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?
3428 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3430 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3431 bool fgHasPostfix; // any postfix ++/-- found?
3432 unsigned fgIncrCount; // number of increment nodes found
3434 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3438 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3439 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3442 bool fgRemoveRestOfBlock; // true if we know that we will throw
3443 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3445 // There are two modes for ordering of the trees.
3446 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3447 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3448 // by traversing the tree according to the order of the operands.
3449 // - In FGOrderLinear, the dominant ordering is the linear order.
3456 FlowGraphOrder fgOrder;
3458 // The following are boolean flags that keep track of the state of internal data structures
3460 bool fgStmtListThreaded;
3461 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3462 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3463 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3464 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3465 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3466 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3467 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3468 // This is derived from the profile data
3469 // or is BB_UNITY_WEIGHT when we don't have profile data
3471 #if FEATURE_EH_FUNCLETS
3472 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3473 #endif // FEATURE_EH_FUNCLETS
3475 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3476 // since fgMorphTree can be called from several places
3477 bool fgExpandInline; // indicates that we are creating tree for the inliner
3479 bool impBoxTempInUse; // the temp below is valid and available
3480 unsigned impBoxTemp; // a temporary that is used for boxing
3483 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3484 // and we are trying to compile again in a "safer", minopts mode?
3488 unsigned impInlinedCodeSize;
3491 //-------------------------------------------------------------------------
3499 void fgRemoveEmptyTry();
3501 void fgRemoveEmptyFinally();
3503 void fgCloneFinally();
3505 void fgCleanupContinuation(BasicBlock* continuation);
3507 void fgUpdateFinallyTargetFlags();
3509 GenTreePtr fgGetCritSectOfStaticMethod();
3511 #if !defined(_TARGET_X86_)
3513 void fgAddSyncMethodEnterExit();
3515 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3517 void fgConvertSyncReturnToLeave(BasicBlock* block);
3519 #endif // !_TARGET_X86_
3521 void fgAddReversePInvokeEnterExit();
3523 bool fgMoreThanOneReturnBlock();
3525 // The number of separate return points in the method.
3526 unsigned fgReturnCount;
3528 void fgAddInternal();
3530 bool fgFoldConditional(BasicBlock* block);
3532 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3533 void fgMorphBlocks();
3535 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3537 void fgCheckArgCnt();
3538 void fgSetOptions();
3541 static fgWalkPreFn fgAssertNoQmark;
3542 void fgPreExpandQmarkChecks(GenTreePtr expr);
3543 void fgPostExpandQmarkChecks();
3544 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3547 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3549 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3550 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3551 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3552 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3553 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3555 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3556 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3557 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3558 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3560 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3561 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3562 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3563 void fgExpandQmarkNodes();
3567 // Do "simple lowering." This functionality is (conceptually) part of "general"
3568 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3569 void fgSimpleLowering();
3571 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3573 GenTreePtr fgInitThisClass();
3575 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3577 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3579 void fgLocalVarLiveness();
3581 void fgLocalVarLivenessInit();
3583 #ifdef LEGACY_BACKEND
3584 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3586 void fgPerNodeLocalVarLiveness(GenTree* node);
3588 void fgPerBlockLocalVarLiveness();
3590 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3592 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3594 // This is used in the liveness computation, as a temporary. When we use the
3595 // arbitrary-length VarSet representation, it is better not to allocate a new one
3597 VARSET_TP fgMarkIntfUnionVS;
3599 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3601 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3603 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3605 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3607 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3609 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3611 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3612 GenTreePtr startNode,
3614 VARSET_VALARG_TP volatileVars,
3615 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3617 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3619 bool fgRemoveDeadStore(GenTree** pTree,
3623 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3625 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3627 // For updating liveset during traversal AFTER fgComputeLife has completed
3628 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3629 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3631 // Returns the set of live variables after endTree,
3632 // assuming that liveSet is the set of live variables BEFORE tree.
