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 Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
104 // Or the more-general IAllocator interface.
105 void* __cdecl operator new(size_t n, IAllocator* alloc);
106 void* __cdecl operator new[](size_t n, IAllocator* alloc);
108 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
109 // caused these to be ambiguous with the global placement new operators.
110 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
111 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
114 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
115 #include "loopcloning.h"
117 /*****************************************************************************/
119 /* This is included here and not earlier as it needs the definition of "CSE"
120 * which is defined in the section above */
122 /*****************************************************************************/
124 unsigned genLog2(unsigned value);
125 unsigned genLog2(unsigned __int64 value);
127 var_types genActualType(var_types type);
128 var_types genUnsignedType(var_types type);
129 var_types genSignedType(var_types type);
131 unsigned ReinterpretHexAsDecimal(unsigned);
133 /*****************************************************************************/
136 #ifdef FEATURE_AVX_SUPPORT
137 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
138 #else // !FEATURE_AVX_SUPPORT
139 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
140 #endif // !FEATURE_AVX_SUPPORT
141 #else // !FEATURE_SIMD
142 const unsigned TEMP_MAX_SIZE = sizeof(double);
143 #endif // !FEATURE_SIMD
144 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
146 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
149 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
152 // The following holds the Local var info (scope information)
153 typedef const char* VarName; // Actual ASCII string
156 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
157 IL_OFFSET vsdLifeEnd; // instr offset of end of life
158 unsigned vsdVarNum; // (remapped) LclVarDsc number
161 VarName vsdName; // name of the var
164 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
165 // Also, it is the index of this entry in the info.compVarScopes array,
166 // which is useful since the array is also accessed via the
167 // compEnterScopeList and compExitScopeList sorted arrays.
170 /*****************************************************************************
172 * The following holds the local variable counts and the descriptor table.
175 // This is the location of a definition.
181 DefLoc() : m_blk(nullptr), m_tree(nullptr)
186 // This class encapsulates all info about a local variable that may vary for different SSA names
191 ValueNumPair m_vnPair;
199 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
204 // The constructor. Most things can just be zero'ed.
205 LclVarDsc(Compiler* comp);
207 // note this only packs because var_types is a typedef of unsigned char
208 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
210 unsigned char lvIsParam : 1; // is this a parameter?
211 unsigned char lvIsRegArg : 1; // is this a register argument?
212 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
214 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
215 // is to help determine whether to use block init in the prolog.
216 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
217 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
218 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
219 // variable is in the same register for the entire function.
220 unsigned char lvTracked : 1; // is this a tracked variable?
221 bool lvTrackedNonStruct()
223 return lvTracked && lvType != TYP_STRUCT;
225 unsigned char lvPinned : 1; // is this a pinned variable?
227 unsigned char lvMustInit : 1; // must be initialized
228 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
229 // global location, etc.
230 // We cannot reason reliably about the value of the variable.
231 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
232 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
236 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
238 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
239 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
240 // read/write its value.
241 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
242 // the variable to be
243 // in the stack (at least at those boundaries.)
244 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
245 // particular byte from an int).
246 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
247 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
249 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
250 unsigned char lvRefAssign : 1; // involved in pointer assignment
251 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
252 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
255 unsigned char lvArgWrite : 1; // variable is a parameter and STARG was used on it
256 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp
258 unsigned char lvIsBoolean : 1; // set if variable is boolean
260 unsigned char lvRngOptDone : 1; // considered for range check opt?
261 unsigned char lvLoopInc : 1; // incremented in the loop?
262 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
263 unsigned char lvArrIndx : 1; // used as an array index?
264 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
265 unsigned char lvArrIndxDom : 1; // index dominates loop exit
267 unsigned char lvSingleDef : 1; // variable has a single def
268 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
269 unsigned char lvVolatileHint : 1; // hint for AssertionProp
272 unsigned char lvSpilled : 1; // enregistered variable was spilled
273 #ifndef _TARGET_64BIT_
274 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
275 #endif // !_TARGET_64BIT_
276 #ifdef _TARGET_64BIT_
277 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
280 unsigned char lvKeepType : 1; // Don't change the type of this variable
281 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
283 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
285 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
286 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
288 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
289 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
290 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
291 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
292 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
294 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
295 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
298 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
299 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
300 // with (lvIsRegArg && lvIsHfa())
301 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
302 #endif // FEATURE_HFA
305 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
306 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
307 // Consider cleaning this up so this workaround is not required.
308 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
309 // I.e. there is no longer any reference to the struct directly.
310 // In this case we can simply remove this struct local.
312 #ifndef LEGACY_BACKEND
313 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
314 #endif // !LEGACY_BACKEND
317 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
318 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
319 unsigned char lvSIMDType : 1; // This is a SIMD struct
320 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
321 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
322 #endif // FEATURE_SIMD
323 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
326 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
328 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
329 // Valid on promoted struct local fields.
332 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
333 unsigned char lvFldOffset;
334 unsigned char lvFldOrdinal;
336 #if FEATURE_MULTIREG_ARGS
337 regNumber lvRegNumForSlot(unsigned slotNum)
343 else if (slotNum == 1)
345 return lvOtherArgReg;
349 assert(false && "Invalid slotNum!");
354 #endif // FEATURE_MULTIREG_ARGS
372 bool lvIsHfaRegArg() const
375 return _lvIsHfaRegArg;
381 void lvSetIsHfaRegArg()
384 _lvIsHfaRegArg = true;
388 bool lvHfaTypeIsFloat() const
391 return _lvHfaTypeIsFloat;
397 void lvSetHfaTypeIsFloat(bool value)
400 _lvHfaTypeIsFloat = value;
404 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
405 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
407 unsigned lvHfaSlots() const
410 assert(lvType == TYP_STRUCT);
412 return lvExactSize / sizeof(float);
413 #else // _TARGET_ARM64_
414 if (lvHfaTypeIsFloat())
416 return lvExactSize / sizeof(float);
420 return lvExactSize / sizeof(double);
422 #endif // _TARGET_ARM64_
425 // lvIsMultiRegArgOrRet()
426 // returns true if this is a multireg LclVar struct used in an argument context
427 // or if this is a multireg LclVar struct assigned from a multireg call
428 bool lvIsMultiRegArgOrRet()
430 return lvIsMultiRegArg || lvIsMultiRegRet;
434 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
435 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
436 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
437 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
438 // to non-zero if the variable gets the same register assignment for its entire
440 #if !defined(_TARGET_64BIT_)
441 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
442 #endif // !defined(_TARGET_64BIT_)
444 regNumberSmall _lvArgReg; // The register in which this argument is passed.
446 #if FEATURE_MULTIREG_ARGS
447 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
448 // Note this is defined but not used by ARM32
449 #endif // FEATURE_MULTIREG_ARGS
451 #ifndef LEGACY_BACKEND
453 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
454 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
456 #endif // !LEGACY_BACKEND
459 // The register number is stored in a small format (8 bits), but the getters return and the setters take
460 // a full-size (unsigned) format, to localize the casts here.
462 /////////////////////
464 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
466 regNumber GetRegNum() const
468 return (regNumber)_lvRegNum;
471 void SetRegNum(regNumber reg)
473 _lvRegNum = (regNumberSmall)reg;
474 assert(_lvRegNum == reg);
477 /////////////////////
479 #if defined(_TARGET_64BIT_)
480 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
482 regNumber GetOtherReg() const
484 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
485 // "unreachable code" warnings
489 void SetOtherReg(regNumber reg)
491 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
492 // "unreachable code" warnings
494 #else // !_TARGET_64BIT_
495 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
497 regNumber GetOtherReg() const
499 return (regNumber)_lvOtherReg;
502 void SetOtherReg(regNumber reg)
504 _lvOtherReg = (regNumberSmall)reg;
505 assert(_lvOtherReg == reg);
507 #endif // !_TARGET_64BIT_
509 /////////////////////
511 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
513 regNumber GetArgReg() const
515 return (regNumber)_lvArgReg;
518 void SetArgReg(regNumber reg)
520 _lvArgReg = (regNumberSmall)reg;
521 assert(_lvArgReg == reg);
524 #if FEATURE_MULTIREG_ARGS
525 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
527 regNumber GetOtherArgReg() const
529 return (regNumber)_lvOtherArgReg;
532 void SetOtherArgReg(regNumber reg)
534 _lvOtherArgReg = (regNumberSmall)reg;
535 assert(_lvOtherArgReg == reg);
537 #endif // FEATURE_MULTIREG_ARGS
540 // Is this is a SIMD struct?
541 bool lvIsSIMDType() const
546 // Is this is a SIMD struct which is used for SIMD intrinsic?
547 bool lvIsUsedInSIMDIntrinsic() const
549 return lvUsedInSIMDIntrinsic;
552 // If feature_simd not enabled, return false
553 bool lvIsSIMDType() const
557 bool lvIsUsedInSIMDIntrinsic() const
563 /////////////////////
565 #ifndef LEGACY_BACKEND
566 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
568 regNumber GetArgInitReg() const
570 return (regNumber)_lvArgInitReg;
573 void SetArgInitReg(regNumber reg)
575 _lvArgInitReg = (regNumberSmall)reg;
576 assert(_lvArgInitReg == reg);
579 /////////////////////
581 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
583 regPairNo GetArgInitRegPair() const
585 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
586 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
590 void SetArgInitRegPair(regPairNo regPair)
592 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
593 _lvArgInitRegPair = (regPairNoSmall)regPair;
594 assert(_lvArgInitRegPair == regPair);
597 /////////////////////
599 bool lvIsRegCandidate() const
601 return lvLRACandidate != 0;
604 bool lvIsInReg() const
606 return lvIsRegCandidate() && (lvRegNum != REG_STK);
609 #else // LEGACY_BACKEND
611 bool lvIsRegCandidate() const
613 return lvTracked != 0;
616 bool lvIsInReg() const
618 return lvRegister != 0;
621 #endif // LEGACY_BACKEND
623 regMaskTP lvRegMask() const
625 regMaskTP regMask = RBM_NONE;
626 if (varTypeIsFloating(TypeGet()))
628 if (lvRegNum != REG_STK)
630 regMask = genRegMaskFloat(lvRegNum, TypeGet());
635 if (lvRegNum != REG_STK)
637 regMask = genRegMask(lvRegNum);
640 // For longs we may have two regs
641 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
643 regMask |= genRegMask(lvOtherReg);
649 regMaskSmall lvPrefReg; // set of regs it prefers to live in
651 unsigned short lvVarIndex; // variable tracking index
652 unsigned short lvRefCnt; // unweighted (real) reference count
653 unsigned lvRefCntWtd; // weighted reference count
654 int lvStkOffs; // stack offset of home
655 unsigned lvExactSize; // (exact) size of the type in bytes
657 // Is this a promoted struct?
658 // This method returns true only for structs (including SIMD structs), not for
659 // locals that are split on a 32-bit target.
660 // It is only necessary to use this:
661 // 1) if only structs are wanted, and
662 // 2) if Lowering has already been done.
663 // Otherwise lvPromoted is valid.
664 bool lvPromotedStruct()
666 #if !defined(_TARGET_64BIT_)
667 return (lvPromoted && !varTypeIsLong(lvType));
668 #else // defined(_TARGET_64BIT_)
670 #endif // defined(_TARGET_64BIT_)
673 unsigned lvSize() // Size needed for storage representation. Only used for structs or TYP_BLK.
675 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
676 // where the struct itself is no longer used because all access is via its member fields.
677 // When that happens, the struct is marked as unused and its type has been changed to
678 // TYP_INT (to keep the GC tracking code from looking at it).
679 // See Compiler::raAssignVars() for details. For example:
680 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
681 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
682 // float V03.f1 (offs=0x00) -> V12 tmp7
683 // f8 (last use) (last use) $345
684 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
685 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
687 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
689 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
690 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
691 // this for arguments, which must be passed according the defined ABI.
692 if ((lvType == TYP_SIMD12) && !lvIsParam)
694 assert(lvExactSize == 12);
697 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
699 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
702 unsigned lvSlotNum; // original slot # (if remapped)
704 typeInfo lvVerTypeInfo; // type info needed for verification
706 BYTE* lvGcLayout; // GC layout info for structs
709 BlockSet lvRefBlks; // Set of blocks that contain refs
710 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
711 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
713 var_types TypeGet() const
715 return (var_types)lvType;
717 bool lvStackAligned() const
719 assert(lvIsStructField);
720 return ((lvFldOffset % sizeof(void*)) == 0);
722 bool lvNormalizeOnLoad() const
724 return varTypeIsSmall(TypeGet()) &&
725 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
726 (lvIsParam || lvAddrExposed || lvIsStructField);
729 bool lvNormalizeOnStore()
731 return varTypeIsSmall(TypeGet()) &&
732 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
733 !(lvIsParam || lvAddrExposed || lvIsStructField);
736 void lvaResetSortAgainFlag(Compiler* pComp);
737 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
738 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
739 void setPrefReg(regNumber regNum, Compiler* pComp);
740 void addPrefReg(regMaskTP regMask, Compiler* pComp);
741 bool IsFloatRegType() const
743 return isFloatRegType(lvType) || lvIsHfaRegArg();
745 var_types GetHfaType() const
747 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
749 void SetHfaType(var_types type)
751 assert(varTypeIsFloating(type));
752 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
755 #ifndef LEGACY_BACKEND
756 var_types lvaArgType();
759 PerSsaArray lvPerSsaData;
762 // Keep track of the # of SsaNames, for a bounds check.
763 unsigned lvNumSsaNames;
766 // Returns the address of the per-Ssa data for the given ssaNum (which is required
767 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
768 // not an SSA variable).
769 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
771 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
772 assert(SsaConfig::RESERVED_SSA_NUM == 0);
773 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
774 assert(zeroBased < lvNumSsaNames);
775 return &lvPerSsaData.GetRef(zeroBased);
780 void PrintVarReg() const
782 if (isRegPairType(TypeGet()))
784 printf("%s:%s", getRegName(lvOtherReg), // hi32
785 getRegName(lvRegNum)); // lo32
789 printf("%s", getRegName(lvRegNum));
794 }; // class LclVarDsc
797 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
802 XX The temporary lclVars allocated by the compiler for code generation XX
804 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
808 /*****************************************************************************
810 * The following keeps track of temporaries allocated in the stack frame
811 * during code-generation (after register allocation). These spill-temps are
812 * only used if we run out of registers while evaluating a tree.
814 * These are different from the more common temps allocated by lvaGrabTemp().
825 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
833 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
837 0); // temps must have a negative number (so they have a different number from all local variables)
838 tdOffs = BAD_TEMP_OFFSET;
842 IMPL_LIMITATION("too many spill temps");
847 bool tdLegalOffset() const
849 return tdOffs != BAD_TEMP_OFFSET;
853 int tdTempOffs() const
855 assert(tdLegalOffset());
858 void tdSetTempOffs(int offs)
861 assert(tdLegalOffset());
863 void tdAdjustTempOffs(int offs)
866 assert(tdLegalOffset());
869 int tdTempNum() const
874 unsigned tdTempSize() const
878 var_types tdTempType() const
884 // interface to hide linearscan implementation from rest of compiler
885 class LinearScanInterface
888 virtual void doLinearScan() = 0;
889 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
892 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
894 // Information about arrays: their element type and size, and the offset of the first element.
895 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
896 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
897 // for example, in value numbering of array index expressions.
900 var_types m_elemType;
901 CORINFO_CLASS_HANDLE m_elemStructType;
903 unsigned m_elemOffset;
905 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
909 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
910 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
915 // This enumeration names the phases into which we divide compilation. The phases should completely
916 // partition a compilation.
919 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
920 #include "compphases.h"
924 extern const char* PhaseNames[];
925 extern const char* PhaseEnums[];
926 extern const LPCWSTR PhaseShortNames[];
928 // The following enum provides a simple 1:1 mapping to CLR API's
929 enum API_ICorJitInfo_Names
931 #define DEF_CLR_API(name) API_##name,
932 #include "ICorJitInfo_API_names.h"
936 //---------------------------------------------------------------
940 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
941 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
942 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
943 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
944 // by "m_timerFailure" being true.
945 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
948 #ifdef FEATURE_JIT_METHOD_PERF
949 // The string names of the phases.
950 static const char* PhaseNames[];
952 static bool PhaseHasChildren[];
953 static int PhaseParent[];
955 unsigned m_byteCodeBytes;
956 unsigned __int64 m_totalCycles;
957 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
958 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
959 #if MEASURE_CLRAPI_CALLS
960 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
961 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
963 // For better documentation, we call EndPhase on
964 // non-leaf phases. We should also call EndPhase on the
965 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
966 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
967 // We add all such "redundant end phase" intervals to this variable below; we print
968 // it out in a report, so we can verify that it is, indeed, very small. If it ever
969 // isn't, this means that we're doing something significant between the end of the last
970 // declared subphase and the end of its parent.
971 unsigned __int64 m_parentPhaseEndSlop;
974 #if MEASURE_CLRAPI_CALLS
975 // The following measures the time spent inside each individual CLR API call.
976 unsigned m_allClrAPIcalls;
977 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
978 unsigned __int64 m_allClrAPIcycles;
979 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
980 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
981 #endif // MEASURE_CLRAPI_CALLS
983 CompTimeInfo(unsigned byteCodeBytes);
987 #ifdef FEATURE_JIT_METHOD_PERF
989 #if MEASURE_CLRAPI_CALLS
990 struct WrapICorJitInfo;
993 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
994 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
995 // The operation of adding a single method's timing to the summary may be performed concurrently by several
996 // threads, so it is protected by a lock.
997 // This class is intended to be used as a singleton type, with only a single instance.
998 class CompTimeSummaryInfo
1000 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1001 static CritSecObject s_compTimeSummaryLock;
1005 CompTimeInfo m_total;
1006 CompTimeInfo m_maximum;
1008 int m_numFilteredMethods;
1009 CompTimeInfo m_filtered;
1011 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1012 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1013 // If any OS interaction fails, returns 0.0.
1014 double CyclesPerSecond();
1016 // This can use what ever data you want to determine if the value to be added
1017 // belongs in the filtered section (it's always included in the unfiltered section)
1018 bool IncludedInFilteredData(CompTimeInfo& info);
1021 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1022 static CompTimeSummaryInfo s_compTimeSummary;
1024 CompTimeSummaryInfo()
1025 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1029 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1030 // This is thread safe.
1031 void AddInfo(CompTimeInfo& info, bool includePhases);
1033 // Print the summary information to "f".
1034 // This is not thread-safe; assumed to be called by only one thread.
1035 void Print(FILE* f);
1038 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1039 // and when the current phase started. This is intended to be part of a Compilation object. This is
1040 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1044 unsigned __int64 m_start; // Start of the compilation.
1045 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1046 #if MEASURE_CLRAPI_CALLS
1047 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1048 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1049 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1050 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1051 static double s_cyclesPerSec; // Cached for speedier measurements
1054 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1056 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1058 static CritSecObject s_csvLock; // Lock to protect the time log file.
1059 void PrintCsvMethodStats(Compiler* comp);
1062 void* operator new(size_t);
1063 void* operator new[](size_t);
1064 void operator delete(void*);
1065 void operator delete[](void*);
1068 // Initialized the timer instance
1069 JitTimer(unsigned byteCodeSize);
1071 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1073 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1076 static void PrintCsvHeader();
1078 // Ends the current phase (argument is for a redundant check).
1079 void EndPhase(Phases phase);
1081 #if MEASURE_CLRAPI_CALLS
1082 // Start and end a timed CLR API call.
1083 void CLRApiCallEnter(unsigned apix);
1084 void CLRApiCallLeave(unsigned apix);
1085 #endif // MEASURE_CLRAPI_CALLS
1087 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1088 // and adds it to "sum".
1089 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1091 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1092 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1093 // "m_info" to true.
1094 bool GetThreadCycles(unsigned __int64* cycles)
1096 bool res = CycleTimer::GetThreadCyclesS(cycles);
1099 m_info.m_timerFailure = true;
1104 #endif // FEATURE_JIT_METHOD_PERF
1106 //------------------- Function/Funclet info -------------------------------
1107 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1109 FUNC_ROOT, // The main/root function (always id==0)
1110 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1111 FUNC_FILTER, // a funclet associated with an EH filter
1114 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1121 BYTE funFlags; // Currently unused, just here for padding
1122 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1123 // funclet. It is only valid if funKind field indicates this is a
1124 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1126 #if defined(_TARGET_AMD64_)
1128 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1129 emitLocation* startLoc;
1130 emitLocation* endLoc;
1131 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1132 emitLocation* coldEndLoc;
1133 UNWIND_INFO unwindHeader;
1134 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1135 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1136 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1137 unsigned unwindCodeSlot;
1139 #ifdef UNIX_AMD64_ABI
1140 jitstd::vector<CFI_CODE>* cfiCodes;
1141 #endif // UNIX_AMD64_ABI
1143 #elif defined(_TARGET_ARMARCH_)
1145 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1146 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1147 // Note: we only have a pointer here instead of the actual object,
1148 // to save memory in the JIT case (compared to the NGEN case),
1149 // where we don't have any cold section.
1150 // Note 2: we currently don't support hot/cold splitting in functions
1151 // with EH, so uwiCold will be NULL for all funclets.
1153 #endif // _TARGET_ARMARCH_
1155 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1156 // that isn't shared between the main function body and funclets.
1159 struct fgArgTabEntry
1162 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1165 otherRegNum = REG_NA;
1166 isStruct = false; // is this a struct arg
1168 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1170 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1172 // it will point at the actual argument in the gtCallLateArgs list.
1173 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1175 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1177 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1179 unsigned numRegs; // Count of number of registers that this argument uses
1181 // A slot is a pointer sized region in the OutArg area.
1182 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1183 unsigned numSlots; // Count of number of slots that this argument uses
1185 unsigned alignment; // 1 or 2 (slots/registers)
1186 unsigned lateArgInx; // index into gtCallLateArgs list
1187 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1189 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1190 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1191 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1192 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1193 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1194 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1195 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1196 // previous arguments.
1197 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1198 // to be on the stack despite its arg list position.
1200 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1201 bool isStruct : 1; // True if this is a struct arg
1203 regNumber otherRegNum; // The (second) register to use when passing this argument.
1205 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1206 #elif defined(_TARGET_X86_)
1207 __declspec(property(get = getIsStruct)) bool isStruct;
1210 return varTypeIsStruct(node);
1212 #endif // _TARGET_X86_
1215 void SetIsHfaRegArg(bool hfaRegArg)
1217 isHfaRegArg = hfaRegArg;
1220 void SetIsBackFilled(bool backFilled)
1222 isBackFilled = backFilled;
1225 bool IsBackFilled() const
1227 return isBackFilled;
1229 #else // !_TARGET_ARM_
1230 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1232 void SetIsHfaRegArg(bool hfaRegArg)
1236 void SetIsBackFilled(bool backFilled)
1240 bool IsBackFilled() const
1244 #endif // !_TARGET_ARM_
1250 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1252 //-------------------------------------------------------------------------
1254 // The class fgArgInfo is used to handle the arguments
1255 // when morphing a GT_CALL node.
