19890 setMethodHasFatPointer();
19891 call->SetFatPointerCandidate();
19892 SpillRetExprHelper helper(this);
19893 helper.StoreRetExprResultsInArgs(call);