3633 // Requires that fgComputeLife has completed, and that tree is in the same
3634 // statement as endTree, and that it comes before endTree in execution order
3636 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3638 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3639 while (tree != nullptr && tree != endTree->gtNext)
3641 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3642 tree = tree->gtNext;
3644 assert(tree == endTree->gtNext);
3648 void fgInterBlockLocalVarLiveness();
3650 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3651 // "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
3652 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3653 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3654 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3655 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3656 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3658 if (m_opAsgnVarDefSsaNums == nullptr)
3660 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3662 return m_opAsgnVarDefSsaNums;
3665 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3666 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3667 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3669 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3671 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3672 // Except: assumes that lcl is a def, and if it is
3673 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3674 // rather than the "use" SSA number recorded in the tree "lcl".
3675 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3677 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3678 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3679 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3680 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3681 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3683 // (byref addrS1 = &s1,
3684 // *(addrS1 * offsetof(f0)) = s2f0,
3686 // *(addrS1 * offsetof(fn)) = s2fn)
3688 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3689 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3690 // give it SSA names and value numbers?
3692 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3693 // end with an instance of the structure below, whose fields are described in the declaration.
3694 struct IndirectAssignmentAnnotation
3696 unsigned m_lclNum; // The local num that is being indirectly assigned.
3697 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3698 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3699 // be the singleton field sequence "g". The individual assignments would
3700 // further append the fields of "s.g" to that.
3701 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3702 // structure has a single field).
3703 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3704 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3707 IndirectAssignmentAnnotation(unsigned lclNum,
3708 FieldSeqNode* fldSeq,
3710 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3711 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3712 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3716 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3717 NodeToIndirAssignMap;
3718 NodeToIndirAssignMap* m_indirAssignMap;
3719 NodeToIndirAssignMap* GetIndirAssignMap()
3721 if (m_indirAssignMap == nullptr)
3723 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3724 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3725 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3727 return m_indirAssignMap;
3730 // Performs SSA conversion.
3733 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3734 void fgResetForSsa();
3736 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3738 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3739 inline bool fgExcludeFromSsa(unsigned lclNum);
3741 // The value numbers for this compilation.
3742 ValueNumStore* vnStore;
3745 ValueNumStore* GetValueNumStore()
3750 // Do value numbering (assign a value number to each
3752 void fgValueNumber();
3754 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3755 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3756 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3757 // match the element type of the array or fldSeq. When this type doesn't match
3758 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3760 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3763 FieldSeqNode* fldSeq,
3767 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3768 // has been parsed to yield the other input arguments. If evaluation of the address
3769 // can raise exceptions, those should be captured in the exception set "excVN."
3770 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3771 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3772 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3773 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3774 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3776 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3777 CORINFO_CLASS_HANDLE elemTypeEq,
3781 FieldSeqNode* fldSeq);
3783 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3784 // by evaluating the array index expression "tree". Returns the value number resulting from
3785 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3786 // "GT_IND" that does the dereference, and it is given the returned value number.
3787 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3789 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3791 // Utility functions for fgValueNumber.
3793 // Perform value-numbering for the trees in "blk".
3794 void fgValueNumberBlock(BasicBlock* blk);
3796 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3797 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3798 // assumed for the heap at the start "entryBlk".
3799 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3801 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3802 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3804 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3805 // value in that SSA #.
3806 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3808 // The input 'tree' is a leaf node that is a constant
3809 // Assign the proper value number to the tree
3810 void fgValueNumberTreeConst(GenTreePtr tree);
3812 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3813 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3815 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3817 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3819 // Does value-numbering for a block assignment.
3820 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3822 // Does value-numbering for a cast tree.
3823 void fgValueNumberCastTree(GenTreePtr tree);
3825 // Does value-numbering for an intrinsic tree.
3826 void fgValueNumberIntrinsic(GenTreePtr tree);
3828 // Does value-numbering for a call. We interpret some helper calls.