1260 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1261 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1262 unsigned argCount; // Updatable arg count value
1263 unsigned nextSlotNum; // Updatable slot count value
1264 unsigned stkLevel; // Stack depth when we make this call (for x86)
1266 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1267 bool hasRegArgs; // true if we have one or more register arguments
1268 bool hasStackArgs; // true if we have one or more stack arguments
1269 bool argsComplete; // marker for state
1270 bool argsSorted; // marker for state
1271 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1274 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1277 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1278 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1280 fgArgTabEntryPtr AddRegArg(
1281 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1283 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1284 fgArgTabEntryPtr AddRegArg(
1291 const bool isStruct,
1292 const regNumber otherRegNum = REG_NA,
1293 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1294 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1296 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1300 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1302 void RemorphReset();
1303 fgArgTabEntryPtr RemorphRegArg(
1304 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1306 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1308 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1310 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1312 void ArgsComplete();
1316 void EvalArgsToTemps();
1318 void RecordStkLevel(unsigned stkLvl);
1319 unsigned RetrieveStkLevel();
1325 fgArgTabEntryPtr* ArgTable()
1329 unsigned GetNextSlotNum()
1339 return hasStackArgs;
1341 bool AreArgsComplete() const
1343 return argsComplete;
1348 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1349 // We have the ability to mark source expressions with "Test Labels."
1350 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1351 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1353 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1356 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1357 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1358 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1359 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1360 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1363 struct TestLabelAndNum
1368 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1373 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1375 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1378 // This class implements the "IAllocator" interface, so that we can use
1379 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1381 class CompAllocator : public IAllocator
1384 #if MEASURE_MEM_ALLOC
1388 CompAllocator(Compiler* comp, CompMemKind cmk)
1390 #if MEASURE_MEM_ALLOC
1396 inline void* Alloc(size_t sz);
1398 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1400 // For the compiler's no-release allocator, free operations are no-ops.
1407 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1408 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1410 XX The big guy. The sections are currently organized as : XX
1412 XX o GenTree and BasicBlock XX
1424 XX o PrologScopeInfo XX
1425 XX o CodeGenerator XX
1430 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1431 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1436 friend class emitter;
1437 friend class UnwindInfo;
1438 friend class UnwindFragmentInfo;
1439 friend class UnwindEpilogInfo;
1440 friend class JitTimer;
1441 friend class LinearScan;
1442 friend class fgArgInfo;
1443 friend class Rationalizer;
1445 friend class Lowering;
1446 friend class CSE_DataFlow;
1447 friend class CSE_Heuristic;
1448 friend class CodeGenInterface;
1449 friend class CodeGen;
1450 friend class LclVarDsc;
1451 friend class TempDsc;
1453 friend class ObjectAllocator;
1455 #ifndef _TARGET_64BIT_
1456 friend class DecomposeLongs;
1457 #endif // !_TARGET_64BIT_
1460 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1461 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1463 XX Misc structs definitions XX
1465 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1466 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1470 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1489 bool dumpIRDataflow;
1490 bool dumpIRBlockHeaders;
1492 LPCWSTR dumpIRPhase;
1493 LPCWSTR dumpIRFormat;
1495 bool shouldUseVerboseTrees();
1496 bool asciiTrees; // If true, dump trees using only ASCII characters
1497 bool shouldDumpASCIITrees();
1498 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1499 bool shouldUseVerboseSsa();
1500 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1501 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1503 const char* VarNameToStr(VarName name)
1508 DWORD expensiveDebugCheckLevel;
1511 #if FEATURE_MULTIREG_RET
1512 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1513 #endif // FEATURE_MULTIREG_RET
1516 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1517 #endif // ARM_SOFTFP
1519 //-------------------------------------------------------------------------
1520 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1521 // HFAs are one to four element structs where each element is the same
1522 // type, either all float or all double. They are treated specially
1523 // in the ARM Procedure Call Standard, specifically, they are passed in
1524 // floating-point registers instead of the general purpose registers.
1527 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1528 bool IsHfa(GenTreePtr tree);
1530 var_types GetHfaType(GenTreePtr tree);
1531 unsigned GetHfaCount(GenTreePtr tree);
1533 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1534 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1536 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1537 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1539 //-------------------------------------------------------------------------
1540 // The following is used for validating format of EH table
1544 typedef struct EHNodeDsc* pEHNodeDsc;
1546 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1547 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1560 EHBlockType ehnBlockType; // kind of EH block
1561 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1562 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1563 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1565 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1566 pEHNodeDsc ehnChild; // leftmost nested block
1568 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1569 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1571 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1572 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1574 inline void ehnSetTryNodeType()
1576 ehnBlockType = TryNode;
1578 inline void ehnSetFilterNodeType()
1580 ehnBlockType = FilterNode;
1582 inline void ehnSetHandlerNodeType()
1584 ehnBlockType = HandlerNode;
1586 inline void ehnSetFinallyNodeType()
1588 ehnBlockType = FinallyNode;
1590 inline void ehnSetFaultNodeType()
1592 ehnBlockType = FaultNode;
1595 inline BOOL ehnIsTryBlock()
1597 return ehnBlockType == TryNode;
1599 inline BOOL ehnIsFilterBlock()
1601 return ehnBlockType == FilterNode;
1603 inline BOOL ehnIsHandlerBlock()
1605 return ehnBlockType == HandlerNode;
1607 inline BOOL ehnIsFinallyBlock()
1609 return ehnBlockType == FinallyNode;
1611 inline BOOL ehnIsFaultBlock()
1613 return ehnBlockType == FaultNode;
1616 // returns true if there is any overlap between the two nodes
1617 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1619 if (node1->ehnStartOffset < node2->ehnStartOffset)
1621 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1625 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1629 // fails with BADCODE if inner is not completely nested inside outer
1630 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1632 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1636 //-------------------------------------------------------------------------
1637 // Exception handling functions
1640 #if !FEATURE_EH_FUNCLETS
1642 bool ehNeedsShadowSPslots()
1644 return (info.compXcptnsCount || opts.compDbgEnC);
1647 // 0 for methods with no EH
1648 // 1 for methods with non-nested EH, or where only the try blocks are nested
1649 // 2 for a method with a catch within a catch
1651 unsigned ehMaxHndNestingCount;
1653 #endif // !FEATURE_EH_FUNCLETS
1655 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1656 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1658 bool bbInCatchHandlerILRange(BasicBlock* blk);
1659 bool bbInFilterILRange(BasicBlock* blk);
1660 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1661 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1662 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1663 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1664 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1666 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1667 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1669 // Returns true if "block" is the start of a try region.
1670 bool bbIsTryBeg(BasicBlock* block);
1672 // Returns true if "block" is the start of a handler or filter region.
1673 bool bbIsHandlerBeg(BasicBlock* block);
1675 // Returns true iff "block" is where control flows if an exception is raised in the
1676 // try region, and sets "*regionIndex" to the index of the try for the handler.
1677 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1678 // block of the filter, but not for the filter's handler.
1679 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1681 bool ehHasCallableHandlers();
1683 // Return the EH descriptor for the given region index.
1684 EHblkDsc* ehGetDsc(unsigned regionIndex);
1686 // Return the EH index given a region descriptor.
1687 unsigned ehGetIndex(EHblkDsc* ehDsc);
1689 // Return the EH descriptor index of the enclosing try, for the given region index.
1690 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1692 // Return the EH descriptor index of the enclosing handler, for the given region index.
1693 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1695 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1696 // block is not in a 'try' region).
1697 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1699 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1700 // if this block is not in a filter or handler region).
1701 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1703 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1704 // nullptr if this block's exceptions propagate to caller).
1705 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1707 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1708 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1709 bool ehIsBlockEHLast(BasicBlock* block);
1711 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1713 // Return the region index of the most nested EH region this block is in.
1714 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1716 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1717 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1719 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1720 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1721 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1722 // (It can never be a filter.)
1723 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1725 // A block has been deleted. Update the EH table appropriately.
1726 void ehUpdateForDeletedBlock(BasicBlock* block);
1728 // Determine whether a block can be deleted while preserving the EH normalization rules.
1729 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1731 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1732 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1734 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1735 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1736 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1737 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1738 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1739 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1740 // lives in a filter.)
1741 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1743 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1744 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1745 // (nullptr if the last block is the last block in the program).
1746 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1747 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1750 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1751 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1752 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1755 #if FEATURE_EH_FUNCLETS
1756 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1757 // if there is a filter that protects a region with a nested EH clause (such as a
1758 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1759 // genFuncletProlog() for more details. However, the VM seems to use it for more
1760 // purposes, maybe including debugging. Until we are sure otherwise, always create
1761 // a PSPSym for functions with any EH.
1762 bool ehNeedsPSPSym() const
1766 #else // _TARGET_X86_
1767 return compHndBBtabCount > 0;
1768 #endif // _TARGET_X86_
1771 bool ehAnyFunclets(); // Are there any funclets in this function?
1772 unsigned ehFuncletCount(); // Return the count of funclets in the function
1774 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1775 #else // !FEATURE_EH_FUNCLETS
1776 bool ehAnyFunclets()
1780 unsigned ehFuncletCount()
1785 unsigned bbThrowIndex(BasicBlock* blk)
1787 return blk->bbTryIndex;
1788 } // Get the index to use as the cache key for sharing throw blocks
1789 #endif // !FEATURE_EH_FUNCLETS
1791 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1792 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1793 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1794 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1795 // convenient to also consider it a predecessor.)
1796 flowList* BlockPredsWithEH(BasicBlock* blk);
1798 // This table is useful for memoization of the method above.
1799 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1801 BlockToFlowListMap* m_blockToEHPreds;
1802 BlockToFlowListMap* GetBlockToEHPreds()
1804 if (m_blockToEHPreds == nullptr)
1806 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1808 return m_blockToEHPreds;
1811 void* ehEmitCookie(BasicBlock* block);
1812 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1814 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1816 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1818 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1820 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1822 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1824 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1826 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1828 void fgAllocEHTable();
1830 void fgRemoveEHTableEntry(unsigned XTnum);
1832 #if FEATURE_EH_FUNCLETS
1834 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1836 #endif // FEATURE_EH_FUNCLETS
1840 #endif // !FEATURE_EH
1842 void fgSortEHTable();
1844 // Causes the EH table to obey some well-formedness conditions, by inserting
1845 // empty BB's when necessary:
1846 // * No block is both the first block of a handler and the first block of a try.
1847 // * No block is the first block of multiple 'try' regions.
1848 // * No block is the last block of multiple EH regions.
1849 void fgNormalizeEH();
1850 bool fgNormalizeEHCase1();
1851 bool fgNormalizeEHCase2();
1852 bool fgNormalizeEHCase3();
1855 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1856 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1857 void fgVerifyHandlerTab();
1858 void fgDispHandlerTab();
1861 bool fgNeedToSortEHTable;
1863 void verInitEHTree(unsigned numEHClauses);
1864 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1865 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1866 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1867 void verCheckNestingLevel(EHNodeDsc* initRoot);
1870 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1871 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1873 XX GenTree and BasicBlock XX
1875 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1877 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1878 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1881 // Functions to create nodes
1882 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1885 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1887 // For binary opers.
1888 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1890 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1892 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1893 var_types type = TYP_I_IMPL,
1894 GenTreePtr op1 = nullptr,
1895 GenTreePtr op2 = nullptr);
1897 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1899 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1901 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1903 GenTreePtr gtNewJmpTableNode();
1904 GenTreePtr gtNewIconHandleNode(
1905 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1907 unsigned gtTokenToIconFlags(unsigned token);
1909 GenTreePtr gtNewIconEmbHndNode(void* value,
1912 unsigned handle1 = 0,
1913 void* handle2 = nullptr,
1914 void* compileTimeHandle = nullptr);
1916 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1917 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1918 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1919 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1921 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1923 GenTreePtr gtNewLconNode(__int64 value);
1925 GenTreePtr gtNewDconNode(double value);
1927 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1929 GenTreePtr gtNewZeroConNode(var_types type);
1931 GenTreePtr gtNewOneConNode(var_types type);
1934 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1935 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1938 GenTreeBlk* gtNewBlkOpNode(
1939 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1941 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1944 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1947 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1948 void gtSetObjGcInfo(GenTreeObj* objNode);
1949 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1950 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1952 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1954 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1956 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1957 CORINFO_METHOD_HANDLE handle,
1959 GenTreeArgList* args,
1960 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1962 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1964 GenTreeArgList* args,
1965 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1967 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1970 GenTreeArgList* args = nullptr);
1972 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1975 GenTreeSIMD* gtNewSIMDNode(
1976 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1977 GenTreeSIMD* gtNewSIMDNode(var_types type,
1980 SIMDIntrinsicID simdIntrinsicID,
1985 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1986 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
1987 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
1989 GenTreePtr gtNewCodeRef(BasicBlock* block);
1991 GenTreePtr gtNewFieldRef(
1992 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
1994 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
1996 GenTreeArgList* gtNewArgList(GenTreePtr op);
1997 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
1998 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2000 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2001 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2002 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2003 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2005 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2007 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2009 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2010 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2011 CORINFO_ACCESS_FLAGS access,
2012 CORINFO_FIELD_INFO* pFieldInfo,
2014 CORINFO_CLASS_HANDLE structType,
2017 GenTreePtr gtNewNothingNode();
2019 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2021 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2023 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2025 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2027 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2029 //------------------------------------------------------------------------
2030 // Other GenTree functions
2032 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2034 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2035 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2036 // IntCnses with value `deepVarVal`.
2037 GenTreePtr gtCloneExpr(
2038 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2040 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2041 // `varNum` to int constants with value `varVal`.
2042 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2044 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2047 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2049 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2051 // Returns "true" iff the complexity (not formally defined, but first interpretation
2052 // is #of nodes in subtree) of "tree" is greater than "limit".
2053 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2054 // before they have been set.)
2055 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2057 bool gtCompareTree(GenTree* op1, GenTree* op2);
2059 GenTreePtr gtReverseCond(GenTree* tree);
2061 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2063 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2065 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2067 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2070 unsigned gtHashValue(GenTree* tree);
2072 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2075 void gtPrepareCost(GenTree* tree);
2076 bool gtIsLikelyRegVar(GenTree* tree);
2078 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2080 // Returns true iff the secondNode can be swapped with firstNode.
2081 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2083 unsigned gtSetEvalOrder(GenTree* tree);
2085 #if FEATURE_STACK_FP_X87
2087 void gtComputeFPlvls(GenTreePtr tree);
2088 #endif // FEATURE_STACK_FP_X87
2090 void gtSetStmtInfo(GenTree* stmt);
2092 // Returns "true" iff "node" has any of the side effects in "flags".
2093 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2095 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2096 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2098 // Appends 'expr' in front of 'list'
2099 // 'list' will typically start off as 'nullptr'
2100 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2101 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2103 void gtExtractSideEffList(GenTreePtr expr,
2105 unsigned flags = GTF_SIDE_EFFECT,
2106 bool ignoreRoot = false);
2108 GenTreePtr gtGetThisArg(GenTreePtr call);
2110 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2111 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2112 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2113 // the given "fldHnd", is such an object pointer.
2114 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2116 // Return true if call is a recursive call; return false otherwise.
2117 bool gtIsRecursiveCall(GenTreeCall* call)
2119 return (call->gtCallMethHnd == info.compMethodHnd);
2122 //-------------------------------------------------------------------------
2124 GenTreePtr gtFoldExpr(GenTreePtr tree);
2127 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2128 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2129 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2130 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2131 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2132 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2133 // optimizations for now.
2134 __attribute__((optnone))
2136 gtFoldExprConst(GenTreePtr tree);
2137 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2138 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2140 //-------------------------------------------------------------------------
2141 // Get the handle, if any.
2142 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2143 // Get the handle, and assert if not found.
2144 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2146 //-------------------------------------------------------------------------
2147 // Functions to display the trees
2150 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2152 void gtDispVN(GenTreePtr tree);
2153 void gtDispConst(GenTreePtr tree);
2154 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2155 void gtDispNodeName(GenTreePtr tree);
2156 void gtDispRegVal(GenTreePtr tree);
2168 void gtDispChild(GenTreePtr child,
2169 IndentStack* indentStack,
2171 __in_opt const char* msg = nullptr,
2172 bool topOnly = false);
2173 void gtDispTree(GenTreePtr tree,
2174 IndentStack* indentStack = nullptr,
2175 __in_opt const char* msg = nullptr,
2176 bool topOnly = false,
2177 bool isLIR = false);
2178 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2179 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2180 char* gtGetLclVarName(unsigned lclNum);
2181 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2182 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2183 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2184 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2185 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2186 void gtDispFieldSeq(FieldSeqNode* pfsn);
2188 void gtDispRange(LIR::ReadOnlyRange const& range);
2190 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2192 void gtDispLIRNode(GenTree* node);
2204 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2205 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2208 static fgWalkPreFn gtAssertColonCond;
2210 static fgWalkPreFn gtMarkColonCond;
2211 static fgWalkPreFn gtClearColonCond;
2213 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2214 bool gtHasCatchArg(GenTreePtr tree);
2215 bool gtHasUnmanagedCall(GenTreePtr tree);
2217 typedef ArrayStack<GenTree*> GenTreeStack;
2219 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2220 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2222 //=========================================================================
2223 // BasicBlock functions
2225 // This is a debug flag we will use to assert when creating block during codegen
2226 // as this interferes with procedure splitting. If you know what you're doing, set
2227 // it to true before creating the block. (DEBUG only)
2228 bool fgSafeBasicBlockCreation;
2231 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2234 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2235 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2239 XX The variables to be used by the code generator. XX
2241 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2242 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2246 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2247 // be placed in the stack frame and it's fields must be laid out sequentially.
2249 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2250 // a local variable that can be enregistered or placed in the stack frame.
2251 // The fields do not need to be laid out sequentially
2253 enum lvaPromotionType
2255 PROMOTION_TYPE_NONE, // The struct local is not promoted
2256 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2257 // and its field locals are independent of its parent struct local.
2258 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2259 // but its field locals depend on its parent struct local.
2262 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2263 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2265 /*****************************************************************************/
2267 enum FrameLayoutState
2270 INITIAL_FRAME_LAYOUT,
2271 PRE_REGALLOC_FRAME_LAYOUT,
2272 REGALLOC_FRAME_LAYOUT,
2273 TENTATIVE_FRAME_LAYOUT,
2278 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2279 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2280 bool lvaSortAgain; // true: We need to sort the lvaTable
2281 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2282 unsigned lvaCount; // total number of locals
2284 unsigned lvaRefCount; // total number of references to locals
2285 LclVarDsc* lvaTable; // variable descriptor table
2286 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2288 LclVarDsc** lvaRefSorted; // table sorted by refcount
2290 unsigned short lvaTrackedCount; // actual # of locals being tracked
2291 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2293 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2294 // Only for AMD64 System V cache the first caller stack homed argument.
2295 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2296 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2299 VARSET_TP lvaTrackedVars; // set of tracked variables
2301 #ifndef _TARGET_64BIT_
2302 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2304 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2306 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2307 // It that changes, this changes. VarSets from different epochs
2308 // cannot be meaningfully combined.
2310 unsigned GetCurLVEpoch()
2315 // reverse map of tracked number to var number
2316 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2318 #ifdef LEGACY_BACKEND
2319 // variable interference graph
2320 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2323 // variable preference graph
2324 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2328 // # of procs compiled a with double-aligned stack
2329 static unsigned s_lvaDoubleAlignedProcsCount;
2333 // Getters and setters for address-exposed and do-not-enregister local var properties.
2334 bool lvaVarAddrExposed(unsigned varNum);
2335 void lvaSetVarAddrExposed(unsigned varNum);
2336 bool lvaVarDoNotEnregister(unsigned varNum);
2338 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2339 enum DoNotEnregisterReason
2344 DNER_VMNeedsStackAddr,
2345 DNER_LiveInOutOfHandler,
2346 DNER_LiveAcrossUnmanagedCall,
2347 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2348 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2349 #ifdef JIT32_GCENCODER
2354 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2356 unsigned lvaVarargsHandleArg;
2358 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2360 #endif // _TARGET_X86_
2362 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2363 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2364 #if FEATURE_FIXED_OUT_ARGS
2365 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2367 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2368 // that tracks whether the lock has been taken
2370 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2371 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2372 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2374 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2375 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2377 #if FEATURE_FIXED_OUT_ARGS
2378 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2379 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2380 #endif // FEATURE_FIXED_OUT_ARGS
2383 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2384 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2385 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2386 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2387 // this variable to be this scratch word whenever struct promotion occurs.
2388 unsigned lvaPromotedStructAssemblyScratchVar;
2389 #endif // _TARGET_ARM_
2392 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2393 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2396 bool lvaGenericsContextUsed;
2398 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2399 // CORINFO_GENERICS_CTXT_FROM_THIS?
2400 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2402 //-------------------------------------------------------------------------
2403 // All these frame offsets are inter-related and must be kept in sync
2405 #if !FEATURE_EH_FUNCLETS
2406 // This is used for the callable handlers
2407 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2408 #endif // FEATURE_EH_FUNCLETS
2410 unsigned lvaCachedGenericContextArgOffs;
2411 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2414 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2416 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2418 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2419 // after the reg predict we will use a computed maxTmpSize
2420 // which is based upon the number of spill temps predicted by reg predict
2421 // All this is necessary because if we under-estimate the size of the spill
2422 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2424 // Pre codegen max spill temp size.
2425 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2427 //-------------------------------------------------------------------------
2429 unsigned lvaGetMaxSpillTempSize();
2431 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2432 #endif // _TARGET_ARM_
2433 void lvaAssignFrameOffsets(FrameLayoutState curState);
2434 void lvaFixVirtualFrameOffsets();
2436 #ifndef LEGACY_BACKEND
2437 void lvaUpdateArgsWithInitialReg();
2438 #endif // !LEGACY_BACKEND
2440 void lvaAssignVirtualFrameOffsetsToArgs();
2441 #ifdef UNIX_AMD64_ABI
2442 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2443 #else // !UNIX_AMD64_ABI
2444 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2445 #endif // !UNIX_AMD64_ABI
2446 void lvaAssignVirtualFrameOffsetsToLocals();
2447 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2448 #ifdef _TARGET_AMD64_
2449 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2450 bool lvaIsCalleeSavedIntRegCountEven();
2452 void lvaAlignFrame();
2453 void lvaAssignFrameOffsetsToPromotedStructs();
2454 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2457 void lvaDumpRegLocation(unsigned lclNum);
2458 void lvaDumpFrameLocation(unsigned lclNum);
2459 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2460 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2461 // layout state defined by lvaDoneFrameLayout
2464 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2465 // to avoid bugs from borderline cases.
2466 #define MAX_FrameSize 0x3FFFFFFF
2467 void lvaIncrementFrameSize(unsigned size);
2469 unsigned lvaFrameSize(FrameLayoutState curState);
2471 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2472 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2474 // Returns the caller-SP-relative offset for the local variable "varNum."