3829 void fgValueNumberCall(GenTreeCall* call);
3831 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3832 void fgUpdateArgListVNs(GenTreeArgList* args);
3834 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3835 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3837 // Requires "helpCall" to be a helper call. Assigns it a value number;
3838 // we understand the semantics of some of the calls. Returns "true" if
3839 // the call may modify the heap (we assume arbitrary memory side effects if so).
3840 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3842 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3843 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3845 // This is the current value number for the "Heap" implicit variable while
3846 // doing value numbering. This is the value number under the "liberal" interpretation
3847 // of heap values; the "conservative" interpretation needs no VN, since every access of
3848 // the heap yields an unknown value.
3849 ValueNum fgCurHeapVN;
3851 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3852 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3853 // is 1, and the rest is an encoding of "elemTyp".
3854 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3856 if (elemStructType != nullptr)
3858 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3859 varTypeIsIntegral(elemTyp));
3860 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3861 return elemStructType;
3865 elemTyp = varTypeUnsignedToSigned(elemTyp);
3866 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3869 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3870 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3871 // the struct type of the element).
3872 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3874 size_t clsHndVal = size_t(clsHnd);
3875 if (clsHndVal & 0x1)
3877 return var_types(clsHndVal >> 1);
3885 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3886 var_types getJitGCType(BYTE gcType);
3888 enum structPassingKind
3890 SPK_Unknown, // Invalid value, never returned
3891 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3892 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3893 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3894 // parameters registers are used, then the stack will be used)
3895 // for X86 passed on the stack, for ARM32 passed in registers
3896 // or the stack or split between registers and the stack.
3897 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3899 }; // The struct is passed/returned by reference to a copy/buffer.
3901 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3902 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3903 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3904 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3906 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3908 // Get the type that is used to pass 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 getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3912 structPassingKind* wbPassStruct,
3913 unsigned structSize = 0);
3915 // Get the type that is used to return values of the given struct type.
3916 // If you have already retrieved the struct size then pass it as the optional third argument
3918 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3919 structPassingKind* wbPassStruct = nullptr,
3920 unsigned structSize = 0);
3923 // Print a representation of "vnp" or "vn" on standard output.
3924 // If "level" is non-zero, we also print out a partial expansion of the value.
3925 void vnpPrint(ValueNumPair vnp, unsigned level);
3926 void vnPrint(ValueNum vn, unsigned level);
3929 // Dominator computation member functions
3930 // Not exposed outside Compiler
3932 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3934 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3936 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3937 // flow graph. We first assume the fields bbIDom on each
3938 // basic block are invalid. This computation is needed later
3939 // by fgBuildDomTree to build the dominance tree structure.
3940 // Based on: A Simple, Fast Dominance Algorithm
3941 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3943 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3944 // Note: this is relatively slow compared to calling fgDominate(),
3945 // especially if dealing with a single block versus block check.
3947 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3949 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3951 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3953 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3955 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3957 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3958 // processed in topological sort, this function takes care of that.
3960 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3962 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3963 // Returns this as a set.
3965 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3966 // root nodes. Returns this as a set.
3969 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
3972 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
3973 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
3976 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
3977 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
3978 // && postOrder(A) >= postOrder(B) making the computation O(1).
3979 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
3981 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
3983 void fgUpdateChangedFlowGraph();
3986 // Compute the predecessors of the blocks in the control flow graph.
3987 void fgComputePreds();
3989 // Remove all predecessor information.
3990 void fgRemovePreds();
3992 // Compute the cheap flow graph predecessors lists. This is used in some early phases
3993 // before the full predecessors lists are computed.
3994 void fgComputeCheapPreds();
3997 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
3999 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4009 // Initialize the per-block variable sets (used for liveness analysis).
4010 void fgInitBlockVarSets();
4012 // true if we've gone through and created GC Poll calls.
4013 bool fgGCPollsCreated;
4014 void fgMarkGCPollBlocks();
4015 void fgCreateGCPolls();
4016 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4018 // Requires that "block" is a block that returns from
4019 // a finally. Returns the number of successors (jump targets of
4020 // of blocks in the covered "try" that did a "LEAVE".)
4021 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4023 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4024 // a finally. Returns its "i"th successor (jump targets of
4025 // of blocks in the covered "try" that did a "LEAVE".)
4026 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4027 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4030 // Factor out common portions of the impls of the methods above.
4031 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4034 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4035 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4036 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4037 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4038 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4039 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4040 // we leave the entry associated with the block, but it will no longer be accessed.)
4041 struct SwitchUniqueSuccSet
4043 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4044 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4047 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4048 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4049 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4050 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4053 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4054 BlockToSwitchDescMap;
4057 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4058 // iteration over only the distinct successors.
4059 BlockToSwitchDescMap* m_switchDescMap;
4062 BlockToSwitchDescMap* GetSwitchDescMap()
4064 if (m_switchDescMap == nullptr)
4066 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4068 return m_switchDescMap;
4071 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4072 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4073 // we don't accidentally look up and return the wrong switch data.
4074 void InvalidateUniqueSwitchSuccMap()
4076 m_switchDescMap = nullptr;
4079 // Requires "switchBlock" to be a block that ends in a switch. Returns
4080 // the corresponding SwitchUniqueSuccSet.
4081 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4083 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4084 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4085 // remove it from "this", and ensure that "to" is a member.
4086 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4088 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4089 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4091 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4093 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4095 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4097 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4099 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4101 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4103 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4105 void fgRemoveBlockAsPred(BasicBlock* block);
4107 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4109 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4111 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4113 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4115 flowList* fgAddRefPred(BasicBlock* block,
4116 BasicBlock* blockPred,
4117 flowList* oldEdge = nullptr,
4118 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4121 void fgFindBasicBlocks();
4123 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4125 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4127 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4128 bool putInTryRegion,
4129 BasicBlock* startBlk,
4131 BasicBlock* nearBlk,
4132 BasicBlock* jumpBlk,
4135 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4137 void fgRemoveEmptyBlocks();
4139 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4141 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4143 void fgCreateLoopPreHeader(unsigned lnum);
4145 void fgUnreachableBlock(BasicBlock* block);
4147 void fgRemoveConditionalJump(BasicBlock* block);
4149 BasicBlock* fgLastBBInMainFunction();
4151 BasicBlock* fgEndBBAfterMainFunction();
4153 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4155 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4157 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4159 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4161 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4163 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4165 bool fgRenumberBlocks();
4167 bool fgExpandRarelyRunBlocks();
4169 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4171 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4173 enum FG_RELOCATE_TYPE
4175 FG_RELOCATE_TRY, // relocate the 'try' region
4176 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4178 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4180 #if FEATURE_EH_FUNCLETS
4181 #if defined(_TARGET_ARM_)
4182 void fgClearFinallyTargetBit(BasicBlock* block);
4183 #endif // defined(_TARGET_ARM_)
4184 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4185 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4186 void fgInsertFuncletPrologBlock(BasicBlock* block);
4187 void fgCreateFuncletPrologBlocks();
4188 void fgCreateFunclets();
4189 #else // !FEATURE_EH_FUNCLETS
4190 bool fgRelocateEHRegions();
4191 #endif // !FEATURE_EH_FUNCLETS
4193 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4195 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4197 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4199 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4201 bool fgOptimizeEmptyBlock(BasicBlock* block);
4203 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4205 bool fgOptimizeBranch(BasicBlock* bJump);
4207 bool fgOptimizeSwitchBranches(BasicBlock* block);
4209 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4211 bool fgOptimizeSwitchJumps();