2475 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2477 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2478 int lvaGetSPRelativeOffset(unsigned varNum);
2480 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2481 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2483 //------------------------ For splitting types ----------------------------
2485 void lvaInitTypeRef();
2487 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2488 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2489 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2490 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2491 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2492 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2494 void lvaInitVarDsc(LclVarDsc* varDsc,
2496 CorInfoType corInfoType,
2497 CORINFO_CLASS_HANDLE typeHnd,
2498 CORINFO_ARG_LIST_HANDLE varList,
2499 CORINFO_SIG_INFO* varSig);
2501 static unsigned lvaTypeRefMask(var_types type);
2503 var_types lvaGetActualType(unsigned lclNum);
2504 var_types lvaGetRealType(unsigned lclNum);
2506 //-------------------------------------------------------------------------
2510 unsigned lvaLclSize(unsigned varNum);
2511 unsigned lvaLclExactSize(unsigned varNum);
2513 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2515 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2516 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2517 // the return result.
2518 bool lvaLclVarRefsAccum(
2519 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2521 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2522 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2523 // and (destructively) unions "trkedVars" into "*result".
2524 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2526 ALLVARSET_VALARG_TP allVars,
2527 VARSET_VALARG_TP trkdVars);
2529 bool lvaHaveManyLocals() const;
2531 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2532 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2533 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2536 void lvaSortByRefCount();
2537 void lvaDumpRefCounts();
2539 void lvaMarkLocalVars(BasicBlock* block);
2541 void lvaMarkLocalVars(); // Local variable ref-counting
2543 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2545 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2547 static fgWalkPreFn lvaIncRefCntsCB;
2548 void lvaIncRefCnts(GenTreePtr tree);
2550 static fgWalkPreFn lvaDecRefCntsCB;
2551 void lvaDecRefCnts(GenTreePtr tree);
2552 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2553 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2554 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2557 struct lvaStressLclFldArgs
2559 Compiler* m_pCompiler;
2563 static fgWalkPreFn lvaStressLclFldCB;
2564 void lvaStressLclFld();
2566 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2567 void lvaDispVarSet(VARSET_VALARG_TP set);
2572 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2574 int lvaFrameAddress(int varNum, bool* pFPbased);
2577 bool lvaIsParameter(unsigned varNum);
2578 bool lvaIsRegArgument(unsigned varNum);
2579 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2580 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2581 // that writes to arg0
2583 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2584 // (this is an overload of lvIsTemp because there are no temp parameters).
2585 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2586 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2587 bool lvaIsImplicitByRefLocal(unsigned varNum)
2589 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2590 LclVarDsc* varDsc = &(lvaTable[varNum]);
2591 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2593 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2596 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2600 // Returns true if this local var is a multireg struct
2601 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2603 // If the class is a TYP_STRUCT, get/set a class handle describing it
2605 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2606 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2608 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2610 // Info about struct fields
2611 struct lvaStructFieldInfo
2613 CORINFO_FIELD_HANDLE fldHnd;
2614 unsigned char fldOffset;
2615 unsigned char fldOrdinal;
2618 CORINFO_CLASS_HANDLE fldTypeHnd;
2621 // Info about struct to be promoted.
2622 struct lvaStructPromotionInfo
2624 CORINFO_CLASS_HANDLE typeHnd;
2626 bool requiresScratchVar;
2629 unsigned char fieldCnt;
2630 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2632 lvaStructPromotionInfo()
2633 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2638 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2639 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2640 lvaStructPromotionInfo* StructPromotionInfo,
2642 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2643 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2644 #if !defined(_TARGET_64BIT_)
2645 void lvaPromoteLongVars();
2646 #endif // !defined(_TARGET_64BIT_)
2647 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2648 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2649 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2650 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2651 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2652 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2653 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2655 BYTE* lvaGetGcLayout(unsigned varNum);
2656 bool lvaTypeIsGC(unsigned varNum);
2657 unsigned lvaGSSecurityCookie; // LclVar number
2658 bool lvaTempsHaveLargerOffsetThanVars();
2660 unsigned lvaSecurityObject; // variable representing the security object on the stack
2661 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2663 #if FEATURE_EH_FUNCLETS
2664 unsigned lvaPSPSym; // variable representing the PSPSym
2667 InlineInfo* impInlineInfo;
2668 InlineStrategy* m_inlineStrategy;
2670 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2671 Compiler* impInlineRoot();
2673 #if defined(DEBUG) || defined(INLINE_DATA)
2674 unsigned __int64 getInlineCycleCount()
2676 return m_compCycles;
2678 #endif // defined(DEBUG) || defined(INLINE_DATA)
2680 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2681 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2683 //=========================================================================
2685 //=========================================================================
2688 //---------------- Local variable ref-counting ----------------------------
2691 BasicBlock* lvaMarkRefsCurBlock;
2692 GenTreePtr lvaMarkRefsCurStmt;
2694 BasicBlock::weight_t lvaMarkRefsWeight;
2696 static fgWalkPreFn lvaMarkLclRefsCallback;
2697 void lvaMarkLclRefs(GenTreePtr tree);
2699 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2700 PerSsaArray lvHeapPerSsaData;
2701 unsigned lvHeapNumSsaNames;
2704 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2705 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2706 // not an SSA variable).
2707 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2709 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2710 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2712 assert(ssaNum < lvHeapNumSsaNames);
2713 return &lvHeapPerSsaData.GetRef(ssaNum);
2717 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2718 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2722 XX Imports the given method and converts it to semantic trees XX
2724 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2725 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2731 void impImport(BasicBlock* method);
2733 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2734 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2735 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2736 CORINFO_CLASS_HANDLE impGetStringClass();
2737 CORINFO_CLASS_HANDLE impGetObjectClass();
2739 //=========================================================================
2741 //=========================================================================
2744 //-------------------- Stack manipulation ---------------------------------
2746 unsigned impStkSize; // Size of the full stack
2748 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2750 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2752 struct SavedStack // used to save/restore stack contents.
2754 unsigned ssDepth; // number of values on stack
2755 StackEntry* ssTrees; // saved tree values
2758 bool impIsPrimitive(CorInfoType type);
2759 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2761 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2762 void impPushOnStackNoType(GenTreePtr tree);
2764 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2765 void impPushNullObjRefOnStack();
2766 StackEntry impPopStack();
2767 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2768 GenTreePtr impPopStack(typeInfo& ti);
2769 StackEntry& impStackTop(unsigned n = 0);
2771 void impSaveStackState(SavedStack* savePtr, bool copy);
2772 void impRestoreStackState(SavedStack* savePtr);
2774 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2775 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2776 CORINFO_CALL_INFO* pCallInfo);
2778 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2780 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2782 bool impCanPInvokeInline();
2783 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2784 void impCheckForPInvokeCall(
2785 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2786 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2787 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2789 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2790 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2791 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2793 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2794 CORINFO_METHOD_HANDLE calleeMethodHnd,
2795 CORINFO_CLASS_HANDLE delegateTypeHnd);
2797 var_types impImportCall(OPCODE opcode,
2798 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2799 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2801 GenTreePtr newobjThis,
2803 CORINFO_CALL_INFO* callInfo,
2804 IL_OFFSET rawILOffset);
2806 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2808 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2810 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2813 var_types impImportJitTestLabelMark(int numArgs);
2816 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2818 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2820 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2821 CORINFO_ACCESS_FLAGS access,
2822 CORINFO_FIELD_INFO* pFieldInfo,
2825 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2827 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2829 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2831 void impImportLeave(BasicBlock* block);
2832 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2833 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2834 CORINFO_CLASS_HANDLE clsHnd,
2835 CORINFO_METHOD_HANDLE method,
2836 CORINFO_SIG_INFO* sig,
2840 CorInfoIntrinsics* pIntrinsicID);
2841 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2842 CORINFO_SIG_INFO* sig,
2845 CorInfoIntrinsics intrinsicID);
2846 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2848 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2850 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2851 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2852 CORINFO_THIS_TRANSFORM transform);
2854 //----------------- Manipulating the trees and stmts ----------------------
2856 GenTreePtr impTreeList; // Trees for the BB being imported
2857 GenTreePtr impTreeLast; // The last tree for the current BB
2861 CHECK_SPILL_ALL = -1,
2862 CHECK_SPILL_NONE = -2
2866 void impBeginTreeList();
2867 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2868 void impEndTreeList(BasicBlock* block);
2869 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2870 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2871 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2872 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2873 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2874 void impAssignTempGen(unsigned tmp,
2877 GenTreePtr* pAfterStmt = nullptr,
2878 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2879 BasicBlock* block = nullptr);
2880 void impAssignTempGen(unsigned tmpNum,
2882 CORINFO_CLASS_HANDLE structHnd,
2884 GenTreePtr* pAfterStmt = nullptr,
2885 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2886 BasicBlock* block = nullptr);
2887 GenTreePtr impCloneExpr(GenTreePtr tree,
2889 CORINFO_CLASS_HANDLE structHnd,
2891 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2892 GenTreePtr impAssignStruct(GenTreePtr dest,
2894 CORINFO_CLASS_HANDLE structHnd,
2896 GenTreePtr* pAfterStmt = nullptr,
2897 BasicBlock* block = nullptr);
2898 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2900 CORINFO_CLASS_HANDLE structHnd,
2902 GenTreePtr* pAfterStmt = nullptr,
2903 BasicBlock* block = nullptr);
2905 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2906 CORINFO_CLASS_HANDLE structHnd,
2910 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2911 BYTE* gcLayout = nullptr,
2912 unsigned* numGCVars = nullptr,
2913 var_types* simdBaseType = nullptr);
2915 GenTreePtr impNormStructVal(GenTreePtr structVal,
2916 CORINFO_CLASS_HANDLE structHnd,
2918 bool forceNormalization = false);
2920 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2921 BOOL* pRuntimeLookup = nullptr,
2922 BOOL mustRestoreHandle = FALSE,
2923 BOOL importParent = FALSE);
2925 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2926 BOOL* pRuntimeLookup = nullptr,
2927 BOOL mustRestoreHandle = FALSE)
2929 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2932 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2933 CORINFO_LOOKUP* pLookup,
2935 void* compileTimeHandle);
2937 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2939 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2940 CORINFO_LOOKUP* pLookup,
2941 void* compileTimeHandle);
2943 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2945 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2946 CorInfoHelpFunc helper,
2948 GenTreeArgList* arg = nullptr,
2949 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
2951 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
2953 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2956 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
2957 CORINFO_CLASS_HANDLE typeClass,
2961 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
2962 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
2963 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
2964 static bool IsMathIntrinsic(GenTreePtr tree);
2967 //----------------- Importing the method ----------------------------------
2969 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
2972 unsigned impCurOpcOffs;
2973 const char* impCurOpcName;
2974 bool impNestedStackSpill;
2976 // For displaying instrs with generated native code (-n:B)
2977 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
2978 void impNoteLastILoffs();
2981 /* IL offset of the stmt currently being imported. It gets set to
2982 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
2983 updated at IL offsets for which we have to report mapping info.
2984 It also includes flag bits, so use jitGetILoffs()
2985 to get the actual IL offset value.
2988 IL_OFFSETX impCurStmtOffs;
2989 void impCurStmtOffsSet(IL_OFFSET offs);
2991 void impNoteBranchOffs();
2993 unsigned impInitBlockLineInfo();
2995 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
2996 bool impIsThis(GenTreePtr obj);
2997 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
2998 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
2999 bool impIsAnySTLOC(OPCODE opcode)
3001 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3002 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3005 GenTreeArgList* impPopList(unsigned count,
3007 CORINFO_SIG_INFO* sig,
3008 GenTreeArgList* prefixTree = nullptr);
3010 GenTreeArgList* impPopRevList(unsigned count,
3012 CORINFO_SIG_INFO* sig,
3013 unsigned skipReverseCount = 0);
3016 * Get current IL offset with stack-empty info incoporated
3018 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3020 //---------------- Spilling the importer stack ----------------------------
3026 SavedStack pdSavedStack;
3027 ThisInitState pdThisPtrInit;
3030 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3031 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3033 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3034 ExpandArray<BYTE> impPendingBlockMembers;
3036 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3037 // Operates on the map in the top-level ancestor.
3038 BYTE impGetPendingBlockMember(BasicBlock* blk)
3040 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3043 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3044 // Operates on the map in the top-level ancestor.
3045 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3047 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3050 bool impCanReimport;
3052 bool impSpillStackEntry(unsigned level,
3056 bool bAssertOnRecursion,
3061 void impSpillStackEnsure(bool spillLeaves = false);
3062 void impEvalSideEffects();
3063 void impSpillSpecialSideEff();
3064 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3065 void impSpillValueClasses();
3066 void impSpillEvalStack();
3067 static fgWalkPreFn impFindValueClasses;
3068 void impSpillLclRefs(ssize_t lclNum);
3070 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3072 void impImportBlockCode(BasicBlock* block);
3074 void impReimportMarkBlock(BasicBlock* block);
3075 void impReimportMarkSuccessors(BasicBlock* block);
3077 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3079 void impImportBlockPending(BasicBlock* block);
3081 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3082 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3083 // for the block, but instead, just re-uses the block's existing EntryState.
3084 void impReimportBlockPending(BasicBlock* block);
3086 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3088 void impImportBlock(BasicBlock* block);
3090 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3091 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3092 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3093 // the variables that will be used -- and for all the predecessors of those successors, and the
3094 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3095 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3096 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3097 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3098 // of local variable numbers, so we represent them with the base local variable number), returns that.
3099 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3100 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3101 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3102 // on which kind of member of the clique the block is).
3103 unsigned impGetSpillTmpBase(BasicBlock* block);
3105 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3106 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3107 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3108 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3109 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3110 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3111 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3112 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3113 // successors receive a native int. Similarly float and double are unified to double.
3114 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3115 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3116 // predecessors, so they insert an upcast if needed).
3117 void impReimportSpillClique(BasicBlock* block);
3119 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3120 // block, and represent the predecessor and successor members of the clique currently being computed.
3121 // *** Access to these will need to be locked in a parallel compiler.
3122 ExpandArray<BYTE> impSpillCliquePredMembers;
3123 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3131 // Abstract class for receiving a callback while walking a spill clique
3132 class SpillCliqueWalker
3135 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3138 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3139 class SetSpillTempsBase : public SpillCliqueWalker
3144 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3147 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3150 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3151 class ReimportSpillClique : public SpillCliqueWalker
3156 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3159 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3162 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3163 // predecessor or successor within the spill clique
3164 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3166 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3167 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3168 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3169 void impRetypeEntryStateTemps(BasicBlock* blk);
3171 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3172 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3174 void impPushVar(GenTree* op, typeInfo tiRetVal);
3175 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3176 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3178 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3180 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3181 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3182 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3185 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3188 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3189 struct BlockListNode
3192 BlockListNode* m_next;
3193 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3196 void* operator new(size_t sz, Compiler* comp);
3198 BlockListNode* impBlockListNodeFreeList;
3200 BlockListNode* AllocBlockListNode();
3201 void FreeBlockListNode(BlockListNode* node);
3203 bool impIsValueType(typeInfo* pTypeInfo);
3204 var_types mangleVarArgsType(var_types type);
3207 regNumber getCallArgIntRegister(regNumber floatReg);
3208 regNumber getCallArgFloatRegister(regNumber intReg);
3209 #endif // FEATURE_VARARG
3212 static unsigned jitTotalMethodCompiled;
3216 static LONG jitNestingLevel;
3219 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3221 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3223 // STATIC inlining decision based on the IL code.
3224 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3225 CORINFO_METHOD_INFO* methInfo,
3227 InlineResult* inlineResult);
3229 void impCheckCanInline(GenTreePtr call,
3230 CORINFO_METHOD_HANDLE fncHandle,
3232 CORINFO_CONTEXT_HANDLE exactContextHnd,
3233 InlineCandidateInfo** ppInlineCandidateInfo,
3234 InlineResult* inlineResult);
3236 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3237 GenTreePtr curArgVal,
3239 InlineResult* inlineResult);
3241 void impInlineInitVars(InlineInfo* pInlineInfo);
3243 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3245 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3247 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3249 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3250 GenTreePtr variableBeingDereferenced,
3251 InlArgInfo* inlArgInfo);
3253 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3255 bool impTailCallRetTypeCompatible(var_types callerRetType,
3256 CORINFO_CLASS_HANDLE callerRetTypeClass,
3257 var_types calleeRetType,
3258 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3260 bool impIsTailCallILPattern(bool tailPrefixed,
3262 const BYTE* codeAddrOfNextOpcode,
3263 const BYTE* codeEnd,
3265 bool* IsCallPopRet = nullptr);
3267 bool impIsImplicitTailCallCandidate(
3268 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3271 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3272 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3276 XX Info about the basic-blocks, their contents and the flow analysis XX
3278 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3279 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3283 BasicBlock* fgFirstBB; // Beginning of the basic block list
3284 BasicBlock* fgLastBB; // End of the basic block list
3285 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3286 #if FEATURE_EH_FUNCLETS
3287 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3289 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3291 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3292 unsigned fgEdgeCount; // # of control flow edges between the BBs
3293 unsigned fgBBcount; // # of BBs in the method
3295 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3297 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3298 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3299 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3300 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3302 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3303 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3304 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3305 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3306 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3307 // index). The arrays are of size fgBBNumMax + 1.
3308 unsigned* fgDomTreePreOrder;
3309 unsigned* fgDomTreePostOrder;
3311 bool fgBBVarSetsInited;
3313 // Allocate array like T* a = new T[fgBBNumMax + 1];
3314 // Using helper so we don't keep forgetting +1.
3315 template <typename T>
3316 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3318 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3321 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3322 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3323 // cannot be meaningfully combined. Note that new blocks can be created with higher
3324 // block numbers without changing the basic block epoch. These blocks *cannot*
3325 // participate in a block set until the blocks are all renumbered, causing the epoch
3326 // to change. This is useful if continuing to use previous block sets is valuable.
3327 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3328 unsigned fgCurBBEpoch;
3330 unsigned GetCurBasicBlockEpoch()
3332 return fgCurBBEpoch;
3335 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3336 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3337 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3338 unsigned fgCurBBEpochSize;
3340 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3341 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3342 unsigned fgBBSetCountInSizeTUnits;
3344 void NewBasicBlockEpoch()
3346 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3348 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3350 fgCurBBEpochSize = fgBBNumMax + 1;
3351 fgBBSetCountInSizeTUnits =
3352 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3355 // All BlockSet objects are now invalid!
3356 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3357 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3361 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3362 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3363 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3364 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3366 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3367 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3368 // array of size_t bitsets), then print that out.
3369 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3376 void EnsureBasicBlockEpoch()
3378 if (fgCurBBEpochSize != fgBBNumMax + 1)
3380 NewBasicBlockEpoch();
3384 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3385 void fgEnsureFirstBBisScratch();
3386 bool fgFirstBBisScratch();
3387 bool fgBBisScratch(BasicBlock* block);
3389 void fgExtendEHRegionBefore(BasicBlock* block);
3390 void fgExtendEHRegionAfter(BasicBlock* block);
3392 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3394 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3396 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3399 BasicBlock* nearBlk,
3400 bool putInFilter = false,
3401 bool runRarely = false,
3402 bool insertAtEnd = false);
3404 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3406 bool runRarely = false,
3407 bool insertAtEnd = false);
3409 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3411 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3412 BasicBlock* afterBlk,
3413 unsigned xcptnIndex,
3414 bool putInTryRegion);
3416 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3417 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3418 void fgUnlinkBlock(BasicBlock* block);
3420 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3421 bool fgMultipleNots;
3424 bool fgModified; // True if the flow graph has been modified recently
3425 bool fgComputePredsDone; // Have we computed the bbPreds list
3426 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3427 bool fgDomsComputed; // Have we computed the dominator sets?
3429 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3430 bool fgHasPostfix; // any postfix ++/-- found?
3431 unsigned fgIncrCount; // number of increment nodes found
3433 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3437 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3438 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3441 bool fgRemoveRestOfBlock; // true if we know that we will throw
3442 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3444 // There are two modes for ordering of the trees.
3445 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3446 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3447 // by traversing the tree according to the order of the operands.
3448 // - In FGOrderLinear, the dominant ordering is the linear order.
3455 FlowGraphOrder fgOrder;
3457 // The following are boolean flags that keep track of the state of internal data structures
3459 bool fgStmtListThreaded;
3460 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3461 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3462 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3463 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3464 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3465 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3466 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3467 // This is derived from the profile data
3468 // or is BB_UNITY_WEIGHT when we don't have profile data
3470 #if FEATURE_EH_FUNCLETS
3471 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3472 #endif // FEATURE_EH_FUNCLETS
3474 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3475 // since fgMorphTree can be called from several places
3476 bool fgExpandInline; // indicates that we are creating tree for the inliner
3478 bool impBoxTempInUse; // the temp below is valid and available
3479 unsigned impBoxTemp; // a temporary that is used for boxing
3482 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3483 // and we are trying to compile again in a "safer", minopts mode?
3487 unsigned impInlinedCodeSize;
3490 //-------------------------------------------------------------------------
3498 void fgRemoveEmptyFinally();
3500 void fgCloneFinally();
3502 GenTreePtr fgGetCritSectOfStaticMethod();
3504 #if !defined(_TARGET_X86_)
3506 void fgAddSyncMethodEnterExit();
3508 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3510 void fgConvertSyncReturnToLeave(BasicBlock* block);
3512 #endif // !_TARGET_X86_
3514 void fgAddReversePInvokeEnterExit();
3516 bool fgMoreThanOneReturnBlock();
3518 // The number of separate return points in the method.
3519 unsigned fgReturnCount;
3521 void fgAddInternal();
3523 bool fgFoldConditional(BasicBlock* block);
3525 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3526 void fgMorphBlocks();
3528 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3530 void fgCheckArgCnt();
3531 void fgSetOptions();
3534 static fgWalkPreFn fgAssertNoQmark;
3535 void fgPreExpandQmarkChecks(GenTreePtr expr);
3536 void fgPostExpandQmarkChecks();
3537 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3540 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3542 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3543 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3544 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3545 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3546 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3548 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3549 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3550 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3551 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3553 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3554 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3555 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3556 void fgExpandQmarkNodes();
3560 // Do "simple lowering." This functionality is (conceptually) part of "general"
3561 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3562 void fgSimpleLowering();
3564 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3566 GenTreePtr fgInitThisClass();
3568 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3570 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3572 void fgLocalVarLiveness();
3574 void fgLocalVarLivenessInit();
3576 #ifdef LEGACY_BACKEND
3577 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3579 void fgPerNodeLocalVarLiveness(GenTree* node);
3581 void fgPerBlockLocalVarLiveness();
3583 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3585 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3587 // This is used in the liveness computation, as a temporary. When we use the
3588 // arbitrary-length VarSet representation, it is better not to allocate a new one
3590 VARSET_TP fgMarkIntfUnionVS;
3592 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3594 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3596 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3598 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3600 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3602 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3604 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3605 GenTreePtr startNode,
3607 VARSET_VALARG_TP volatileVars,
3608 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3610 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3612 bool fgRemoveDeadStore(GenTree** pTree,
3616 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3618 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3620 // For updating liveset during traversal AFTER fgComputeLife has completed
3621 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3622 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3624 // Returns the set of live variables after endTree,
3625 // assuming that liveSet is the set of live variables BEFORE tree.
3626 // Requires that fgComputeLife has completed, and that tree is in the same
3627 // statement as endTree, and that it comes before endTree in execution order
3629 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3631 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3632 while (tree != nullptr && tree != endTree->gtNext)
3634 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3635 tree = tree->gtNext;
3637 assert(tree == endTree->gtNext);
3641 void fgInterBlockLocalVarLiveness();
3643 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3644 // "x", and a def of a new SSA name for "x". The tree only has one local variable for "x", so it has to choose
3645 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3646 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3647 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3648 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3649 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3651 if (m_opAsgnVarDefSsaNums == nullptr)
3653 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3655 return m_opAsgnVarDefSsaNums;
3658 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3659 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3660 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3662 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3664 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3665 // Except: assumes that lcl is a def, and if it is
3666 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3667 // rather than the "use" SSA number recorded in the tree "lcl".
3668 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3670 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3671 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3672 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3673 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3674 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3676 // (byref addrS1 = &s1,
3677 // *(addrS1 * offsetof(f0)) = s2f0,
3679 // *(addrS1 * offsetof(fn)) = s2fn)
3681 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3682 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3683 // give it SSA names and value numbers?
3685 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3686 // end with an instance of the structure below, whose fields are described in the declaration.
3687 struct IndirectAssignmentAnnotation
3689 unsigned m_lclNum; // The local num that is being indirectly assigned.
3690 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3691 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3692 // be the singleton field sequence "g". The individual assignments would
3693 // further append the fields of "s.g" to that.
3694 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3695 // structure has a single field).
3696 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3697 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3700 IndirectAssignmentAnnotation(unsigned lclNum,
3701 FieldSeqNode* fldSeq,
3703 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3704 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3705 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3709 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3710 NodeToIndirAssignMap;
3711 NodeToIndirAssignMap* m_indirAssignMap;
3712 NodeToIndirAssignMap* GetIndirAssignMap()
3714 if (m_indirAssignMap == nullptr)
3716 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3717 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3718 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3720 return m_indirAssignMap;
3723 // Performs SSA conversion.
3726 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3727 void fgResetForSsa();
3729 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3731 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3732 inline bool fgExcludeFromSsa(unsigned lclNum);
3734 // The value numbers for this compilation.
3735 ValueNumStore* vnStore;
3738 ValueNumStore* GetValueNumStore()
3743 // Do value numbering (assign a value number to each
3745 void fgValueNumber();
3747 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3748 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3749 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3750 // match the element type of the array or fldSeq. When this type doesn't match
3751 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3753 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3756 FieldSeqNode* fldSeq,
3760 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3761 // has been parsed to yield the other input arguments. If evaluation of the address
3762 // can raise exceptions, those should be captured in the exception set "excVN."
3763 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3764 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3765 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3766 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3767 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3769 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3770 CORINFO_CLASS_HANDLE elemTypeEq,
3774 FieldSeqNode* fldSeq);
3776 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3777 // by evaluating the array index expression "tree". Returns the value number resulting from
3778 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3779 // "GT_IND" that does the dereference, and it is given the returned value number.
3780 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3782 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3784 // Utility functions for fgValueNumber.
3786 // Perform value-numbering for the trees in "blk".
3787 void fgValueNumberBlock(BasicBlock* blk);
3789 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3790 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3791 // assumed for the heap at the start "entryBlk".
3792 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3794 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3795 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3797 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3798 // value in that SSA #.
3799 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3801 // The input 'tree' is a leaf node that is a constant
3802 // Assign the proper value number to the tree
3803 void fgValueNumberTreeConst(GenTreePtr tree);
3805 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3806 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3808 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3810 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3812 // Does value-numbering for a block assignment.
3813 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3815 // Does value-numbering for a cast tree.
3816 void fgValueNumberCastTree(GenTreePtr tree);
3818 // Does value-numbering for an intrinsic tree.
3819 void fgValueNumberIntrinsic(GenTreePtr tree);
3821 // Does value-numbering for a call. We interpret some helper calls.
3822 void fgValueNumberCall(GenTreeCall* call);
3824 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3825 void fgUpdateArgListVNs(GenTreeArgList* args);
3827 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3828 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3830 // Requires "helpCall" to be a helper call. Assigns it a value number;
3831 // we understand the semantics of some of the calls. Returns "true" if
3832 // the call may modify the heap (we assume arbitrary memory side effects if so).
3833 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3835 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3836 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3838 // This is the current value number for the "Heap" implicit variable while
3839 // doing value numbering. This is the value number under the "liberal" interpretation
3840 // of heap values; the "conservative" interpretation needs no VN, since every access of
3841 // the heap yields an unknown value.
3842 ValueNum fgCurHeapVN;
3844 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3845 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3846 // is 1, and the rest is an encoding of "elemTyp".
3847 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3849 if (elemStructType != nullptr)
3851 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3852 varTypeIsIntegral(elemTyp));
3853 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3854 return elemStructType;
3858 elemTyp = varTypeUnsignedToSigned(elemTyp);
3859 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3862 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3863 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3864 // the struct type of the element).
3865 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3867 size_t clsHndVal = size_t(clsHnd);
3868 if (clsHndVal & 0x1)
3870 return var_types(clsHndVal >> 1);
3878 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3879 var_types getJitGCType(BYTE gcType);
3881 enum structPassingKind
3883 SPK_Unknown, // Invalid value, never returned
3884 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3885 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3886 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3887 // parameters registers are used, then the stack will be used)
3888 // for X86 passed on the stack, for ARM32 passed in registers
3889 // or the stack or split between registers and the stack.
3890 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3892 }; // The struct is passed/returned by reference to a copy/buffer.
3894 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3895 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3896 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3897 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3899 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3901 // Get the type that is used to pass values of the given struct type.
3902 // If you have already retrieved the struct size then pass it as the optional third argument
3904 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3905 structPassingKind* wbPassStruct,
3906 unsigned structSize = 0);
3908 // Get the type that is used to return values of the given struct type.
3909 // If you have already retrieved the struct size then pass it as the optional third argument
3911 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3912 structPassingKind* wbPassStruct = nullptr,
3913 unsigned structSize = 0);
3916 // Print a representation of "vnp" or "vn" on standard output.
3917 // If "level" is non-zero, we also print out a partial expansion of the value.
3918 void vnpPrint(ValueNumPair vnp, unsigned level);
3919 void vnPrint(ValueNum vn, unsigned level);
3922 // Dominator computation member functions
3923 // Not exposed outside Compiler
3925 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3927 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3929 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3930 // flow graph. We first assume the fields bbIDom on each
3931 // basic block are invalid. This computation is needed later
3932 // by fgBuildDomTree to build the dominance tree structure.
3933 // Based on: A Simple, Fast Dominance Algorithm
3934 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3936 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3937 // Note: this is relatively slow compared to calling fgDominate(),
3938 // especially if dealing with a single block versus block check.
3940 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3942 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3944 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3946 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3948 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3950 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3951 // processed in topological sort, this function takes care of that.
3953 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3955 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3956 // Returns this as a set.
3958 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3959 // root nodes. Returns this as a set.
3962 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
3965 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
3966 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
3969 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
3970 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
3971 // && postOrder(A) >= postOrder(B) making the computation O(1).
3972 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
3974 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
3976 void fgUpdateChangedFlowGraph();
3979 // Compute the predecessors of the blocks in the control flow graph.
3980 void fgComputePreds();
3982 // Remove all predecessor information.
3983 void fgRemovePreds();
3985 // Compute the cheap flow graph predecessors lists. This is used in some early phases
3986 // before the full predecessors lists are computed.
3987 void fgComputeCheapPreds();
3990 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
3992 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4002 // Initialize the per-block variable sets (used for liveness analysis).
4003 void fgInitBlockVarSets();
4005 // true if we've gone through and created GC Poll calls.
4006 bool fgGCPollsCreated;
4007 void fgMarkGCPollBlocks();
4008 void fgCreateGCPolls();
4009 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4011 // Requires that "block" is a block that returns from
4012 // a finally. Returns the number of successors (jump targets of
4013 // of blocks in the covered "try" that did a "LEAVE".)
4014 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4016 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4017 // a finally. Returns its "i"th successor (jump targets of
4018 // of blocks in the covered "try" that did a "LEAVE".)
4019 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4020 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4023 // Factor out common portions of the impls of the methods above.
4024 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4027 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4028 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4029 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4030 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4031 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4032 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4033 // we leave the entry associated with the block, but it will no longer be accessed.)
4034 struct SwitchUniqueSuccSet
4036 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4037 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4040 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4041 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4042 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4043 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4046 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4047 BlockToSwitchDescMap;
4050 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4051 // iteration over only the distinct successors.
4052 BlockToSwitchDescMap* m_switchDescMap;
4055 BlockToSwitchDescMap* GetSwitchDescMap()
4057 if (m_switchDescMap == nullptr)
4059 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4061 return m_switchDescMap;
4064 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4065 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4066 // we don't accidentally look up and return the wrong switch data.
4067 void InvalidateUniqueSwitchSuccMap()
4069 m_switchDescMap = nullptr;
4072 // Requires "switchBlock" to be a block that ends in a switch. Returns
4073 // the corresponding SwitchUniqueSuccSet.
4074 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4076 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4077 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4078 // remove it from "this", and ensure that "to" is a member.
4079 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4081 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4082 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4084 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4086 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4088 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4090 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4092 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4094 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4096 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4098 void fgRemoveBlockAsPred(BasicBlock* block);
4100 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4102 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4104 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4106 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4108 flowList* fgAddRefPred(BasicBlock* block,
4109 BasicBlock* blockPred,
4110 flowList* oldEdge = nullptr,
4111 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4114 void fgFindBasicBlocks();
4116 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4118 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4120 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4121 bool putInTryRegion,
4122 BasicBlock* startBlk,
4124 BasicBlock* nearBlk,
4125 BasicBlock* jumpBlk,
4128 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4130 void fgRemoveEmptyBlocks();
4132 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4134 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4136 void fgCreateLoopPreHeader(unsigned lnum);
4138 void fgUnreachableBlock(BasicBlock* block);
4140 void fgRemoveConditionalJump(BasicBlock* block);
4142 BasicBlock* fgLastBBInMainFunction();
4144 BasicBlock* fgEndBBAfterMainFunction();
4146 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4148 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4150 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4152 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4154 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4156 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4158 bool fgRenumberBlocks();
4160 bool fgExpandRarelyRunBlocks();
4162 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4164 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4166 enum FG_RELOCATE_TYPE
4168 FG_RELOCATE_TRY, // relocate the 'try' region
4169 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4171 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4173 #if FEATURE_EH_FUNCLETS
4174 #if defined(_TARGET_ARM_)
4175 void fgClearFinallyTargetBit(BasicBlock* block);
4176 #endif // defined(_TARGET_ARM_)
4177 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4178 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4179 void fgInsertFuncletPrologBlock(BasicBlock* block);
4180 void fgCreateFuncletPrologBlocks();
4181 void fgCreateFunclets();
4182 #else // !FEATURE_EH_FUNCLETS
4183 bool fgRelocateEHRegions();
4184 #endif // !FEATURE_EH_FUNCLETS
4186 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4188 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4190 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4192 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4194 bool fgOptimizeEmptyBlock(BasicBlock* block);
4196 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4198 bool fgOptimizeBranch(BasicBlock* bJump);
4200 bool fgOptimizeSwitchBranches(BasicBlock* block);
4202 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4204 bool fgOptimizeSwitchJumps();
4206 void fgPrintEdgeWeights();
4208 void fgComputeEdgeWeights();
4210 void fgReorderBlocks();
4212 void fgDetermineFirstColdBlock();
4214 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4216 bool fgUpdateFlowGraph(bool doTailDup = false);
4218 void fgFindOperOrder();
4220 // method that returns if you should split here
4221 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4223 void fgSetBlockOrder();
4225 void fgRemoveReturnBlock(BasicBlock* block);
4227 /* Helper code that has been factored out */
4228 inline void fgConvertBBToThrowBB(BasicBlock* block);
4230 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4231 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4232 GenTreePtr fgMakeTmpArgNode(
4233 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4235 // The following check for loops that don't execute calls
4236 bool fgLoopCallMarked;
4238 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4239 void fgLoopCallMark();
4241 void fgMarkLoopHead(BasicBlock* block);
4243 unsigned fgGetCodeEstimate(BasicBlock* block);
4246 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4247 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4248 bool fgDumpFlowGraph(Phases phase);
4250 #endif // DUMP_FLOWGRAPHS
4255 void fgDispBBLiveness(BasicBlock* block);
4256 void fgDispBBLiveness();
4257 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4258 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4259 void fgDispBasicBlocks(bool dumpTrees = false);
4260 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4261 void fgDumpBlock(BasicBlock* block);
4262 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4264 static fgWalkPreFn fgStress64RsltMulCB;
4265 void fgStress64RsltMul();
4266 void fgDebugCheckUpdate();
4267 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4268 void fgDebugCheckBlockLinks();
4269 void fgDebugCheckLinks(bool morphTrees = false);
4270 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4271 void fgDebugCheckFlags(GenTreePtr tree);
4272 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4273 void fgDebugCheckTryFinallyExits();
4276 #ifdef LEGACY_BACKEND
4277 static void fgOrderBlockOps(GenTreePtr tree,
4281 GenTreePtr* opsPtr, // OUT
4282 regMaskTP* regsPtr); // OUT
4283 #endif // LEGACY_BACKEND
4285 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4286 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4288 inline bool fgIsInlining()
4290 return fgExpandInline;
4293 void fgTraverseRPO();
4295 //--------------------- Walking the trees in the IR -----------------------
4300 fgWalkPreFn* wtprVisitorFn;
4301 fgWalkPostFn* wtpoVisitorFn;
4302 void* pCallbackData; // user-provided data
4303 bool wtprLclsOnly; // whether to only visit lclvar nodes
4304 GenTreePtr parent; // parent of current node, provided to callback
4305 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4307 bool printModified; // callback can use this
4311 template <bool computeStack>
4312 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4314 // general purpose tree-walker that is capable of doing pre- and post- order
4315 // callbacks at the same time
4316 template <bool doPreOrder, bool doPostOrder>
4317 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4319 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4320 fgWalkPreFn* visitor,
4321 void* pCallBackData = nullptr,
4322 bool lclVarsOnly = false,
4323 bool computeStack = false);
4325 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4326 fgWalkPreFn* preVisitor,
4327 fgWalkPostFn* postVisitor,
4328 void* pCallBackData = nullptr);
4330 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4334 template <bool computeStack>
4335 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4337 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4338 fgWalkPostFn* visitor,
4339 void* pCallBackData = nullptr,
4340 bool computeStack = false);
4342 // An fgWalkPreFn that looks for expressions that have inline throws in
4343 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4344 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4345 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4346 // properly propagated to parent trees). It returns WALK_CONTINUE
4348 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4349 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4350 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4352 /**************************************************************************
4354 *************************************************************************/
4357 friend class SsaBuilder;
4358 friend struct ValueNumberState;
4360 //--------------------- Detect the basic blocks ---------------------------
4362 BasicBlock** fgBBs; // Table of pointers to the BBs
4364 void fgInitBBLookup();
4365 BasicBlock* fgLookupBB(unsigned addr);
4367 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4369 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4371 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4373 void fgLinkBasicBlocks();
4375 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4377 void fgCheckBasicBlockControlFlow();
4379 void fgControlFlowPermitted(BasicBlock* blkSrc,
4380 BasicBlock* blkDest,
4381 BOOL IsLeave = false /* is the src a leave block */);
4383 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4385 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4387 void fgAdjustForAddressExposedOrWrittenThis();
4389 bool fgProfileData_ILSizeMismatch;
4390 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4391 ULONG fgProfileBufferCount;
4392 ULONG fgNumProfileRuns;
4394 unsigned fgStressBBProf()
4397 unsigned result = JitConfig.JitStressBBProf();
4400 if (compStressCompile(STRESS_BB_PROFILE, 15))
4411 bool fgHaveProfileData();
4412 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4414 bool fgIsUsingProfileWeights()
4416 return (fgHaveProfileData() || fgStressBBProf());
4418 void fgInstrumentMethod();
4420 //-------- Insert a statement at the start or end of a basic block --------
4424 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4428 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4430 public: // Used by linear scan register allocation
4431 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4434 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4435 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4437 public: // Used by linear scan register allocation
4438 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4441 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4443 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4445 // Create a new temporary variable to hold the result of *ppTree,
4446 // and transform the graph accordingly.
4447 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4448 GenTree* fgMakeMultiUse(GenTree** ppTree);
4451 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4452 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4453 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4455 //-------- Determine the order in which the trees will be evaluated -------
4457 unsigned fgTreeSeqNum;
4458 GenTree* fgTreeSeqLst;
4459 GenTree* fgTreeSeqBeg;
4461 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4462 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4463 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4464 void fgSetStmtSeq(GenTree* tree);
4465 void fgSetBlockOrder(BasicBlock* block);
4467 //------------------------- Morphing --------------------------------------
4469 unsigned fgPtrArgCntCur;
4470 unsigned fgPtrArgCntMax;
4471 hashBv* fgOutgoingArgTemps;
4472 hashBv* fgCurrentlyInUseArgTemps;
4474 bool compCanEncodePtrArgCntMax();
4476 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4479 void fgMoveOpsLeft(GenTreePtr tree);
4482 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4484 bool fgIsThrow(GenTreePtr tree);
4486 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4487 bool fgIsBlockCold(BasicBlock* block);
4489 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4491 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4493 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4495 bool fgMorphRelopToQmark(GenTreePtr tree);
4497 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4498 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4499 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4500 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4501 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4502 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4503 // small; hence the other fields of MorphAddrContext.
4504 enum MorphAddrContextKind
4509 struct MorphAddrContext
4511 MorphAddrContextKind m_kind;
4512 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4513 // top-level indirection and here have been constants.
4514 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4515 // In that case, is the sum of those constant offsets.
4517 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4522 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4523 static MorphAddrContext s_CopyBlockMAC;
4526 GenTreePtr fgCopySIMDNode(GenTreeSIMD* simdNode);
4527 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4528 var_types* baseTypeOut,
4530 unsigned* simdSizeOut,
4531 bool ignoreUsedInSIMDIntrinsic = false);
4532 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4533 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4534 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4535 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4537 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4538 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4539 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4541 #endif // FEATURE_SIMD
4542 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4543 GenTreePtr fgMorphCast(GenTreePtr tree);
4544 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4545 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4547 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4550 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4551 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4553 void fgFixupStructReturn(GenTreePtr call);
4554 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4555 bool fgAddrCouldBeNull(GenTreePtr addr);
4556 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4557 bool fgCanFastTailCall(GenTreeCall* call);
4558 void fgMorphTailCall(GenTreeCall* call);
4559 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4560 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4561 fgArgTabEntryPtr argTabEntry,
4563 IL_OFFSETX callILOffset,
4564 GenTreePtr tmpAssignmentInsertionPoint,
4565 GenTreePtr paramAssignmentInsertionPoint);
4566 static int fgEstimateCallStackSize(GenTreeCall* call);
4567 GenTreePtr fgMorphCall(GenTreeCall* call);
4568 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4569 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4571 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4572 static fgWalkPreFn fgFindNonInlineCandidate;
4574 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4575 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4576 void fgAssignSetVarDef(GenTreePtr tree);
4577 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4578 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4579 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4580 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4581 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4582 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4583 void fgMorphUnsafeBlk(GenTreeObj* obj);
4584 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4585 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4586 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4587 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4588 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4589 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4590 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4592 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4593 GenTreePtr fgMorphConst(GenTreePtr tree);
4596 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4599 #if LOCAL_ASSERTION_PROP
4600 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4601 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4603 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4605 GenTreeStmt* fgMorphStmt;
4607 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4608 // used when morphing big offset.
4610 //----------------------- Liveness analysis -------------------------------
4612 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4613 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4615 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4616 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4617 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4619 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4621 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4622 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4624 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4625 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4627 void fgExtendDbgScopes();
4628 void fgExtendDbgLifetimes();
4631 void fgDispDebugScopes();
4634 //-------------------------------------------------------------------------
4636 // The following keeps track of any code we've added for things like array
4637 // range checking or explicit calls to enable GC, and so on.
4642 AddCodeDsc* acdNext;
4643 BasicBlock* acdDstBlk; // block to which we jump
4645 SpecialCodeKind acdKind; // what kind of a special block is this?
4646 unsigned short acdStkLvl;
4650 static unsigned acdHelper(SpecialCodeKind codeKind);
4652 AddCodeDsc* fgAddCodeList;
4654 bool fgRngChkThrowAdded;
4655 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4657 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4659 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4662 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4665 bool fgIsCodeAdded();
4667 bool fgIsThrowHlpBlk(BasicBlock* block);
4668 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4670 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4672 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4673 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4674 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4675 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4676 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4678 #if FEATURE_MULTIREG_RET
4679 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4680 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4681 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4682 #endif // FEATURE_MULTIREG_RET
4684 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4687 static fgWalkPreFn fgDebugCheckInlineCandidates;
4690 void fgPromoteStructs();
4691 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4692 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4693 void fgMarkImplicitByRefArgs();
4694 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4695 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4696 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4697 void fgMarkAddressExposedLocals();
4698 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4700 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4702 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4704 // The given local variable, required to be a struct variable, is being assigned via
4705 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4706 // the variable is not enregistered, and is therefore not promoted independently.
4707 void fgLclFldAssign(unsigned lclNum);
4709 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4710 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4711 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4712 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4715 bool fgPrintInlinedMethods;
4718 bool fgIsBigOffset(size_t offset);
4720 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4721 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4722 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4723 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4724 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4727 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4728 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4732 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4740 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4743 void optRemoveRangeCheck(
4744 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4745 bool optIsRangeCheckRemovable(GenTreePtr tree);
4748 static fgWalkPreFn optValidRangeCheckIndex;
4749 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4752 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4754 /**************************************************************************
4756 *************************************************************************/
4759 // Do hoisting for all loops.
4760 void optHoistLoopCode();
4762 // To represent sets of VN's that have already been hoisted in outer loops.
4763 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4764 typedef VNToBoolMap VNSet;
4766 struct LoopHoistContext
4769 // The set of variables hoisted in the current loop (or nullptr if there are none).
4770 VNSet* m_pHoistedInCurLoop;
4773 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4774 VNSet m_hoistedInParentLoops;
4775 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4776 // Previous decisions on loop-invariance of value numbers in the current loop.
4777 VNToBoolMap m_curLoopVnInvariantCache;
4779 VNSet* GetHoistedInCurLoop(Compiler* comp)
4781 if (m_pHoistedInCurLoop == nullptr)
4783 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4785 return m_pHoistedInCurLoop;
4788 VNSet* ExtractHoistedInCurLoop()
4790 VNSet* res = m_pHoistedInCurLoop;
4791 m_pHoistedInCurLoop = nullptr;
4795 LoopHoistContext(Compiler* comp)
4796 : m_pHoistedInCurLoop(nullptr)
4797 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4798 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4803 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4804 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4805 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4806 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4808 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4809 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4810 // "m_hoistedInParentLoops".
4812 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4814 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4815 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4816 // expressions to "hoistInLoop".
4817 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4819 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4820 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4822 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4823 // that are invariant in loop "lnum" (an index into the optLoopTable)
4824 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4825 // expressions to "hoistInLoop".