4213 void fgPrintEdgeWeights();
4215 void fgComputeEdgeWeights();
4217 void fgReorderBlocks();
4219 void fgDetermineFirstColdBlock();
4221 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4223 bool fgUpdateFlowGraph(bool doTailDup = false);
4225 void fgFindOperOrder();
4227 // method that returns if you should split here
4228 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4230 void fgSetBlockOrder();
4232 void fgRemoveReturnBlock(BasicBlock* block);
4234 /* Helper code that has been factored out */
4235 inline void fgConvertBBToThrowBB(BasicBlock* block);
4237 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4238 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4239 GenTreePtr fgMakeTmpArgNode(
4240 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4242 // The following check for loops that don't execute calls
4243 bool fgLoopCallMarked;
4245 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4246 void fgLoopCallMark();
4248 void fgMarkLoopHead(BasicBlock* block);
4250 unsigned fgGetCodeEstimate(BasicBlock* block);
4253 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4254 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4255 bool fgDumpFlowGraph(Phases phase);
4257 #endif // DUMP_FLOWGRAPHS
4262 void fgDispBBLiveness(BasicBlock* block);
4263 void fgDispBBLiveness();
4264 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4265 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4266 void fgDispBasicBlocks(bool dumpTrees = false);
4267 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4268 void fgDumpBlock(BasicBlock* block);
4269 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4271 static fgWalkPreFn fgStress64RsltMulCB;
4272 void fgStress64RsltMul();
4273 void fgDebugCheckUpdate();
4274 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4275 void fgDebugCheckBlockLinks();
4276 void fgDebugCheckLinks(bool morphTrees = false);
4277 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4278 void fgDebugCheckFlags(GenTreePtr tree);
4279 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4280 void fgDebugCheckTryFinallyExits();
4283 #ifdef LEGACY_BACKEND
4284 static void fgOrderBlockOps(GenTreePtr tree,
4288 GenTreePtr* opsPtr, // OUT
4289 regMaskTP* regsPtr); // OUT
4290 #endif // LEGACY_BACKEND
4292 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4293 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4295 inline bool fgIsInlining()
4297 return fgExpandInline;
4300 void fgTraverseRPO();
4302 //--------------------- Walking the trees in the IR -----------------------
4307 fgWalkPreFn* wtprVisitorFn;
4308 fgWalkPostFn* wtpoVisitorFn;
4309 void* pCallbackData; // user-provided data
4310 bool wtprLclsOnly; // whether to only visit lclvar nodes
4311 GenTreePtr parent; // parent of current node, provided to callback
4312 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4314 bool printModified; // callback can use this
4318 template <bool computeStack>
4319 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4321 // general purpose tree-walker that is capable of doing pre- and post- order
4322 // callbacks at the same time
4323 template <bool doPreOrder, bool doPostOrder>
4324 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4326 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4327 fgWalkPreFn* visitor,
4328 void* pCallBackData = nullptr,
4329 bool lclVarsOnly = false,
4330 bool computeStack = false);
4332 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4333 fgWalkPreFn* preVisitor,
4334 fgWalkPostFn* postVisitor,
4335 void* pCallBackData = nullptr);
4337 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4341 template <bool computeStack>
4342 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4344 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4345 fgWalkPostFn* visitor,
4346 void* pCallBackData = nullptr,
4347 bool computeStack = false);
4349 // An fgWalkPreFn that looks for expressions that have inline throws in
4350 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4351 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4352 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4353 // properly propagated to parent trees). It returns WALK_CONTINUE
4355 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4356 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4357 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4359 /**************************************************************************
4361 *************************************************************************/
4364 friend class SsaBuilder;
4365 friend struct ValueNumberState;
4367 //--------------------- Detect the basic blocks ---------------------------
4369 BasicBlock** fgBBs; // Table of pointers to the BBs
4371 void fgInitBBLookup();
4372 BasicBlock* fgLookupBB(unsigned addr);
4374 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4376 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4378 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4380 void fgLinkBasicBlocks();
4382 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4384 void fgCheckBasicBlockControlFlow();
4386 void fgControlFlowPermitted(BasicBlock* blkSrc,
4387 BasicBlock* blkDest,
4388 BOOL IsLeave = false /* is the src a leave block */);
4390 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4392 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4394 void fgAdjustForAddressExposedOrWrittenThis();
4396 bool fgProfileData_ILSizeMismatch;
4397 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4398 ULONG fgProfileBufferCount;
4399 ULONG fgNumProfileRuns;
4401 unsigned fgStressBBProf()
4404 unsigned result = JitConfig.JitStressBBProf();
4407 if (compStressCompile(STRESS_BB_PROFILE, 15))
4418 bool fgHaveProfileData();
4419 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4421 bool fgIsUsingProfileWeights()
4423 return (fgHaveProfileData() || fgStressBBProf());
4425 void fgInstrumentMethod();
4427 //-------- Insert a statement at the start or end of a basic block --------
4431 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4435 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4437 public: // Used by linear scan register allocation
4438 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4441 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4442 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4444 public: // Used by linear scan register allocation
4445 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4448 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4450 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4452 // Create a new temporary variable to hold the result of *ppTree,
4453 // and transform the graph accordingly.
4454 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4455 GenTree* fgMakeMultiUse(GenTree** ppTree);
4458 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4459 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4460 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4462 //-------- Determine the order in which the trees will be evaluated -------
4464 unsigned fgTreeSeqNum;
4465 GenTree* fgTreeSeqLst;
4466 GenTree* fgTreeSeqBeg;
4468 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4469 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4470 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4471 void fgSetStmtSeq(GenTree* tree);
4472 void fgSetBlockOrder(BasicBlock* block);
4474 //------------------------- Morphing --------------------------------------
4476 unsigned fgPtrArgCntCur;
4477 unsigned fgPtrArgCntMax;
4478 hashBv* fgOutgoingArgTemps;
4479 hashBv* fgCurrentlyInUseArgTemps;
4481 bool compCanEncodePtrArgCntMax();
4483 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4486 void fgMoveOpsLeft(GenTreePtr tree);
4489 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4491 bool fgIsThrow(GenTreePtr tree);
4493 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4494 bool fgIsBlockCold(BasicBlock* block);
4496 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4498 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4500 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4502 bool fgMorphRelopToQmark(GenTreePtr tree);
4504 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4505 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4506 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4507 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4508 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4509 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4510 // small; hence the other fields of MorphAddrContext.
4511 enum MorphAddrContextKind
4516 struct MorphAddrContext
4518 MorphAddrContextKind m_kind;
4519 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4520 // top-level indirection and here have been constants.
4521 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4522 // In that case, is the sum of those constant offsets.
4524 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4529 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4530 static MorphAddrContext s_CopyBlockMAC;
4533 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4534 var_types* baseTypeOut,
4536 unsigned* simdSizeOut,
4537 bool ignoreUsedInSIMDIntrinsic = false);
4538 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4539 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4540 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4541 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4543 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4544 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4545 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4547 #endif // FEATURE_SIMD
4548 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4549 GenTreePtr fgMorphCast(GenTreePtr tree);
4550 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4551 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4553 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4556 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4557 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4559 void fgFixupStructReturn(GenTreePtr call);
4560 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4561 bool fgAddrCouldBeNull(GenTreePtr addr);
4562 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4563 bool fgCanFastTailCall(GenTreeCall* call);
4564 void fgMorphTailCall(GenTreeCall* call);
4565 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4566 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4567 fgArgTabEntryPtr argTabEntry,
4569 IL_OFFSETX callILOffset,
4570 GenTreePtr tmpAssignmentInsertionPoint,
4571 GenTreePtr paramAssignmentInsertionPoint);
4572 static int fgEstimateCallStackSize(GenTreeCall* call);
4573 GenTreePtr fgMorphCall(GenTreeCall* call);
4574 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4575 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4577 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4578 static fgWalkPreFn fgFindNonInlineCandidate;
4580 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4581 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4582 void fgAssignSetVarDef(GenTreePtr tree);
4583 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4584 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4585 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4586 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4587 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4588 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4589 void fgMorphUnsafeBlk(GenTreeObj* obj);
4590 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4591 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4592 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4593 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4594 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4595 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4596 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4598 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4599 GenTreePtr fgMorphConst(GenTreePtr tree);
4602 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4605 #if LOCAL_ASSERTION_PROP
4606 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4607 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4609 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4611 GenTreeStmt* fgMorphStmt;
4613 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4614 // used when morphing big offset.