4826 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4827 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4828 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4829 bool optHoistLoopExprsForTree(GenTreePtr tree,
4831 LoopHoistContext* hoistCtxt,
4832 bool* firstBlockAndBeforeSideEffect,
4835 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4836 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4838 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4839 // Constants and init values are always loop invariant.
4840 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4841 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4843 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4844 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4845 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4846 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4847 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4849 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4850 // in the loop table.
4851 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4853 // Records the set of "side effects" of all loops: fields (object instance and static)
4854 // written to, and SZ-array element type equivalence classes updated.
4855 void optComputeLoopSideEffects();
4858 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4859 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4860 // static) written to, and SZ-array element type equivalence classes updated.
4861 void optComputeLoopNestSideEffects(unsigned lnum);
4863 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4864 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4866 // Hoist the expression "expr" out of loop "lnum".
4867 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4870 void optOptimizeBools();
4873 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4875 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4878 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4880 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4881 // the loop into a "do-while" loop
4882 // Also finds all natural loops and records them in the loop table
4884 // Optionally clone loops in the loop table.
4885 void optCloneLoops();
4887 // Clone loop "loopInd" in the loop table.
4888 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4890 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4891 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4892 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4894 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4896 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4899 // This enumeration describes what is killed by a call.
4903 CALLINT_NONE, // no interference (most helpers)
4904 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4905 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4906 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4907 CALLINT_ALL, // kills everything (normal method call)
4911 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4912 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4913 // in bbNext order; we use comparisons on the bbNum to decide order.)
4914 // The blocks that define the body are
4915 // first <= top <= entry <= bottom .
4916 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4917 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4918 // Compiler::optFindNaturalLoops().
4921 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4922 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4923 // loop, but not the outer loop.)
4924 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4926 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4927 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4928 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4930 callInterf lpAsgCall; // "callInterf" for calls in the loop
4931 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4932 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4934 unsigned short lpFlags; // Mask of the LPFLG_* constants
4936 unsigned char lpExitCnt; // number of exits from the loop
4938 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4939 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4940 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4941 // (Actually, an "immediately" nested loop --
4942 // no other child of this loop is a parent of lpChild.)
4943 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4944 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4945 // by following "lpChild" then "lpSibling" links.
4947 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4948 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4950 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4951 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4952 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4954 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4955 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4957 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4958 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4959 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4960 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4962 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
4963 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
4964 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
4966 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
4967 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
4968 // type are assigned to.
4970 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
4971 // If this is set, the fields below may not be accurate (since they become irrelevant.)
4972 bool lpContainsCall; // True if executing the loop body *may* execute a call
4974 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
4975 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
4977 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
4979 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
4980 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
4982 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
4984 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
4985 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
4987 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
4988 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
4990 JitSimplerHashBehavior>
4992 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
4993 // instance fields modified
4996 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
4997 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
4999 JitSimplerHashBehavior>
5001 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5002 // arrays of that type are modified
5005 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5006 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5008 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5009 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5010 // (shifted left, with a low-order bit set to distinguish.)
5011 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5012 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5014 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5016 GenTreePtr lpIterTree; // The "i <op>= const" tree
5017 unsigned lpIterVar(); // iterator variable #
5018 int lpIterConst(); // the constant with which the iterator is incremented
5019 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5020 void VERIFY_lpIterTree();
5022 var_types lpIterOperType(); // For overflow instructions
5025 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5026 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5030 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5032 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5033 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5034 void VERIFY_lpTestTree();
5036 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5037 GenTreePtr lpIterator(); // the iterator node in the loop test
5038 GenTreePtr lpLimit(); // the limit node in the loop test
5040 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5041 // LPFLG_CONST_LIMIT
5042 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5044 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5045 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5046 // LPFLG_ARRLEN_LIMIT
5048 // Returns "true" iff "*this" contains the blk.
5049 bool lpContains(BasicBlock* blk)
5051 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5053 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5054 // to be equal, but requiring bottoms to be different.)
5055 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5057 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5060 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5061 // bottoms to be different.)
5062 bool lpContains(const LoopDsc& lp2)
5064 return lpContains(lp2.lpFirst, lp2.lpBottom);
5067 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5068 // (allowing firsts to be equal, but requiring bottoms to be different.)
5069 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5071 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5074 // Returns "true" iff "*this" is (properly) contained by "lp2"
5075 // (allowing firsts to be equal, but requiring bottoms to be different.)
5076 bool lpContainedBy(const LoopDsc& lp2)
5078 return lpContains(lp2.lpFirst, lp2.lpBottom);
5081 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5082 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5084 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5086 // Returns "true" iff "*this" is disjoint from "lp2".
5087 bool lpDisjoint(const LoopDsc& lp2)
5089 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5091 // Returns "true" iff the loop is well-formed (see code for defn).
5094 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5095 lpEntry->bbNum <= lpBottom->bbNum &&
5096 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5101 bool fgMightHaveLoop(); // returns true if there are any backedges
5102 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5105 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5106 unsigned char optLoopCount; // number of tracked loops
5109 unsigned optCallCount; // number of calls made in the method
5110 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5111 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5112 unsigned optLoopsCloned; // number of loops cloned in the current method.
5115 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5116 void optPrintLoopInfo(unsigned loopNum,
5118 BasicBlock* lpFirst,
5120 BasicBlock* lpEntry,
5121 BasicBlock* lpBottom,
5122 unsigned char lpExitCnt,
5124 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5125 void optPrintLoopInfo(unsigned lnum);
5126 void optPrintLoopRecording(unsigned lnum);
5128 void optCheckPreds();
5131 void optSetBlockWeights();
5133 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5135 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5137 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5139 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5140 unsigned optIsLoopIncrTree(GenTreePtr incr);
5141 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5142 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5143 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5144 bool optExtractInitTestIncr(BasicBlock* head,
5149 GenTreePtr* ppIncr);
5151 void optRecordLoop(BasicBlock* head,
5157 unsigned char exitCnt);
5159 void optFindNaturalLoops();
5161 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5162 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5163 bool optCanonicalizeLoopNest(unsigned char loopInd);
5165 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5166 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5167 bool optCanonicalizeLoop(unsigned char loopInd);
5169 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5170 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5171 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5172 bool optLoopContains(unsigned l1, unsigned l2);
5174 // Requires "loopInd" to be a valid index into the loop table.
5175 // Updates the loop table by changing loop "loopInd", whose head is required
5176 // to be "from", to be "to". Also performs this transformation for any
5177 // loop nested in "loopInd" that shares the same head as "loopInd".
5178 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5180 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5181 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5182 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5184 // Marks the containsCall information to "lnum" and any parent loops.
5185 void AddContainsCallAllContainingLoops(unsigned lnum);
5186 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5187 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5188 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5189 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5190 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5191 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5193 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5194 // of "from".) Copies the jump destination from "from" to "to".
5195 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5197 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5198 unsigned optLoopDepth(unsigned lnum)
5200 unsigned par = optLoopTable[lnum].lpParent;
5201 if (par == BasicBlock::NOT_IN_LOOP)
5207 return 1 + optLoopDepth(par);
5211 void fgOptWhileLoop(BasicBlock* block);
5213 bool optComputeLoopRep(int constInit,
5216 genTreeOps iterOper,
5218 genTreeOps testOper,
5221 unsigned* iterCount);
5222 #if FEATURE_STACK_FP_X87
5225 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5226 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5227 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5228 #endif // FEATURE_STACK_FP_X87
5231 static fgWalkPreFn optIsVarAssgCB;
5234 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5236 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5238 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5240 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5242 /**************************************************************************
5243 * Optimization conditions
5244 *************************************************************************/
5246 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5247 bool optPentium4(void);
5248 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5249 bool optAvoidIntMult(void);
5254 // The following is the upper limit on how many expressions we'll keep track
5255 // of for the CSE analysis.
5257 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5259 static const int MIN_CSE_COST = 2;
5261 // Keeps tracked cse indices
5262 BitVecTraits* cseTraits;
5266 /* Generic list of nodes - used by the CSE logic */
5274 typedef struct treeLst* treeLstPtr;
5278 treeStmtLst* tslNext;
5279 GenTreePtr tslTree; // tree node
5280 GenTreePtr tslStmt; // statement containing the tree
5281 BasicBlock* tslBlock; // block containing the statement
5284 typedef struct treeStmtLst* treeStmtLstPtr;
5286 // The following logic keeps track of expressions via a simple hash table.
5290 CSEdsc* csdNextInBucket; // used by the hash table
5292 unsigned csdHashValue; // the orginal hashkey
5294 unsigned csdIndex; // 1..optCSECandidateCount
5295 char csdLiveAcrossCall; // 0 or 1
5297 unsigned short csdDefCount; // definition count
5298 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5300 unsigned csdDefWtCnt; // weighted def count
5301 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5303 GenTreePtr csdTree; // treenode containing the 1st occurance
5304 GenTreePtr csdStmt; // stmt containing the 1st occurance
5305 BasicBlock* csdBlock; // block containing the 1st occurance
5307 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5308 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5311 static const size_t s_optCSEhashSize;
5312 CSEdsc** optCSEhash;
5317 CSEdsc* optCSEfindDsc(unsigned index);
5318 void optUnmarkCSE(GenTreePtr tree);
5320 // user defined callback data for the tree walk function optCSE_MaskHelper()
5321 struct optCSE_MaskData
5323 EXPSET_TP CSE_defMask;
5324 EXPSET_TP CSE_useMask;
5327 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5328 static fgWalkPreFn optCSE_MaskHelper;
5330 // This function walks all the node for an given tree
5331 // and return the mask of CSE definitions and uses for the tree
5333 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5335 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5336 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5337 bool optCSE_canSwap(GenTree* tree);
5339 static fgWalkPostFn optPropagateNonCSE;
5340 static fgWalkPreFn optHasNonCSEChild;
5342 static fgWalkPreFn optUnmarkCSEs;
5344 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5345 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5347 void optCleanupCSEs();
5350 void optEnsureClearCSEInfo();
5353 #endif // FEATURE_ANYCSE
5355 #if FEATURE_VALNUM_CSE
5356 /**************************************************************************
5357 * Value Number based CSEs
5358 *************************************************************************/
5361 void optOptimizeValnumCSEs();
5364 void optValnumCSE_Init();
5365 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5366 unsigned optValnumCSE_Locate();
5367 void optValnumCSE_InitDataFlow();
5368 void optValnumCSE_DataFlow();
5369 void optValnumCSE_Availablity();
5370 void optValnumCSE_Heuristic();
5371 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5373 #endif // FEATURE_VALNUM_CSE
5376 bool optDoCSE; // True when we have found a duplicate CSE tree
5377 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5378 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5379 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5380 unsigned optCSEstart; // The first local variable number that is a CSE
5381 unsigned optCSEcount; // The total count of CSE's introduced.
5382 unsigned optCSEweight; // The weight of the current block when we are
5383 // scanning for CSE expressions
5385 bool optIsCSEcandidate(GenTreePtr tree);
5387 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5389 bool lclNumIsTrueCSE(unsigned lclNum) const
5391 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5394 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5396 bool lclNumIsCSE(unsigned lclNum) const
5398 return lvaTable[lclNum].lvIsCSE;
5402 bool optConfigDisableCSE();
5403 bool optConfigDisableCSE2();
5405 void optOptimizeCSEs();
5407 #endif // FEATURE_ANYCSE
5415 unsigned ivaVar; // Variable we are interested in, or -1
5416 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5417 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5418 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5419 callInterf ivaMaskCall; // What kind of calls are there?
5422 static callInterf optCallInterf(GenTreePtr call);
5425 // VN based copy propagation.
5426 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5427 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5428 LclNumToGenTreePtrStack;
5430 // Kill set to track variables with intervening definitions.
5431 VARSET_TP optCopyPropKillSet;
5433 // Copy propagation functions.
5434 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5435 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5436 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5437 bool optIsSsaLocal(GenTreePtr tree);
5438 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5439 void optVnCopyProp();
5441 /**************************************************************************
5442 * Early value propagation
5443 *************************************************************************/
5449 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5453 static unsigned GetHashCode(SSAName ssaNm)
5455 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5458 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5460 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5464 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5465 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5466 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5467 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5468 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5470 unsigned optMethodFlags;
5472 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5473 // No throughput diff was found with backward walk bound between 3-8.
5474 static const int optEarlyPropRecurBound = 5;
5476 enum class optPropKind
5484 bool gtIsVtableRef(GenTreePtr tree);
5485 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5486 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5487 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5488 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5489 bool optEarlyPropRewriteTree(GenTreePtr tree);
5490 bool optDoEarlyPropForBlock(BasicBlock* block);
5491 bool optDoEarlyPropForFunc();
5492 void optEarlyProp();
5493 void optFoldNullCheck(GenTreePtr tree);
5494 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5497 /**************************************************************************
5498 * Value/Assertion propagation
5499 *************************************************************************/
5501 // Data structures for assertion prop
5502 BitVecTraits* apTraits;
5506 enum optAssertionKind
5521 O1K_ARRLEN_OPER_BND,
5522 O1K_ARRLEN_LOOP_BND,
5523 O1K_CONSTANT_LOOP_BND,
5544 optAssertionKind assertionKind;
5547 unsigned lclNum; // assigned to or property of this local var number
5555 struct AssertionDscOp1
5557 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5564 struct AssertionDscOp2
5566 optOp2Kind kind; // a const or copy assignment
5570 ssize_t iconVal; // integer
5571 unsigned iconFlags; // gtFlags
5573 struct Range // integer subrange
5587 bool IsArrLenArithBound()
5589 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5591 bool IsArrLenBound()
5593 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5595 bool IsConstantBound()
5597 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5598 op1.kind == O1K_CONSTANT_LOOP_BND);
5600 bool IsBoundsCheckNoThrow()
5602 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5605 bool IsCopyAssertion()
5607 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5610 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5612 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5613 a1->op2.kind == a2->op2.kind;
5616 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5618 if (kind == OAK_EQUAL)
5620 return kind2 == OAK_NOT_EQUAL;
5622 else if (kind == OAK_NOT_EQUAL)
5624 return kind2 == OAK_EQUAL;
5629 static ssize_t GetLowerBoundForIntegralType(var_types type)
5649 static ssize_t GetUpperBoundForIntegralType(var_types type)
5673 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5675 return (op1.kind == that->op1.kind) &&
5676 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5679 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5681 if (op2.kind != that->op2.kind)
5687 case O2K_IND_CNS_INT:
5689 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5691 case O2K_CONST_LONG:
5692 return (op2.lconVal == that->op2.lconVal);
5694 case O2K_CONST_DOUBLE:
5695 // exact match because of positive and negative zero.
5696 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5698 case O2K_LCLVAR_COPY:
5700 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5701 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5704 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5707 // we will return false
5711 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5717 bool Complementary(AssertionDsc* that, bool vnBased)
5719 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5720 HasSameOp2(that, vnBased);
5723 bool Equals(AssertionDsc* that, bool vnBased)
5725 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5729 typedef unsigned short AssertionIndex;
5732 static fgWalkPreFn optAddCopiesCallback;
5733 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5734 unsigned optAddCopyLclNum;
5735 GenTreePtr optAddCopyAsgnNode;
5737 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5738 bool optAssertionPropagated; // set to true if we modified the trees
5739 bool optAssertionPropagatedCurrentStmt;
5741 GenTreePtr optAssertionPropCurrentTree;
5743 AssertionIndex* optComplementaryAssertionMap;
5744 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5745 // using the value of a local var) for each local var
5746 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5747 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5748 AssertionIndex optMaxAssertionCount;
5751 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5752 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5753 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5754 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5755 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5756 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5758 AssertionIndex GetAssertionCount()
5760 return optAssertionCount;
5762 ASSERT_TP* bbJtrueAssertionOut;
5763 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5764 ValueNumToAssertsMap;
5765 ValueNumToAssertsMap* optValueNumToAsserts;
5767 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5769 // Assertion prop helpers.
5770 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5771 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5772 void optAssertionInit(bool isLocalProp);
5773 void optAssertionTraitsInit(AssertionIndex assertionCount);
5774 #if LOCAL_ASSERTION_PROP
5775 void optAssertionReset(AssertionIndex limit);
5776 void optAssertionRemove(AssertionIndex index);
5779 // Assertion prop data flow functions.
5780 void optAssertionPropMain();
5781 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5782 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5783 ASSERT_TP* optInitAssertionDataflowFlags();
5784 ASSERT_TP* optComputeAssertionGen();
5786 // Assertion Gen functions.
5787 void optAssertionGen(GenTreePtr tree);
5788 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5789 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5790 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5791 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5792 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5793 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5795 // Assertion creation functions.
5796 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5797 AssertionIndex optCreateAssertion(GenTreePtr op1,
5799 optAssertionKind assertionKind,
5800 AssertionDsc* assertion);
5801 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5803 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5804 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5805 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5807 void optPrintVnAssertionMapping();
5809 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5811 // Used for respective assertion propagations.
5812 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5813 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5814 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5815 bool optAssertionIsNonNull(GenTreePtr op,
5816 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5818 // Used for Relop propagation.
5819 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5820 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5821 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5823 // Assertion prop for lcl var functions.
5824 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5825 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5827 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5828 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5829 const GenTreePtr tree,
5830 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5831 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5833 // Assertion propagation functions.
5834 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5835 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5836 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5837 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5838 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5839 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5840 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5841 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5842 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5843 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5844 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5845 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5847 // Implied assertion functions.
5848 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5849 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5850 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5851 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5853 ASSERT_VALRET_TP optNewFullAssertSet();
5854 ASSERT_VALRET_TP optNewEmptyAssertSet();
5857 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5858 void optDebugCheckAssertion(AssertionDsc* assertion);
5859 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5861 void optAddCopies();
5862 #endif // ASSERTION_PROP
5864 /**************************************************************************
5866 *************************************************************************/
5869 struct LoopCloneVisitorInfo
5871 LoopCloneContext* context;
5874 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5875 : context(context), loopNum(loopNum), stmt(nullptr)
5880 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5881 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5882 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5883 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5884 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5885 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5886 void optObtainLoopCloningOpts(LoopCloneContext* context);
5887 bool optIsLoopClonable(unsigned loopInd);
5889 bool optCanCloneLoops();
5892 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5894 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5895 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5896 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5897 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5901 void optInsertLoopCloningStress(BasicBlock* head);
5903 #if COUNT_RANGECHECKS
5904 static unsigned optRangeChkRmv;
5905 static unsigned optRangeChkAll;
5914 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5919 RngChkDsc* rcdNextInBucket; // used by the hash table
5921 unsigned short rcdHashValue; // to make matching faster
5922 unsigned short rcdIndex; // 0..optRngChkCount-1
5924 GenTreePtr rcdTree; // the array index tree
5927 unsigned optRngChkCount;
5928 static const size_t optRngChkHashSize;
5930 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5931 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5933 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5936 bool optLoopsMarked;
5939 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5940 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5944 XX Does the register allocation and puts the remaining lclVars on the stack XX
5946 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5947 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5951 #ifndef LEGACY_BACKEND
5956 #else // LEGACY_BACKEND
5961 #endif // LEGACY_BACKEND
5963 #ifdef LEGACY_BACKEND
5965 void raAssignVars(); // register allocation
5966 #endif // LEGACY_BACKEND
5968 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
5970 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
5972 void raMarkStkVars();
5975 // Some things are used by both LSRA and regpredict allocators.
5977 FrameType rpFrameType;
5978 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
5980 #ifdef LEGACY_BACKEND
5981 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
5983 #endif // LEGACY_BACKEND
5985 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
5987 #if FEATURE_FP_REGALLOC
5988 enum enumConfigRegisterFP
5990 CONFIG_REGISTER_FP_NONE = 0x0,
5991 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
5992 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
5993 CONFIG_REGISTER_FP_FULL = 0x3,
5995 enumConfigRegisterFP raConfigRegisterFP();
5996 #endif // FEATURE_FP_REGALLOC
5999 regMaskTP raConfigRestrictMaskFP();
6002 #ifndef LEGACY_BACKEND
6003 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6004 #else // LEGACY_BACKEND
6005 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6006 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6007 bool raNewBlocks; // True is we added killing blocks for FPU registers
6008 unsigned rpPasses; // Number of passes made by the register predicter
6009 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6010 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6011 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6012 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6013 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6014 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6015 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6016 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6017 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6018 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6019 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6020 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6022 bool rpRegAllocDone; // Set to true after we have completed register allocation
6024 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6026 void raSetupArgMasks(RegState* r);
6028 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6030 void raDumpVarIntf(); // Dump the variable to variable interference graph
6031 void raDumpRegIntf(); // Dump the variable to register interference graph
6033 void raAdjustVarIntf();
6035 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6037 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6039 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6040 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6042 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6044 static fgWalkPreFn rpMarkRegIntf;
6046 regMaskTP rpPredictAddressMode(
6047 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6049 void rpPredictRefAssign(unsigned lclNum);
6051 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6053 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6055 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6057 void rpPredictRegUse(); // Entry point
6059 unsigned raPredictTreeRegUse(GenTreePtr tree);
6060 unsigned raPredictListRegUse(GenTreePtr list);
6062 void raSetRegVarOrder(var_types regType,
6063 regNumber* customVarOrder,
6064 unsigned* customVarOrderSize,
6066 regMaskTP avoidReg);
6068 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6069 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6070 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6071 void raAddToStkPredict(unsigned val)
6073 unsigned newStkPredict = rpStkPredict + val;
6074 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6075 rpStkPredict = UINT_MAX - 1;
6077 rpStkPredict = newStkPredict;
6081 #if !FEATURE_FP_REGALLOC
6082 void raDispFPlifeInfo();
6086 regMaskTP genReturnRegForTree(GenTreePtr tree);
6087 #endif // LEGACY_BACKEND
6089 /* raIsVarargsStackArg is called by raMaskStkVars and by
6090 lvaSortByRefCount. It identifies the special case
6091 where a varargs function has a parameter passed on the
6092 stack, other than the special varargs handle. Such parameters
6093 require special treatment, because they cannot be tracked
6094 by the GC (their offsets in the stack are not known
6098 bool raIsVarargsStackArg(unsigned lclNum)
6102 LclVarDsc* varDsc = &lvaTable[lclNum];
6104 assert(varDsc->lvIsParam);
6106 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6108 #else // _TARGET_X86_
6112 #endif // _TARGET_X86_
6115 #ifdef LEGACY_BACKEND
6116 // Records the current prediction, if it's better than any previous recorded prediction.
6117 void rpRecordPrediction();
6118 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6119 void rpUseRecordedPredictionIfBetter();
6121 // Data members used in the methods above.
6122 unsigned rpBestRecordedStkPredict;
6123 struct VarRegPrediction
6125 bool m_isEnregistered;
6126 regNumberSmall m_regNum;
6127 regNumberSmall m_otherReg;
6129 VarRegPrediction* rpBestRecordedPrediction;
6130 #endif // LEGACY_BACKEND
6133 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6134 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6138 XX Get to the class and method info from the Execution Engine given XX
6139 XX tokens for the class and method XX
6141 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6142 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6146 /* These are the different addressing modes used to access a local var.
6147 * The JIT has to report the location of the locals back to the EE
6148 * for debugging purposes.
6154 VLT_REG_BYREF, // this type is currently only used for value types on X64
6157 VLT_STK_BYREF, // this type is currently only used for value types on X64
6171 siVarLocType vlType;
6174 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6176 // VLT_REG_BYREF -- the specified register contains the address of the variable
6184 // VLT_STK -- Any 32 bit value which is on the stack
6185 // eg. [ESP+0x20], or [EBP-0x28]
6186 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6187 // eg. mov EAX, [ESP+0x20]; [EAX]
6191 regNumber vlsBaseReg;
6192 NATIVE_OFFSET vlsOffset;
6195 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6204 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6205 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6213 regNumber vlrssBaseReg;
6214 NATIVE_OFFSET vlrssOffset;
6218 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6219 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6225 regNumber vlsrsBaseReg;
6226 NATIVE_OFFSET vlsrsOffset;
6232 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6233 // eg 2 DWords at [ESP+0x10]
6237 regNumber vls2BaseReg;
6238 NATIVE_OFFSET vls2Offset;
6241 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6242 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6249 // VLT_FIXED_VA -- fixed argument of a varargs function.
6250 // The argument location depends on the size of the variable
6251 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6252 // location of the first arg. This argument can then be accessed
6253 // relative to the position of the first arg
6257 unsigned vlfvOffset;
6264 void* rpValue; // pointer to the in-process
6265 // location of the value.
6271 bool vlIsInReg(regNumber reg);
6272 bool vlIsOnStk(regNumber reg, signed offset);
6275 /*************************************************************************/
6280 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6281 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6282 CORINFO_CALLINFO_FLAGS flags,
6283 CORINFO_CALL_INFO* pResult);
6284 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6286 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6287 CORINFO_ACCESS_FLAGS flags,
6288 CORINFO_FIELD_INFO* pResult);
6292 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6294 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6296 bool IsSuperPMIException(unsigned code)
6298 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6300 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6301 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6302 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6303 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6304 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6305 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6306 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6307 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6311 case EXCEPTIONCODE_DebugBreakorAV:
6312 case EXCEPTIONCODE_MC:
6313 case EXCEPTIONCODE_LWM:
6314 case EXCEPTIONCODE_SASM:
6315 case EXCEPTIONCODE_SSYM:
6316 case EXCEPTIONCODE_CALLUTILS:
6317 case EXCEPTIONCODE_TYPEUTILS:
6318 case EXCEPTIONCODE_ASSERT:
6325 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6326 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6328 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6329 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6332 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6333 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6334 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6336 // VOM info, method sigs
6338 void eeGetSig(unsigned sigTok,
6339 CORINFO_MODULE_HANDLE scope,
6340 CORINFO_CONTEXT_HANDLE context,
6341 CORINFO_SIG_INFO* retSig);
6343 void eeGetCallSiteSig(unsigned sigTok,
6344 CORINFO_MODULE_HANDLE scope,
6345 CORINFO_CONTEXT_HANDLE context,
6346 CORINFO_SIG_INFO* retSig);
6348 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6350 // Method entry-points, instrs
6352 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6354 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6356 CORINFO_EE_INFO eeInfo;
6357 bool eeInfoInitialized;
6359 CORINFO_EE_INFO* eeGetEEInfo();
6361 // Gets the offset of a SDArray's first element
6362 unsigned eeGetArrayDataOffset(var_types type);
6363 // Gets the offset of a MDArray's first element
6364 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6366 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6368 // Returns the page size for the target machine as reported by the EE.
6369 inline size_t eeGetPageSize()
6371 #if COR_JIT_EE_VERSION > 460
6372 return eeGetEEInfo()->osPageSize;
6373 #else // COR_JIT_EE_VERSION <= 460
6374 return CORINFO_PAGE_SIZE;
6375 #endif // COR_JIT_EE_VERSION > 460
6378 // Returns the frame size at which we will generate a loop to probe the stack.
6379 inline size_t getVeryLargeFrameSize()
6382 // The looping probe code is 40 bytes, whereas the straight-line probing for
6383 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6384 // or greater, to generate smaller code.
6385 return 2 * eeGetPageSize();
6387 return 3 * eeGetPageSize();
6391 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6393 #if COR_JIT_EE_VERSION > 460
6394 return eeGetEEInfo()->targetAbi == abi;
6396 return CORINFO_DESKTOP_ABI == abi;
6400 inline bool generateCFIUnwindCodes()
6402 #ifdef UNIX_AMD64_ABI
6403 return IsTargetAbi(CORINFO_CORERT_ABI);
6411 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6413 // Debugging support - Line number info
6415 void eeGetStmtOffsets();
6417 unsigned eeBoundariesCount;
6419 struct boundariesDsc
6421 UNATIVE_OFFSET nativeIP;
6423 unsigned sourceReason;
6424 } * eeBoundaries; // Boundaries to report to EE
6425 void eeSetLIcount(unsigned count);
6426 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6430 static void eeDispILOffs(IL_OFFSET offs);
6431 static void eeDispLineInfo(const boundariesDsc* line);
6432 void eeDispLineInfos();
6435 // Debugging support - Local var info
6439 unsigned eeVarsCount;
6441 struct VarResultInfo
6443 UNATIVE_OFFSET startOffset;
6444 UNATIVE_OFFSET endOffset;
6448 void eeSetLVcount(unsigned count);
6449 void eeSetLVinfo(unsigned which,
6450 UNATIVE_OFFSET startOffs,
6451 UNATIVE_OFFSET length,
6456 const siVarLoc& loc);
6460 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6461 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6464 // ICorJitInfo wrappers
6466 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6468 void eeAllocUnwindInfo(BYTE* pHotCode,
6474 CorJitFuncKind funcKind);
6476 void eeSetEHcount(unsigned cEH);
6478 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6480 WORD eeGetRelocTypeHint(void* target);
6482 // ICorStaticInfo wrapper functions
6484 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6486 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6488 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6491 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6492 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6493 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6494 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6496 template <typename ParamType>
6497 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6499 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6502 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6504 // Utility functions
6506 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6509 const wchar_t* eeGetCPString(size_t stringHandle);
6512 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6514 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6515 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6517 static fgWalkPreFn CountSharedStaticHelper;
6518 static bool IsSharedStaticHelper(GenTreePtr tree);
6519 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6521 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6522 // returns true/false if 'field' is a Jit Data offset
6523 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6524 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6525 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6527 /*****************************************************************************/
6532 enum TEMP_USAGE_TYPE
6538 static var_types tmpNormalizeType(var_types type);
6539 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6540 void tmpRlsTemp(TempDsc* temp);
6541 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6544 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6545 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6549 bool tmpAllFree() const;
6552 #ifndef LEGACY_BACKEND
6553 void tmpPreAllocateTemps(var_types type, unsigned count);
6554 #endif // !LEGACY_BACKEND
6557 #ifdef LEGACY_BACKEND
6558 unsigned tmpIntSpillMax; // number of int-sized spill temps
6559 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6560 #endif // LEGACY_BACKEND
6562 unsigned tmpCount; // Number of temps
6563 unsigned tmpSize; // Size of all the temps
6566 // Used by RegSet::rsSpillChk()
6567 unsigned tmpGetCount; // Temps which haven't been released yet
6570 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6572 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6573 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6576 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6577 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6581 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6582 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6586 CodeGenInterface* codeGen;
6588 // The following holds information about instr offsets in terms of generated code.
6592 IPmappingDsc* ipmdNext; // next line# record
6593 IL_OFFSETX ipmdILoffsx; // the instr offset
6594 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6595 bool ipmdIsLabel; // Can this code be a branch label?
6598 // Record the instr offset mapping to the generated code
6600 IPmappingDsc* genIPmappingList;
6601 IPmappingDsc* genIPmappingLast;
6603 // Managed RetVal - A side hash table meant to record the mapping from a
6604 // GT_CALL node to its IL offset. This info is used to emit sequence points
6605 // that can be used by debugger to determine the native offset at which the
6606 // managed RetVal will be available.
6608 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6609 // favor of a side table for two reasons: 1) We need IL offset for only those
6610 // GT_CALL nodes (created during importation) that correspond to an IL call and
6611 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6612 // structure and IL offset is needed only when generating debuggable code. Therefore
6613 // it is desirable to avoid memory size penalty in retail scenarios.
6614 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6615 CallSiteILOffsetTable;
6616 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6618 unsigned genReturnLocal; // Local number for the return value when applicable.
6619 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6621 // The following properties are part of CodeGenContext. Getters are provided here for
6622 // convenience and backward compatibility, but the properties can only be set by invoking
6623 // the setter on CodeGenContext directly.
6625 __declspec(property(get = getEmitter)) emitter* genEmitter;
6626 emitter* getEmitter()
6628 return codeGen->getEmitter();
6631 const bool isFramePointerUsed()
6633 return codeGen->isFramePointerUsed();
6636 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6637 bool getInterruptible()
6639 return codeGen->genInterruptible;
6641 void setInterruptible(bool value)
6643 codeGen->setInterruptible(value);
6647 const bool genDoubleAlign()
6649 return codeGen->doDoubleAlign();
6651 DWORD getCanDoubleAlign();
6652 bool shouldDoubleAlign(unsigned refCntStk,
6654 unsigned refCntWtdReg,
6655 unsigned refCntStkParam,
6656 unsigned refCntWtdStkDbl);
6657 #endif // DOUBLE_ALIGN
6659 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6660 bool getFullPtrRegMap()
6662 return codeGen->genFullPtrRegMap;
6664 void setFullPtrRegMap(bool value)
6666 codeGen->setFullPtrRegMap(value);
6669 // Things that MAY belong either in CodeGen or CodeGenContext
6671 #if FEATURE_EH_FUNCLETS
6672 FuncInfoDsc* compFuncInfos;
6673 unsigned short compCurrFuncIdx;
6674 unsigned short compFuncInfoCount;
6676 unsigned short compFuncCount()
6678 assert(fgFuncletsCreated);
6679 return compFuncInfoCount;
6682 #else // !FEATURE_EH_FUNCLETS
6684 // This is a no-op when there are no funclets!
6685 void genUpdateCurrentFunclet(BasicBlock* block)
6690 FuncInfoDsc compFuncInfoRoot;
6692 static const unsigned compCurrFuncIdx = 0;
6694 unsigned short compFuncCount()
6699 #endif // !FEATURE_EH_FUNCLETS
6701 FuncInfoDsc* funCurrentFunc();
6702 void funSetCurrentFunc(unsigned funcIdx);
6703 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6704 unsigned int funGetFuncIdx(BasicBlock* block);
6708 VARSET_TP compCurLife; // current live variables
6709 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6711 template <bool ForCodeGen>
6712 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6714 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6716 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6719 template <bool ForCodeGen>
6720 void compUpdateLife(GenTreePtr tree);
6722 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6723 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6724 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6725 template <bool ForCodeGen>
6726 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6728 template <bool ForCodeGen>
6729 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6731 // Gets a register mask that represent the kill set for a helper call since
6732 // not all JIT Helper calls follow the standard ABI on the target architecture.
6733 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6735 // Gets a register mask that represent the kill set for a NoGC helper call.
6736 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6739 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6740 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6741 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6742 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6743 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6744 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6745 #endif // _TARGET_ARM_
6747 // If "tree" is a indirection (GT_IND, or GT_OBJ) whose arg is an ADDR, whose arg is a LCL_VAR, return that LCL_VAR
6749 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6751 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6752 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6753 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6754 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6755 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6756 // for the tracked var indices of the field vars, as in a live var set).
6757 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6759 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6761 if (m_promotedStructDeathVars == nullptr)
6763 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6765 return m_promotedStructDeathVars;
6769 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6770 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6774 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6775 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6778 #if !defined(__GNUC__)
6779 #pragma region Unwind information
6784 // Infrastructure functions: start/stop/reserve/emit.
6787 void unwindBegProlog();
6788 void unwindEndProlog();
6789 void unwindBegEpilog();
6790 void unwindEndEpilog();
6791 void unwindReserve();
6792 void unwindEmit(void* pHotCode, void* pColdCode);
6795 // Specific unwind information functions: called by code generation to indicate a particular
6796 // prolog or epilog unwindable instruction has been generated.
6799 void unwindPush(regNumber reg);
6800 void unwindAllocStack(unsigned size);
6801 void unwindSetFrameReg(regNumber reg, unsigned offset);
6802 void unwindSaveReg(regNumber reg, unsigned offset);
6804 #if defined(_TARGET_ARM_)
6805 void unwindPushMaskInt(regMaskTP mask);
6806 void unwindPushMaskFloat(regMaskTP mask);
6807 void unwindPopMaskInt(regMaskTP mask);
6808 void unwindPopMaskFloat(regMaskTP mask);
6809 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6810 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6811 // called via unwindPadding().
6812 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6813 // instruction and the current location.
6814 #endif // _TARGET_ARM_
6816 #if defined(_TARGET_ARM64_)
6818 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6819 // instruction and the current location.
6820 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6821 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6822 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6823 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6824 void unwindSaveNext(); // unwind code: save_next
6825 void unwindReturn(regNumber reg); // ret lr
6826 #endif // defined(_TARGET_ARM64_)
6829 // Private "helper" functions for the unwind implementation.
6833 #if FEATURE_EH_FUNCLETS
6834 void unwindGetFuncLocations(FuncInfoDsc* func,
6835 bool getHotSectionData,
6836 /* OUT */ emitLocation** ppStartLoc,
6837 /* OUT */ emitLocation** ppEndLoc);
6838 #endif // FEATURE_EH_FUNCLETS
6840 void unwindReserveFunc(FuncInfoDsc* func);
6841 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6843 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6845 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6846 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6848 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6850 #if defined(_TARGET_AMD64_)
6852 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6854 void unwindBegPrologWindows();
6855 void unwindPushWindows(regNumber reg);
6856 void unwindAllocStackWindows(unsigned size);
6857 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6858 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6860 #ifdef UNIX_AMD64_ABI
6861 void unwindBegPrologCFI();
6862 void unwindPushCFI(regNumber reg);
6863 void unwindAllocStackCFI(unsigned size);
6864 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6865 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6866 int mapRegNumToDwarfReg(regNumber reg);
6867 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6868 #endif // UNIX_AMD64_ABI
6869 #elif defined(_TARGET_ARM_)
6871 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6872 void unwindPushPopMaskFloat(regMaskTP mask);
6873 void unwindSplit(FuncInfoDsc* func);
6875 #endif // _TARGET_ARM_
6877 #if !defined(__GNUC__)
6878 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6882 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6883 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6887 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6888 XX that contains the distinguished, well-known SIMD type definitions). XX
6890 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6891 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6894 // Get highest available instruction set for floating point codegen
6895 InstructionSet getFloatingPointInstructionSet()
6897 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6900 return InstructionSet_AVX;
6905 return InstructionSet_SSE3_4;
6909 assert(canUseSSE2());
6910 return InstructionSet_SSE2;
6912 assert(!"getFPInstructionSet() is not implemented for target arch");
6914 return InstructionSet_NONE;
6918 // Get highest available instruction set for SIMD codegen
6919 InstructionSet getSIMDInstructionSet()
6921 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6922 return getFloatingPointInstructionSet();
6924 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6926 return InstructionSet_NONE;
6932 // Should we support SIMD intrinsics?
6935 // Have we identified any SIMD types?
6936 // This is currently used by struct promotion to avoid getting type information for a struct
6937 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6939 bool _usesSIMDTypes;
6940 bool usesSIMDTypes()
6942 return _usesSIMDTypes;
6944 void setUsesSIMDTypes(bool value)
6946 _usesSIMDTypes = value;
6949 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6950 // that require indexed access to the individual fields of the vector, which is not well supported
6951 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6952 unsigned lvaSIMDInitTempVarNum;
6955 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6956 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6957 CORINFO_CLASS_HANDLE SIMDIntHandle;
6958 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6959 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6960 CORINFO_CLASS_HANDLE SIMDShortHandle;
6961 CORINFO_CLASS_HANDLE SIMDByteHandle;
6962 CORINFO_CLASS_HANDLE SIMDLongHandle;
6963 CORINFO_CLASS_HANDLE SIMDUIntHandle;
6964 CORINFO_CLASS_HANDLE SIMDULongHandle;
6965 CORINFO_CLASS_HANDLE SIMDVector2Handle;
6966 CORINFO_CLASS_HANDLE SIMDVector3Handle;
6967 CORINFO_CLASS_HANDLE SIMDVector4Handle;
6968 CORINFO_CLASS_HANDLE SIMDVectorHandle;
6970 // Get the handle for a SIMD type.
6971 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
6973 if (simdBaseType == TYP_FLOAT)
6978 return SIMDVector2Handle;
6980 return SIMDVector3Handle;
6982 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
6984 return SIMDVector4Handle;
6993 assert(simdType == getSIMDVectorType());
6994 switch (simdBaseType)
6997 return SIMDFloatHandle;
6999 return SIMDDoubleHandle;
7001 return SIMDIntHandle;
7003 return SIMDUShortHandle;
7005 return SIMDUShortHandle;
7007 return SIMDUByteHandle;
7009 return SIMDShortHandle;
7011 return SIMDByteHandle;
7013 return SIMDLongHandle;
7015 return SIMDUIntHandle;
7017 return SIMDULongHandle;
7019 assert(!"Didn't find a class handle for simdType");
7021 return NO_CLASS_HANDLE;
7025 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7026 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7027 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7029 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7030 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7031 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7032 bool isSIMDTypeLocal(GenTree* tree)
7034 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7037 // Returns true if the type of the tree is a byref of TYP_SIMD
7038 bool isAddrOfSIMDType(GenTree* tree)
7040 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7042 switch (tree->OperGet())
7045 return varTypeIsSIMD(tree->gtGetOp1());
7047 case GT_LCL_VAR_ADDR:
7048 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7051 return isSIMDTypeLocal(tree);
7058 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7060 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7061 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7062 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7065 // Returns base type of a TYP_SIMD local.
7066 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7067 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7069 if (isSIMDTypeLocal(tree))
7071 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7077 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7079 return info.compCompHnd->isInSIMDModule(clsHnd);
7082 bool isSIMDClass(typeInfo* pTypeInfo)
7084 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7087 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7088 // if it is not a SIMD type or is an unsupported base type.
7089 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7091 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7093 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7096 // Get SIMD Intrinsic info given the method handle.
7097 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7098 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7099 CORINFO_METHOD_HANDLE methodHnd,
7100 CORINFO_SIG_INFO* sig,
7103 var_types* baseType,
7104 unsigned* sizeBytes);
7106 // Pops and returns GenTree node from importers type stack.
7107 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7108 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7110 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7111 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7113 // Creates a GT_SIMD tree for Select operation
7114 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7116 unsigned simdVectorSize,
7121 // Creates a GT_SIMD tree for Min/Max operation
7122 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7123 CORINFO_CLASS_HANDLE typeHnd,
7125 unsigned simdVectorSize,
7129 // Transforms operands and returns the SIMD intrinsic to be applied on
7130 // transformed operands to obtain given relop result.
7131 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7132 CORINFO_CLASS_HANDLE typeHnd,
7133 unsigned simdVectorSize,
7134 var_types* baseType,
7138 // Creates a GT_SIMD tree for Abs intrinsic.
7139 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7141 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7142 // Transforms operands and returns the SIMD intrinsic to be applied on
7143 // transformed operands to obtain == comparison result.
7144 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7145 unsigned simdVectorSize,
7149 // Transforms operands and returns the SIMD intrinsic to be applied on
7150 // transformed operands to obtain > comparison result.
7151 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7152 unsigned simdVectorSize,
7156 // Transforms operands and returns the SIMD intrinsic to be applied on
7157 // transformed operands to obtain >= comparison result.
7158 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7159 unsigned simdVectorSize,
7163 // Transforms operands and returns the SIMD intrinsic to be applied on
7164 // transformed operands to obtain >= comparison result in case of int32
7165 // and small int base type vectors.
7166 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7167 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7168 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7170 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7171 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7172 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7173 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7174 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7176 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7177 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7178 GenTreePtr newobjThis,
7179 CORINFO_CLASS_HANDLE clsHnd,
7180 CORINFO_METHOD_HANDLE method,
7181 CORINFO_SIG_INFO* sig,
7184 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7186 // Whether SIMD vector occupies part of SIMD register.
7187 // SSE2: vector2f/3f are considered sub register SIMD types.
7188 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7189 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7191 unsigned sizeBytes = 0;
7192 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7193 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7196 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7198 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7201 // Get the type for the hardware SIMD vector.
7202 // This is the maximum SIMD type supported for this target.
7203 var_types getSIMDVectorType()
7205 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7212 assert(canUseSSE2());
7216 assert(!"getSIMDVectorType() unimplemented on target arch");
7221 // Get the size of the SIMD type in bytes
7222 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7224 unsigned sizeBytes = 0;
7225 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7229 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7230 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7232 // Get the the number of elements of basetype of SIMD vector given by its type handle
7233 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7235 // Get preferred alignment of SIMD type.
7236 int getSIMDTypeAlignment(var_types simdType);
7238 // Get the number of bytes in a SIMD Vector for the current compilation.
7239 unsigned getSIMDVectorRegisterByteLength()
7241 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7244 return YMM_REGSIZE_BYTES;
7248 assert(canUseSSE2());
7249 return XMM_REGSIZE_BYTES;
7252 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7257 // The minimum and maximum possible number of bytes in a SIMD vector.
7258 unsigned int maxSIMDStructBytes()
7260 return getSIMDVectorRegisterByteLength();
7262 unsigned int minSIMDStructBytes()
7264 return emitTypeSize(TYP_SIMD8);
7267 #ifdef FEATURE_AVX_SUPPORT
7268 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7269 static const unsigned maxPossibleSIMDStructBytes = 32;
7270 #else // !FEATURE_AVX_SUPPORT
7271 static const unsigned maxPossibleSIMDStructBytes = 16;
7272 #endif // !FEATURE_AVX_SUPPORT
7274 // Returns the codegen type for a given SIMD size.
7275 var_types getSIMDTypeForSize(unsigned size)
7277 var_types simdType = TYP_UNDEF;
7280 simdType = TYP_SIMD8;
7282 else if (size == 12)
7284 simdType = TYP_SIMD12;
7286 else if (size == 16)
7288 simdType = TYP_SIMD16;
7290 #ifdef FEATURE_AVX_SUPPORT
7291 else if (size == 32)
7293 simdType = TYP_SIMD32;
7295 #endif // FEATURE_AVX_SUPPORT
7298 noway_assert(!"Unexpected size for SIMD type");
7303 unsigned getSIMDInitTempVarNum()
7305 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7307 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7308 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7310 return lvaSIMDInitTempVarNum;
7313 #endif // FEATURE_SIMD
7316 //------------------------------------------------------------------------
7317 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7319 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7320 // candidate for enregistration.
7322 unsigned largestEnregisterableStructSize()
7325 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7326 if (vectorRegSize > TARGET_POINTER_SIZE)
7328 return vectorRegSize;
7331 #endif // FEATURE_SIMD
7333 return TARGET_POINTER_SIZE;
7338 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7339 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7340 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7342 // Is this var is of type simd struct?
7343 bool lclVarIsSIMDType(unsigned varNum)
7345 LclVarDsc* varDsc = lvaTable + varNum;
7346 return varDsc->lvIsSIMDType();
7349 // Is this Local node a SIMD local?