4616 //----------------------- Liveness analysis -------------------------------
4618 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4619 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4621 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4622 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4623 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4625 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4627 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4628 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4630 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4631 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4633 void fgExtendDbgScopes();
4634 void fgExtendDbgLifetimes();
4637 void fgDispDebugScopes();
4640 //-------------------------------------------------------------------------
4642 // The following keeps track of any code we've added for things like array
4643 // range checking or explicit calls to enable GC, and so on.
4648 AddCodeDsc* acdNext;
4649 BasicBlock* acdDstBlk; // block to which we jump
4651 SpecialCodeKind acdKind; // what kind of a special block is this?
4652 unsigned short acdStkLvl;
4656 static unsigned acdHelper(SpecialCodeKind codeKind);
4658 AddCodeDsc* fgAddCodeList;
4660 bool fgRngChkThrowAdded;
4661 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4663 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4665 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4668 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4671 bool fgIsCodeAdded();
4673 bool fgIsThrowHlpBlk(BasicBlock* block);
4674 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4676 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4678 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4679 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4680 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4681 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4682 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4684 #if FEATURE_MULTIREG_RET
4685 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4686 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4687 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4688 #endif // FEATURE_MULTIREG_RET
4690 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4693 static fgWalkPreFn fgDebugCheckInlineCandidates;
4696 void fgPromoteStructs();
4697 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4698 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4699 void fgMarkImplicitByRefArgs();
4700 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4701 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4702 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4703 void fgMarkAddressExposedLocals();
4704 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4706 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4708 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4710 // The given local variable, required to be a struct variable, is being assigned via
4711 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4712 // the variable is not enregistered, and is therefore not promoted independently.
4713 void fgLclFldAssign(unsigned lclNum);
4715 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4716 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4717 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4718 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4721 bool fgPrintInlinedMethods;
4724 bool fgIsBigOffset(size_t offset);
4726 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4727 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4728 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4729 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4730 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4734 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4739 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4746 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4749 void optRemoveRangeCheck(
4750 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4751 bool optIsRangeCheckRemovable(GenTreePtr tree);
4754 static fgWalkPreFn optValidRangeCheckIndex;
4755 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4758 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4760 /**************************************************************************
4762 *************************************************************************/
4765 // Do hoisting for all loops.
4766 void optHoistLoopCode();
4768 // To represent sets of VN's that have already been hoisted in outer loops.
4769 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4770 typedef VNToBoolMap VNSet;
4772 struct LoopHoistContext
4775 // The set of variables hoisted in the current loop (or nullptr if there are none).
4776 VNSet* m_pHoistedInCurLoop;
4779 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4780 VNSet m_hoistedInParentLoops;
4781 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4782 // Previous decisions on loop-invariance of value numbers in the current loop.
4783 VNToBoolMap m_curLoopVnInvariantCache;
4785 VNSet* GetHoistedInCurLoop(Compiler* comp)
4787 if (m_pHoistedInCurLoop == nullptr)
4789 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4791 return m_pHoistedInCurLoop;
4794 VNSet* ExtractHoistedInCurLoop()
4796 VNSet* res = m_pHoistedInCurLoop;
4797 m_pHoistedInCurLoop = nullptr;
4801 LoopHoistContext(Compiler* comp)
4802 : m_pHoistedInCurLoop(nullptr)
4803 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4804 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4809 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4810 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4811 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4812 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4814 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4815 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4816 // "m_hoistedInParentLoops".