7350 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7352 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7355 // Returns true if the TYP_SIMD locals on stack are aligned at their
7356 // preferred byte boundary specified by getSIMDTypeAlignment().
7358 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7359 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7360 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7361 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7362 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7363 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7364 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7367 bool isSIMDTypeLocalAligned(unsigned varNum)
7369 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7370 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7373 int off = lvaFrameAddress(varNum, &ebpBased);
7374 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7375 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7376 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7379 #endif // FEATURE_SIMD
7384 // Whether SSE2 is available
7385 bool canUseSSE2() const
7387 #ifdef _TARGET_XARCH_
7388 return opts.compCanUseSSE2;
7394 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7395 bool CanUseSSE3_4() const
7397 #ifdef _TARGET_XARCH_
7398 return opts.compCanUseSSE3_4;
7404 bool canUseAVX() const
7406 #ifdef FEATURE_AVX_SUPPORT
7407 return opts.compCanUseAVX;
7414 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7419 XX Generic info about the compilation and the method being compiled. XX
7420 XX It is responsible for driving the other phases. XX
7421 XX It is also responsible for all the memory management. XX
7423 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7424 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7428 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7430 InlineResult* compInlineResult; // The result of importing the inlinee method.
7432 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7433 bool compJmpOpUsed; // Does the method do a JMP
7434 bool compLongUsed; // Does the method use TYP_LONG
7435 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7436 bool compTailCallUsed; // Does the method do a tailcall
7437 bool compLocallocUsed; // Does the method use localloc.
7438 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7439 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7440 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7442 // NOTE: These values are only reliable after
7443 // the importing is completely finished.
7445 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7446 // we can iterate over these efficiently.
7448 #if CPU_USES_BLOCK_MOVE
7449 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7453 // State information - which phases have completed?
7454 // These are kept together for easy discoverability
7456 bool bRangeAllowStress;
7457 bool compCodeGenDone;
7458 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7459 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7460 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7461 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7464 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7465 bool fgLocalVarLivenessChanged;
7467 bool compStackProbePrologDone;
7469 #ifndef LEGACY_BACKEND
7471 #endif // !LEGACY_BACKEND
7472 bool compRationalIRForm;
7474 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7476 bool compGeneratingProlog;
7477 bool compGeneratingEpilog;
7478 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7479 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7480 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7481 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7482 bool getNeedsGSSecurityCookie() const
7484 return compNeedsGSSecurityCookie;
7486 void setNeedsGSSecurityCookie()
7488 compNeedsGSSecurityCookie = true;
7491 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7492 // frame layout calculations, this is the level we are currently
7495 //---------------------------- JITing options -----------------------------
7508 JitFlags* jitFlags; // all flags passed from the EE
7509 unsigned compFlags; // method attributes
7511 codeOptimize compCodeOpt; // what type of code optimizations
7515 #ifdef _TARGET_XARCH_
7516 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7517 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7519 #ifdef FEATURE_AVX_SUPPORT
7520 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7521 #endif // FEATURE_AVX_SUPPORT
7522 #endif // _TARGET_XARCH_
7524 // optimize maximally and/or favor speed over size?
7526 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7527 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7528 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7529 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7530 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7532 // Maximun number of locals before turning off the inlining
7533 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7536 unsigned instrCount;
7537 unsigned lvRefCount;
7538 bool compMinOptsIsSet;
7540 bool compMinOptsIsUsed;
7542 inline bool MinOpts()
7544 assert(compMinOptsIsSet);
7545 compMinOptsIsUsed = true;
7548 inline bool IsMinOptsSet()
7550 return compMinOptsIsSet;
7553 inline bool MinOpts()
7557 inline bool IsMinOptsSet()
7559 return compMinOptsIsSet;
7562 inline void SetMinOpts(bool val)
7564 assert(!compMinOptsIsUsed);
7565 assert(!compMinOptsIsSet || (compMinOpts == val));
7567 compMinOptsIsSet = true;
7570 // true if the CLFLG_* for an optimization is set.
7571 inline bool OptEnabled(unsigned optFlag)
7573 return !!(compFlags & optFlag);
7576 #ifdef FEATURE_READYTORUN_COMPILER
7577 inline bool IsReadyToRun()
7579 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7582 inline bool IsReadyToRun()
7588 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7589 // PInvoke transitions inline (e.g. when targeting CoreRT).
7590 inline bool ShouldUsePInvokeHelpers()
7592 #if COR_JIT_EE_VERSION > 460
7593 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7599 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7601 inline bool IsReversePInvoke()
7603 #if COR_JIT_EE_VERSION > 460
7604 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7610 // true if we must generate code compatible with JIT32 quirks
7611 inline bool IsJit32Compat()
7613 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7614 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7620 // true if we must generate code compatible with Jit64 quirks
7621 inline bool IsJit64Compat()
7623 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7624 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7625 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7632 bool compScopeInfo; // Generate the LocalVar info ?
7633 bool compDbgCode; // Generate debugger-friendly code?
7634 bool compDbgInfo; // Gather debugging info?
7637 #ifdef PROFILING_SUPPORTED
7638 bool compNoPInvokeInlineCB;
7640 static const bool compNoPInvokeInlineCB;
7644 bool compGcChecks; // Check arguments and return values to ensure they are sane
7645 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7646 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7650 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7651 // to be allocated on the stack.
7652 // It will be set to true in the following cases:
7653 // 1. When the method being compiled has a declarative security
7654 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7655 // This is also the case when we inject a prolog and epilog in the method.
7657 // 2. When the method being compiled has imperative security (i.e. the method
7658 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7660 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7662 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7663 // which gets reported as a GC root to stackwalker.
7664 // (See also ICodeManager::GetAddrOfSecurityObject.)
7671 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7672 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7676 #ifdef UNIX_AMD64_ABI
7677 // This flag is indicating if there is a need to align the frame.
7678 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7679 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7680 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7681 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7682 // there are calls and making sure the frame alignment logic is executed.
7683 bool compNeedToAlignFrame;
7684 #endif // UNIX_AMD64_ABI
7686 bool compProcedureSplitting; // Separate cold code from hot code
7688 bool genFPorder; // Preserve FP order (operations are non-commutative)
7689 bool genFPopt; // Can we do frame-pointer-omission optimization?
7690 bool altJit; // True if we are an altjit and are compiling this method
7693 bool optRepeat; // Repeat optimizer phases k times
7694 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7695 bool dspCode; // Display native code generated
7696 bool dspEHTable; // Display the EH table reported to the VM
7697 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7698 bool dspEmit; // Display emitter output
7699 bool dspLines; // Display source-code lines intermixed with native code output
7700 bool dmpHex; // Display raw bytes in hex of native code output
7701 bool varNames; // Display variables names in native code output
7702 bool disAsm; // Display native code as it is generated
7703 bool disAsmSpilled; // Display native code when any register spilling occurs
7704 bool disDiffable; // Makes the Disassembly code 'diff-able'
7705 bool disAsm2; // Display native code after it is generated using external disassembler
7706 bool dspOrder; // Display names of each of the methods that we ngen/jit
7707 bool dspUnwind; // Display the unwind info output
7708 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7709 bool compLongAddress; // Force using large pseudo instructions for long address
7710 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7711 bool dspGCtbls; // Display the GC tables
7715 bool doLateDisasm; // Run the late disassembler
7716 #endif // LATE_DISASM
7718 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7719 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7720 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7721 static const bool dspGCtbls = true;
7724 // We need stack probes to guarantee that we won't trigger a stack overflow
7725 // when calling unmanaged code until they get a chance to set up a frame, because
7726 // the EE will have no idea where it is.
7728 // We will only be doing this currently for hosted environments. Unfortunately
7729 // we need to take care of stubs, so potentially, we will have to do the probes
7730 // for any call. We have a plan for not needing for stubs though
7731 bool compNeedStackProbes;
7733 #ifdef PROFILING_SUPPORTED
7734 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7735 // This option helps make the JIT behave as if it is running under a profiler.
7736 bool compJitELTHookEnabled;
7737 #endif // PROFILING_SUPPORTED
7739 #if FEATURE_TAILCALL_OPT
7740 // Whether opportunistic or implicit tail call optimization is enabled.
7741 bool compTailCallOpt;
7742 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7743 bool compTailCallLoopOpt;
7747 static const bool compUseSoftFP = true;
7748 #else // !ARM_SOFTFP
7749 static const bool compUseSoftFP = false;
7752 GCPollType compGCPollType;
7756 static bool s_pAltJitExcludeAssembliesListInitialized;
7757 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7762 template <typename T>
7765 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7768 template <typename T>
7771 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7774 static int dspTreeID(GenTree* tree)
7776 return tree->gtTreeID;
7778 static void printTreeID(GenTree* tree)
7780 if (tree == nullptr)
7786 printf("[%06d]", dspTreeID(tree));
7793 #define STRESS_MODES \
7797 /* "Variations" stress areas which we try to mix up with each other. */ \
7798 /* These should not be exhaustively used as they might */ \
7799 /* hide/trivialize other areas */ \
7801 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7802 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7803 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7804 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7805 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7806 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7807 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7808 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7809 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7810 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7811 STRESS_MODE(NULL_OBJECT_CHECK) \
7812 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7813 STRESS_MODE(RANDOM_INLINE) \
7815 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7817 /* "Check" stress areas that can be exhaustively used if we */ \
7818 /* dont care about performance at all */ \
7820 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7821 STRESS_MODE(CHK_FLOW_UPDATE) \
7822 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7824 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7828 #define STRESS_MODE(mode) STRESS_##mode,
7835 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7836 BYTE compActiveStressModes[STRESS_COUNT];
7839 #define MAX_STRESS_WEIGHT 100
7841 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7845 bool compInlineStress()
7847 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7850 bool compRandomInlineStress()
7852 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7857 bool compTailCallStress()
7860 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7866 codeOptimize compCodeOpt()
7869 // Switching between size & speed has measurable throughput impact
7870 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7871 // DEBUG, but should generate identical code between CHK & RET builds,
7872 // so that's not acceptable.
7873 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7874 // Investigate the cause of the throughput regression.
7876 return opts.compCodeOpt;
7878 return BLENDED_CODE;
7882 //--------------------- Info about the procedure --------------------------
7886 COMP_HANDLE compCompHnd;
7887 CORINFO_MODULE_HANDLE compScopeHnd;
7888 CORINFO_CLASS_HANDLE compClassHnd;
7889 CORINFO_METHOD_HANDLE compMethodHnd;
7890 CORINFO_METHOD_INFO* compMethodInfo;
7892 BOOL hasCircularClassConstraints;
7893 BOOL hasCircularMethodConstraints;
7895 #if defined(DEBUG) || defined(LATE_DISASM)
7896 const char* compMethodName;
7897 const char* compClassName;
7898 const char* compFullName;
7899 #endif // defined(DEBUG) || defined(LATE_DISASM)
7901 #if defined(DEBUG) || defined(INLINE_DATA)
7902 // Method hash is logcally const, but computed
7904 mutable unsigned compMethodHashPrivate;
7905 unsigned compMethodHash() const;
7906 #endif // defined(DEBUG) || defined(INLINE_DATA)
7908 #ifdef PSEUDORANDOM_NOP_INSERTION
7909 // things for pseudorandom nop insertion
7910 unsigned compChecksum;
7914 // The following holds the FLG_xxxx flags for the method we're compiling.
7917 // The following holds the class attributes for the method we're compiling.
7918 unsigned compClassAttr;
7920 const BYTE* compCode;
7921 IL_OFFSET compILCodeSize; // The IL code size
7922 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7923 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7924 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7925 // (2) the code is hot/cold split, and we issued less code than we expected
7926 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7928 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7929 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7930 bool compIsContextful : 1; // contextful method
7931 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7932 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7933 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7934 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7935 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7937 var_types compRetType; // Return type of the method as declared in IL
7938 var_types compRetNativeType; // Normalized return type as per target arch ABI
7939 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7940 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7941 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7942 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7943 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7944 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7945 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7946 unsigned compMaxStack;
7947 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7948 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7950 unsigned compCallUnmanaged; // count of unmanaged calls
7951 unsigned compLvFrameListRoot; // lclNum for the Frame root
7952 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7953 // You should generally use compHndBBtabCount instead: it is the
7954 // current number of EH clauses (after additions like synchronized
7955 // methods and funclets, and removals like unreachable code deletion).
7957 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7958 // and the VM expects that, or the JIT is a "self-host" compiler
7959 // (e.g., x86 hosted targeting x86) and the VM expects that.
7961 /* The following holds IL scope information about local variables.
7964 unsigned compVarScopesCount;
7965 VarScopeDsc* compVarScopes;
7967 /* The following holds information about instr offsets for
7968 * which we need to report IP-mappings
7971 IL_OFFSET* compStmtOffsets; // sorted
7972 unsigned compStmtOffsetsCount;
7973 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
7975 #define CPU_X86 0x0100 // The generic X86 CPU
7976 #define CPU_X86_PENTIUM_4 0x0110
7978 #define CPU_X64 0x0200 // The generic x64 CPU
7979 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
7980 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
7982 #define CPU_ARM 0x0300 // The generic ARM CPU
7984 unsigned genCPU; // What CPU are we running on
7987 // Returns true if the method being compiled returns a non-void and non-struct value.
7988 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
7989 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
7990 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
7991 // Methods returning such structs are considered to return non-struct return value and
7992 // this method returns true in that case.
7993 bool compMethodReturnsNativeScalarType()
7995 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
7998 // Returns true if the method being compiled returns RetBuf addr as its return value
7999 bool compMethodReturnsRetBufAddr()
8001 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8002 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8004 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8005 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8006 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8007 // methods with hidden RetBufArg.
8009 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8010 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8011 // returning the address of RetBuf.
8013 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8014 // to be returned in RAX.
8015 CLANG_FORMAT_COMMENT_ANCHOR;
8017 #ifdef _TARGET_AMD64_
8018 return (info.compRetBuffArg != BAD_VAR_NUM);
8019 #else // !_TARGET_AMD64_
8020 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8021 #endif // !_TARGET_AMD64_
8024 // Returns true if the method returns a value in more than one return register
8025 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8026 // TODO-ARM64: Does this apply for ARM64 too?
8027 bool compMethodReturnsMultiRegRetType()
8029 #if FEATURE_MULTIREG_RET
8030 #if defined(_TARGET_X86_)
8031 // On x86 only 64-bit longs are returned in multiple registers
8032 return varTypeIsLong(info.compRetNativeType);
8033 #else // targets: X64-UNIX, ARM64 or ARM32
8034 // On all other targets that support multireg return values:
8035 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8036 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8037 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8038 #endif // TARGET_XXX
8040 #else // not FEATURE_MULTIREG_RET
8042 // For this architecture there are no multireg returns
8045 #endif // FEATURE_MULTIREG_RET
8048 #if FEATURE_MULTIREG_ARGS
8049 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8050 // return the gcPtr layout for the pointers sized fields
8051 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8052 #endif // FEATURE_MULTIREG_ARGS
8054 // Returns true if the method being compiled returns a value
8055 bool compMethodHasRetVal()
8057 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8058 compMethodReturnsMultiRegRetType();
8063 void compDispLocalVars();
8067 //-------------------------- Global Compiler Data ------------------------------------
8070 static unsigned s_compMethodsCount; // to produce unique label names
8071 unsigned compGenTreeID;
8074 BasicBlock* compCurBB; // the current basic block in process
8075 GenTreePtr compCurStmt; // the current statement in process
8077 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8080 // The following is used to create the 'method JIT info' block.
8081 size_t compInfoBlkSize;
8082 BYTE* compInfoBlkAddr;
8084 EHblkDsc* compHndBBtab; // array of EH data
8085 unsigned compHndBBtabCount; // element count of used elements in EH data array
8086 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8088 #if defined(_TARGET_X86_)
8090 //-------------------------------------------------------------------------
8091 // Tracking of region covered by the monitor in synchronized methods
8092 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8093 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8095 #endif // !_TARGET_X86_
8097 Phases previousCompletedPhase; // the most recently completed phase
8099 //-------------------------------------------------------------------------
8100 // The following keeps track of how many bytes of local frame space we've
8101 // grabbed so far in the current function, and how many argument bytes we
8102 // need to pop when we return.
8105 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8107 // Count of callee-saved regs we pushed in the prolog.
8108 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8109 // In case of Amd64 this doesn't include float regs saved on stack.
8110 unsigned compCalleeRegsPushed;
8112 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8113 // Mask of callee saved float regs on stack.
8114 regMaskTP compCalleeFPRegsSavedMask;
8116 #ifdef _TARGET_AMD64_
8117 // Quirk for VS debug-launch scenario to work:
8118 // Bytes of padding between save-reg area and locals.
8119 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8120 unsigned compVSQuirkStackPaddingNeeded;
8121 bool compQuirkForPPPflag;
8124 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8126 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8127 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8128 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8130 //-------------------------------------------------------------------------
8132 static void compStartup(); // One-time initialization
8133 static void compShutdown(); // One-time finalization
8135 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8138 static void compDisplayStaticSizes(FILE* fout);
8140 //------------ Some utility functions --------------
8142 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8143 void** ppIndirection); /* OUT */
8145 // Several JIT/EE interface functions return a CorInfoType, and also return a
8146 // class handle as an out parameter if the type is a value class. Returns the
8147 // size of the type these describe.
8148 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8151 // Components used by the compiler may write unit test suites, and
8152 // have them run within this method. They will be run only once per process, and only
8153 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8154 // These should fail by asserting.
8155 void compDoComponentUnitTestsOnce();
8158 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8159 CORINFO_MODULE_HANDLE classPtr,
8160 COMP_HANDLE compHnd,
8161 CORINFO_METHOD_INFO* methodInfo,
8162 void** methodCodePtr,
8163 ULONG* methodCodeSize,
8164 JitFlags* compileFlags);
8165 void compCompileFinish();
8166 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8167 COMP_HANDLE compHnd,
8168 CORINFO_METHOD_INFO* methodInfo,
8169 void** methodCodePtr,
8170 ULONG* methodCodeSize,
8171 JitFlags* compileFlags,
8172 CorInfoInstantiationVerification instVerInfo);
8174 ArenaAllocator* compGetAllocator();
8176 #if MEASURE_MEM_ALLOC
8178 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8182 unsigned allocCnt; // # of allocs
8183 UINT64 allocSz; // total size of those alloc.
8184 UINT64 allocSzMax; // Maximum single allocation.
8185 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8186 UINT64 nraTotalSizeAlloc;
8187 UINT64 nraTotalSizeUsed;
8189 static const char* s_CompMemKindNames[]; // Names of the kinds.
8191 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8193 for (int i = 0; i < CMK_Count; i++)
8195 allocSzByKind[i] = 0;
8198 MemStats(const MemStats& ms)
8199 : allocCnt(ms.allocCnt)
8200 , allocSz(ms.allocSz)
8201 , allocSzMax(ms.allocSzMax)
8202 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8203 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8205 for (int i = 0; i < CMK_Count; i++)
8207 allocSzByKind[i] = ms.allocSzByKind[i];
8211 // Until we have ubiquitous constructors.
8214 this->MemStats::MemStats();
8217 void AddAlloc(size_t sz, CompMemKind cmk)
8221 if (sz > allocSzMax)
8225 allocSzByKind[cmk] += sz;
8228 void Print(FILE* f); // Print these stats to f.
8229 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8231 MemStats genMemStats;
8233 struct AggregateMemStats : public MemStats
8237 AggregateMemStats() : MemStats(), nMethods(0)
8241 void Add(const MemStats& ms)
8244 allocCnt += ms.allocCnt;
8245 allocSz += ms.allocSz;
8246 allocSzMax = max(allocSzMax, ms.allocSzMax);
8247 for (int i = 0; i < CMK_Count; i++)
8249 allocSzByKind[i] += ms.allocSzByKind[i];
8251 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8252 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8255 void Print(FILE* f); // Print these stats to jitstdout.
8258 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8259 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8260 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8262 #endif // MEASURE_MEM_ALLOC
8264 #if LOOP_HOIST_STATS
8265 unsigned m_loopsConsidered;
8266 bool m_curLoopHasHoistedExpression;
8267 unsigned m_loopsWithHoistedExpressions;
8268 unsigned m_totalHoistedExpressions;
8270 void AddLoopHoistStats();
8271 void PrintPerMethodLoopHoistStats();
8273 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8274 static unsigned s_loopsConsidered;
8275 static unsigned s_loopsWithHoistedExpressions;
8276 static unsigned s_totalHoistedExpressions;
8278 static void PrintAggregateLoopHoistStats(FILE* f);
8279 #endif // LOOP_HOIST_STATS
8281 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8282 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8283 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8284 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8285 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8286 void compFreeMem(void*);
8288 bool compIsForImportOnly();
8289 bool compIsForInlining();
8290 bool compDonotInline();
8293 const char* compLocalVarName(unsigned varNum, unsigned offs);
8294 VarName compVarName(regNumber reg, bool isFloatReg = false);
8295 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8296 const char* compRegPairName(regPairNo regPair);
8297 const char* compRegNameForSize(regNumber reg, size_t size);
8298 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8299 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8300 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8303 //-------------------------------------------------------------------------
8305 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8307 struct VarScopeMapInfo
8309 VarScopeListNode* head;
8310 VarScopeListNode* tail;
8311 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8313 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8320 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8321 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8323 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8324 VarNumToScopeDscMap;
8326 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8327 VarNumToScopeDscMap* compVarScopeMap;
8329 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8331 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8333 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8335 void compInitVarScopeMap();
8337 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8338 // enter scope, sorted by instr offset
8339 unsigned compNextEnterScope;
8341 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8342 // go out of scope, sorted by instr offset
8343 unsigned compNextExitScope;
8345 void compInitScopeLists();
8347 void compResetScopeLists();
8349 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8351 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8353 void compProcessScopesUntil(unsigned offset,
8355 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8356 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8359 void compDispScopeLists();
8362 bool compIsProfilerHookNeeded();
8364 //-------------------------------------------------------------------------
8365 /* Statistical Data Gathering */
8367 void compJitStats(); // call this function and enable
8368 // various ifdef's below for statistical data
8371 void compCallArgStats();
8372 static void compDispCallArgStats(FILE* fout);
8375 //-------------------------------------------------------------------------
8382 ArenaAllocator* compAllocator;
8385 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8386 // suitable for use by utilcode collection types.
8387 IAllocator* compAsIAllocator;
8389 #if MEASURE_MEM_ALLOC
8390 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8391 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8392 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8394 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8396 #endif // MEASURE_MEM_ALLOC
8398 void compFunctionTraceStart();
8399 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8402 size_t compMaxUncheckedOffsetForNullObject;
8404 void compInitOptions(JitFlags* compileFlags);
8406 void compSetProcessor();
8407 void compInitDebuggingInfo();
8408 void compSetOptimizationLevel();
8409 #ifdef _TARGET_ARMARCH_
8410 bool compRsvdRegCheck(FrameLayoutState curState);
8412 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8414 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8415 void ResetOptAnnotations();
8417 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8418 void RecomputeLoopInfo();
8420 #ifdef PROFILING_SUPPORTED
8421 // Data required for generating profiler Enter/Leave/TailCall hooks
8423 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8424 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8425 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8428 #ifdef _TARGET_AMD64_
8429 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8432 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8433 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8435 IAllocator* getAllocator()
8437 return compAsIAllocator;
8440 #if MEASURE_MEM_ALLOC
8441 IAllocator* getAllocatorBitset()
8443 return compAsIAllocatorBitset;
8445 IAllocator* getAllocatorGC()
8447 return compAsIAllocatorGC;
8449 IAllocator* getAllocatorLoopHoist()
8451 return compAsIAllocatorLoopHoist;
8453 #else // !MEASURE_MEM_ALLOC
8454 IAllocator* getAllocatorBitset()
8456 return compAsIAllocator;
8458 IAllocator* getAllocatorGC()
8460 return compAsIAllocator;
8462 IAllocator* getAllocatorLoopHoist()
8464 return compAsIAllocator;
8466 #endif // !MEASURE_MEM_ALLOC
8469 IAllocator* getAllocatorDebugOnly()
8471 #if MEASURE_MEM_ALLOC
8472 return compAsIAllocatorDebugOnly;
8473 #else // !MEASURE_MEM_ALLOC
8474 return compAsIAllocator;
8475 #endif // !MEASURE_MEM_ALLOC
8480 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8481 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8485 XX Checks for type compatibility and merges types XX
8487 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8488 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8492 // Set to TRUE if verification cannot be skipped for this method
8493 // If we detect unverifiable code, we will lazily check
8494 // canSkipMethodVerification() to see if verification is REALLY needed.
8495 BOOL tiVerificationNeeded;
8497 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8498 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8499 BOOL tiIsVerifiableCode;
8501 // Set to TRUE if runtime callout is needed for this method
8502 BOOL tiRuntimeCalloutNeeded;
8504 // Set to TRUE if security prolog/epilog callout is needed for this method
8505 // Note: This flag is different than compNeedSecurityCheck.
8506 // compNeedSecurityCheck means whether or not a security object needs
8507 // to be allocated on the stack, which is currently true for EnC as well.
8508 // tiSecurityCalloutNeeded means whether or not security callouts need
8509 // to be inserted in the jitted code.
8510 BOOL tiSecurityCalloutNeeded;
8512 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8513 // This support is necessary to suport attributes that are not described in
8514 // for example, signatures. For example, the permanent home byref (byref that
8515 // points to the gc heap), isn't a property of method signatures, therefore,
8516 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8517 // but when deciding if we need to reimport a block, we need to take these
8519 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8521 // Returns TRUE if child is equal to or a subtype of parent.
8522 // normalisedForStack indicates that both types are normalised for the stack
8523 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8525 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8526 // *pDest is modified to represent the merged type. Sets "*changed" to true
8527 // if this changes "*pDest".
8528 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8530 // Set pDest from the primitive value type.
8531 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8533 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8536 // <BUGNUM> VSW 471305
8537 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8538 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8539 // We use a "short" as we need to push/pop this scope.
8541 short compRegSetCheckLevel;
8545 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8546 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8548 XX IL verification stuff XX
8551 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8556 // The following is used to track liveness of local variables, initialization
8557 // of valueclass constructors, and type safe use of IL instructions.
8559 // dynamic state info needed for verification
8560 EntryState verCurrentState;
8562 // this ptr of object type .ctors are considered intited only after
8563 // the base class ctor is called, or an alternate ctor is called.
8564 // An uninited this ptr can be used to access fields, but cannot
8565 // be used to call a member function.
8566 BOOL verTrackObjCtorInitState;
8568 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8570 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8571 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8572 void verInitCurrentState();
8573 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8575 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8576 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8577 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8579 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8580 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8581 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8582 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8583 typeInfo verMakeTypeInfo(CorInfoType ciType,
8584 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8585 BOOL verIsSDArray(typeInfo ti);
8586 typeInfo verGetArrayElemType(typeInfo ti);
8588 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8589 BOOL verNeedsVerification();
8590 BOOL verIsByRefLike(const typeInfo& ti);
8591 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8593 // generic type variables range over types that satisfy IsBoxable
8594 BOOL verIsBoxable(const typeInfo& ti);
8596 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8597 DEBUGARG(unsigned line));
8598 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8599 DEBUGARG(unsigned line));
8600 bool verCheckTailCallConstraint(OPCODE opcode,
8601 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8602 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8603 // on a type parameter?
8604 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8605 // return false to the caller.
8606 // If false, it will throw.
8608 bool verIsBoxedValueType(typeInfo ti);
8610 void verVerifyCall(OPCODE opcode,
8611 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8612 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8614 bool readonlyCall, // is this a "readonly." call?
8615 const BYTE* delegateCreateStart,
8616 const BYTE* codeAddr,
8617 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8619 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8621 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8622 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8623 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8624 const CORINFO_FIELD_INFO& fieldInfo,
8625 const typeInfo* tiThis,
8627 BOOL allowPlainStructAsThis = FALSE);
8628 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8629 void verVerifyThisPtrInitialised();
8630 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8632 // Register allocator
8633 void raInitStackFP();
8634 void raEnregisterVarsPrePassStackFP();
8635 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8636 void raEnregisterVarsPostPassStackFP();
8637 void raGenerateFPRefCounts();
8638 void raEnregisterVarsStackFP();
8639 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8641 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8642 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8644 // returns true if enregistering v1 would save more mem accesses than v2
8645 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8648 void raDumpHeightsStackFP();
8649 void raDumpVariableRegIntfFloat();
8652 #if FEATURE_STACK_FP_X87
8654 // Currently, we use FP transition blocks in only 2 situations:
8656 // -conditional jump on longs where FP stack differs with target: it's not strictly
8657 // necessary, but its low frequency and the code would get complicated if we try to
8658 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8659 // minimize the way we generate the jump code.
8660 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8661 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8663 // However, transition blocks have 2 problems
8665 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8666 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8667 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8668 // in the right place without preordering them), this causes us to have to generate the transition
8669 // blocks in the cold area if we want procedure splitting.
8672 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8673 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8674 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8675 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8676 // a big change in the exception.
8678 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8679 // optimizations. For these 2 cases:
8681 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8682 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8683 // a switch statement.
8685 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8686 // current procedure splitting and exception code have.
8687 bool compMayHaveTransitionBlocks;
8689 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8691 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8693 unsigned raCntStkStackFP;
8694 unsigned raCntWtdStkDblStackFP;
8695 unsigned raCntStkParamDblStackFP;
8697 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8698 // TODO: Do we want to put this in LclVarDsc?
8699 unsigned raPayloadStackFP[lclMAX_TRACKED];
8700 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8702 // Useful for debugging
8703 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8705 #endif // FEATURE_STACK_FP_X87
8708 // One line log function. Default level is 0. Increasing it gives you
8709 // more log information
8711 // levels are currently unused: #define JITDUMP(level,...) ();
8712 void JitLogEE(unsigned level, const char* fmt, ...);
8714 bool compDebugBreak;
8716 bool compJitHaltMethod();
8721 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8722 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8724 XX GS Security checks for unsafe buffers XX
8726 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8727 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8730 struct ShadowParamVarInfo
8732 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8733 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8735 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8737 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8738 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8739 // slots and update all trees to refer to shadow slots is done immediately after
8740 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8741 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8742 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8743 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8744 // creating a shadow slot even though this routine returns true.
8746 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8747 // required. There are two cases under which a reg arg could potentially be used from its
8749 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8750 // b) LSRA spills it
8752 // Possible solution to address case (a)
8753 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8754 // in this routine. Note that live out of exception handler is something we may not be
8755 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8756 // Therefore, for methods with exception handling and need GS cookie check we might have
8757 // to take conservative approach.
8759 // Possible solution to address case (b)
8760 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8761 // create a new spill temp if the method needs GS cookie check.
8762 return varDsc->lvIsParam;
8763 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8764 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8771 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8776 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8777 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8778 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8780 void gsGSChecksInitCookie(); // Grabs cookie variable
8781 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8782 bool gsFindVulnerableParams(); // Shadow param analysis code
8783 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8785 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8786 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8788 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8789 // This can be overwritten by setting complus_JITInlineSize env variable.
8791 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8794 #ifdef FEATURE_JIT_METHOD_PERF
8795 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8796 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8798 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8799 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8801 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8803 #if MEASURE_CLRAPI_CALLS
8804 // Thin wrappers that call into JitTimer (if present).
8805 inline void CLRApiCallEnter(unsigned apix);
8806 inline void CLRApiCallLeave(unsigned apix);
8809 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8810 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8815 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8816 // These variables are associated with maintaining SQM data about compile time.
8817 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8818 // in the current compilation.
8819 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8820 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8821 // the inlining phase in the current compilation.
8822 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8824 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8825 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8826 // type-loading and class initialization).
8827 void RecordStateAtEndOfInlining();
8828 // Assumes being called at the end of compilation. Update the SQM state.
8829 void RecordStateAtEndOfCompilation();
8831 #ifdef FEATURE_CLRSQM
8832 // Does anything SQM related necessary at process shutdown time.
8833 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8834 #endif // FEATURE_CLRSQM
8837 #if FUNC_INFO_LOGGING
8838 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8839 // filename to write it to.
8840 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8841 #endif // FUNC_INFO_LOGGING
8843 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8845 // Is the compilation in a full trust context?
8846 bool compIsFullTrust();
8848 #ifndef FEATURE_TRACELOGGING
8849 // Should we actually fire the noway assert body and the exception handler?
8850 bool compShouldThrowOnNoway();
8851 #else // FEATURE_TRACELOGGING
8852 // Should we actually fire the noway assert body and the exception handler?
8853 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8855 // Telemetry instance to use per method compilation.
8856 JitTelemetry compJitTelemetry;
8858 // Get common parameters that have to be logged with most telemetry data.
8859 void compGetTelemetryDefaults(const char** assemblyName,
8860 const char** scopeName,
8861 const char** methodName,
8862 unsigned* methodHash);
8863 #endif // !FEATURE_TRACELOGGING
8867 NodeToTestDataMap* m_nodeTestData;
8869 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8870 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8871 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8872 // Current kept in this.
8874 NodeToTestDataMap* GetNodeTestData()
8876 Compiler* compRoot = impInlineRoot();
8877 if (compRoot->m_nodeTestData == nullptr)
8879 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8881 return compRoot->m_nodeTestData;
8884 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8886 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8887 // currently occur in the AST graph.
8888 NodeToIntMap* FindReachableNodesInNodeTestData();
8890 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8891 // test data, associate that data with "to".
8892 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8894 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8895 // have annotations, attach similar annotations to the corresponding nodes in "to".
8896 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8898 // These are the methods that test that the various conditions implied by the
8899 // test attributes are satisfied.
8900 void JitTestCheckSSA(); // SSA builder tests.
8901 void JitTestCheckVN(); // Value numbering tests.
8904 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8906 FieldSeqStore* m_fieldSeqStore;
8908 FieldSeqStore* GetFieldSeqStore()
8910 Compiler* compRoot = impInlineRoot();
8911 if (compRoot->m_fieldSeqStore == nullptr)
8913 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8914 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8915 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8917 return compRoot->m_fieldSeqStore;
8920 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8922 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8923 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8924 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8925 // attach the field sequence directly to the address node.
8926 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8928 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8930 // Don't need to worry about inlining here
8931 if (m_zeroOffsetFieldMap == nullptr)
8933 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8935 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8936 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8938 return m_zeroOffsetFieldMap;
8941 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8942 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8943 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8944 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8945 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8946 // record the the field sequence using the ZeroOffsetFieldMap described above.
8948 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8949 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8950 // CoreRT. Such case is handled same as the default case.
8951 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8953 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8955 NodeToArrayInfoMap* m_arrayInfoMap;
8957 NodeToArrayInfoMap* GetArrayInfoMap()
8959 Compiler* compRoot = impInlineRoot();
8960 if (compRoot->m_arrayInfoMap == nullptr)
8962 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8963 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8964 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
8966 return compRoot->m_arrayInfoMap;
8969 NodeToUnsignedMap* m_heapSsaMap;
8971 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
8972 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
8973 // all the possible heap states are possible initial states of the corresponding catch block(s).)
8974 NodeToUnsignedMap* GetHeapSsaMap()
8976 Compiler* compRoot = impInlineRoot();
8977 if (compRoot->m_heapSsaMap == nullptr)
8979 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8980 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8981 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
8983 return compRoot->m_heapSsaMap;
8986 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
8987 CORINFO_CLASS_HANDLE m_refAnyClass;
8988 CORINFO_FIELD_HANDLE GetRefanyDataField()
8990 if (m_refAnyClass == nullptr)
8992 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
8994 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
8996 CORINFO_FIELD_HANDLE GetRefanyTypeField()
8998 if (m_refAnyClass == nullptr)
9000 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9002 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9006 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9008 #if ALLVARSET_COUNTOPS
9009 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9012 static HelperCallProperties s_helperCallProperties;
9014 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9015 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9016 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9018 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9021 unsigned __int8* offset0,
9022 unsigned __int8* offset1);
9023 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9024 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9026 void fgMorphMultiregStructArgs(GenTreeCall* call);
9027 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9029 }; // end of class Compiler
9031 // Inline methods of CompAllocator.
9032 void* CompAllocator::Alloc(size_t sz)
9034 #if MEASURE_MEM_ALLOC
9035 return m_comp->compGetMem(sz, m_cmk);
9037 return m_comp->compGetMem(sz);
9041 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9043 #if MEASURE_MEM_ALLOC
9044 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9046 return m_comp->compGetMemArray(elems, elemSize);
9050 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9051 inline LclVarDsc::LclVarDsc(Compiler* comp)
9052 : // Initialize the ArgRegs to REG_STK.
9053 // The morph will do the right thing to change
9054 // to the right register if passed in register.
9057 #if FEATURE_MULTIREG_ARGS
9058 _lvOtherArgReg(REG_STK)
9060 #endif // FEATURE_MULTIREG_ARGS
9062 lvRefBlks(BlockSetOps::UninitVal())
9064 #endif // ASSERTION_PROP
9065 lvPerSsaData(comp->getAllocator())
9070 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9071 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9073 XX Miscellaneous Compiler stuff XX
9075 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9076 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9079 // Values used to mark the types a stack slot is used for
9081 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9082 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9083 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9084 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9085 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9086 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9087 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9088 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9090 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9092 /*****************************************************************************
9094 * Variables to keep track of total code amounts.
9099 extern size_t grossVMsize;
9100 extern size_t grossNCsize;
9101 extern size_t totalNCsize;
9103 extern unsigned genMethodICnt;
9104 extern unsigned genMethodNCnt;
9105 extern size_t gcHeaderISize;
9106 extern size_t gcPtrMapISize;
9107 extern size_t gcHeaderNSize;
9108 extern size_t gcPtrMapNSize;
9110 #endif // DISPLAY_SIZES
9112 /*****************************************************************************
9114 * Variables to keep track of basic block counts (more data on 1 BB methods)
9117 #if COUNT_BASIC_BLOCKS
9118 extern Histogram bbCntTable;
9119 extern Histogram bbOneBBSizeTable;
9122 /*****************************************************************************
9124 * Used by optFindNaturalLoops to gather statistical information such as
9125 * - total number of natural loops
9126 * - number of loops with 1, 2, ... exit conditions
9127 * - number of loops that have an iterator (for like)
9128 * - number of loops that have a constant iterator
9133 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9134 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9135 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9136 extern unsigned totalLoopCount; // counts the total number of natural loops
9137 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9138 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9139 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9140 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9142 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9143 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9144 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9145 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9146 extern Histogram loopCountTable; // Histogram of loop counts
9147 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9149 #endif // COUNT_LOOPS
9151 /*****************************************************************************
9152 * variables to keep track of how many iterations we go in a dataflow pass
9157 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9158 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9160 #endif // DATAFLOW_ITER
9162 #if MEASURE_BLOCK_SIZE
9163 extern size_t genFlowNodeSize;
9164 extern size_t genFlowNodeCnt;
9165 #endif // MEASURE_BLOCK_SIZE
9167 #if MEASURE_NODE_SIZE
9168 struct NodeSizeStats
9173 genTreeNodeSize = 0;
9174 genTreeNodeActualSize = 0;
9177 size_t genTreeNodeCnt;
9178 size_t genTreeNodeSize; // The size we allocate
9179 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9180 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9181 // a smaller node to a larger one. TODO-Cleanup: add stats on
9182 // SetOper()/ChangeOper() usage to quanitfy this.
9184 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9185 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9186 extern Histogram genTreeNcntHist;
9187 extern Histogram genTreeNsizHist;
9188 #endif // MEASURE_NODE_SIZE
9190 /*****************************************************************************
9191 * Count fatal errors (including noway_asserts).
9195 extern unsigned fatal_badCode;
9196 extern unsigned fatal_noWay;
9197 extern unsigned fatal_NOMEM;
9198 extern unsigned fatal_noWayAssertBody;
9200 extern unsigned fatal_noWayAssertBodyArgs;
9202 extern unsigned fatal_NYI;
9203 #endif // MEASURE_FATAL
9205 /*****************************************************************************
9209 #ifdef _TARGET_XARCH_
9211 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9212 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9213 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9215 const instruction INS_AND = INS_and;
9216 const instruction INS_OR = INS_or;
9217 const instruction INS_XOR = INS_xor;
9218 const instruction INS_NEG = INS_neg;
9219 const instruction INS_TEST = INS_test;
9220 const instruction INS_MUL = INS_imul;
9221 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9222 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9223 const instruction INS_BREAKPOINT = INS_int3;
9224 const instruction INS_ADDC = INS_adc;
9225 const instruction INS_SUBC = INS_sbb;
9226 const instruction INS_NOT = INS_not;
9232 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9233 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9234 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9236 const instruction INS_AND = INS_and;
9237 const instruction INS_OR = INS_orr;
9238 const instruction INS_XOR = INS_eor;
9239 const instruction INS_NEG = INS_rsb;
9240 const instruction INS_TEST = INS_tst;
9241 const instruction INS_MUL = INS_mul;
9242 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9243 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9244 const instruction INS_BREAKPOINT = INS_bkpt;
9245 const instruction INS_ADDC = INS_adc;
9246 const instruction INS_SUBC = INS_sbc;
9247 const instruction INS_NOT = INS_mvn;
9251 #ifdef _TARGET_ARM64_
9253 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9254 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9255 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9257 const instruction INS_AND = INS_and;
9258 const instruction INS_OR = INS_orr;
9259 const instruction INS_XOR = INS_eor;
9260 const instruction INS_NEG = INS_neg;
9261 const instruction INS_TEST = INS_tst;
9262 const instruction INS_MUL = INS_mul;
9263 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9264 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9265 const instruction INS_BREAKPOINT = INS_bkpt;
9266 const instruction INS_ADDC = INS_adc;
9267 const instruction INS_SUBC = INS_sbc;
9268 const instruction INS_NOT = INS_mvn;
9272 /*****************************************************************************/
9274 extern const BYTE genTypeSizes[];
9275 extern const BYTE genTypeAlignments[];
9276 extern const BYTE genTypeStSzs[];
9277 extern const BYTE genActualTypes[];
9279 /*****************************************************************************/
9281 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9282 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9285 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9286 #elif defined(_TARGET_ARM64_)
9287 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9290 /*****************************************************************************/
9292 #define REG_CORRUPT regNumber(REG_NA + 1)
9293 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9294 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9296 /*****************************************************************************/
9298 extern BasicBlock dummyBB;
9300 /*****************************************************************************/
9301 /*****************************************************************************/
9303 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9304 // nodes of a statement in execution order.
9305 // __stmt: a GT_STMT type GenTree*
9306 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9308 #define foreach_treenode_execution_order(__node, __stmt) \
9309 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9311 // foreach_block: An iterator over all blocks in the function.
9312 // __compiler: the Compiler* object
9313 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9315 #define foreach_block(__compiler, __block) \
9316 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9318 /*****************************************************************************/
9319 /*****************************************************************************/
9323 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9325 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9326 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9328 XX Debugging helpers XX
9330 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9331 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9334 /*****************************************************************************/
9335 /* The following functions are intended to be called from the debugger, to dump
9336 * various data structures. The can be used in the debugger Watch or Quick Watch
9337 * windows. They are designed to be short to type and take as few arguments as
9338 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9339 * See the function definition comment for more details.
9342 void cBlock(Compiler* comp, BasicBlock* block);
9343 void cBlocks(Compiler* comp);
9344 void cBlocksV(Compiler* comp);
9345 void cTree(Compiler* comp, GenTree* tree);
9346 void cTrees(Compiler* comp);
9347 void cEH(Compiler* comp);
9348 void cVar(Compiler* comp, unsigned lclNum);
9349 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9350 void cVars(Compiler* comp);
9351 void cVarsFinal(Compiler* comp);
9352 void cBlockPreds(Compiler* comp, BasicBlock* block);
9353 void cReach(Compiler* comp);
9354 void cDoms(Compiler* comp);
9355 void cLiveness(Compiler* comp);
9356 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9358 void cFuncIR(Compiler* comp);
9359 void cBlockIR(Compiler* comp, BasicBlock* block);
9360 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9361 void cTreeIR(Compiler* comp, GenTree* tree);
9362 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9363 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9364 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9365 int cOperandIR(Compiler* comp, GenTree* operand);
9366 int cLeafIR(Compiler* comp, GenTree* tree);
9367 int cIndirIR(Compiler* comp, GenTree* tree);
9368 int cListIR(Compiler* comp, GenTree* list);
9369 int cSsaNumIR(Compiler* comp, GenTree* tree);
9370 int cValNumIR(Compiler* comp, GenTree* tree);
9371 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9373 void dBlock(BasicBlock* block);
9376 void dTree(GenTree* tree);
9379 void dVar(unsigned lclNum);
9380 void dVarDsc(LclVarDsc* varDsc);
9383 void dBlockPreds(BasicBlock* block);
9387 void dCVarSet(VARSET_VALARG_TP vars);
9389 void dVarSet(VARSET_VALARG_TP vars);
9390 void dRegMask(regMaskTP mask);
9393 void dBlockIR(BasicBlock* block);
9394 void dTreeIR(GenTree* tree);
9395 void dLoopIR(Compiler::LoopDsc* loop);
9396 void dLoopNumIR(unsigned loopNum);
9397 int dTabStopIR(int curr, int tabstop);
9398 int dTreeTypeIR(GenTree* tree);
9399 int dTreeKindsIR(GenTree* tree);
9400 int dTreeFlagsIR(GenTree* tree);
9401 int dOperandIR(GenTree* operand);
9402 int dLeafIR(GenTree* tree);
9403 int dIndirIR(GenTree* tree);
9404 int dListIR(GenTree* list);
9405 int dSsaNumIR(GenTree* tree);
9406 int dValNumIR(GenTree* tree);
9407 int dDependsIR(GenTree* comma);
9410 GenTree* dFindTree(GenTree* tree, unsigned id);
9411 GenTree* dFindTree(unsigned id);
9412 GenTreeStmt* dFindStmt(unsigned id);
9413 BasicBlock* dFindBlock(unsigned bbNum);
9417 #include "compiler.hpp" // All the shared inline functions
9419 /*****************************************************************************/
9420 #endif //_COMPILER_H_
9421 /*****************************************************************************/