4818 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4820 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4821 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4822 // expressions to "hoistInLoop".
4823 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4825 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4826 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4828 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4829 // that are invariant in loop "lnum" (an index into the optLoopTable)
4830 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4831 // expressions to "hoistInLoop".
4832 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4833 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4834 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4835 bool optHoistLoopExprsForTree(GenTreePtr tree,
4837 LoopHoistContext* hoistCtxt,
4838 bool* firstBlockAndBeforeSideEffect,
4841 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4842 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4844 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4845 // Constants and init values are always loop invariant.
4846 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4847 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4849 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4850 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4851 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4852 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4853 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4855 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4856 // in the loop table.
4857 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4859 // Records the set of "side effects" of all loops: fields (object instance and static)
4860 // written to, and SZ-array element type equivalence classes updated.
4861 void optComputeLoopSideEffects();
4864 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4865 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4866 // static) written to, and SZ-array element type equivalence classes updated.
4867 void optComputeLoopNestSideEffects(unsigned lnum);
4869 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4870 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4872 // Hoist the expression "expr" out of loop "lnum".
4873 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4876 void optOptimizeBools();
4879 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4881 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4884 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4886 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4887 // the loop into a "do-while" loop
4888 // Also finds all natural loops and records them in the loop table
4890 // Optionally clone loops in the loop table.
4891 void optCloneLoops();
4893 // Clone loop "loopInd" in the loop table.
4894 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4896 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4897 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4898 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4900 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4902 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4905 // This enumeration describes what is killed by a call.
4909 CALLINT_NONE, // no interference (most helpers)
4910 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4911 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4912 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4913 CALLINT_ALL, // kills everything (normal method call)
4917 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4918 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4919 // in bbNext order; we use comparisons on the bbNum to decide order.)
4920 // The blocks that define the body are
4921 // first <= top <= entry <= bottom .
4922 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4923 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4924 // Compiler::optFindNaturalLoops().
4927 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4928 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4929 // loop, but not the outer loop.)
4930 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4932 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4933 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4934 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4936 callInterf lpAsgCall; // "callInterf" for calls in the loop
4937 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4938 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4940 unsigned short lpFlags; // Mask of the LPFLG_* constants
4942 unsigned char lpExitCnt; // number of exits from the loop
4944 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4945 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4946 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4947 // (Actually, an "immediately" nested loop --
4948 // no other child of this loop is a parent of lpChild.)
4949 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4950 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4951 // by following "lpChild" then "lpSibling" links.
4953 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4954 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4956 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4957 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4958 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4960 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4961 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4963 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4964 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4965 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4966 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4968 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
4969 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
4970 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
4972 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
4973 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
4974 // type are assigned to.
4976 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
4977 // If this is set, the fields below may not be accurate (since they become irrelevant.)
4978 bool lpContainsCall; // True if executing the loop body *may* execute a call
4980 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
4981 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
4983 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
4985 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
4986 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
4988 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
4990 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
4991 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
4993 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
4994 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
4996 JitSimplerHashBehavior>
4998 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
4999 // instance fields modified
5002 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5003 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5005 JitSimplerHashBehavior>
5007 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5008 // arrays of that type are modified
5011 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5012 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5014 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5015 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5016 // (shifted left, with a low-order bit set to distinguish.)
5017 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5018 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5020 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5022 GenTreePtr lpIterTree; // The "i <op>= const" tree
5023 unsigned lpIterVar(); // iterator variable #
5024 int lpIterConst(); // the constant with which the iterator is incremented
5025 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5026 void VERIFY_lpIterTree();
5028 var_types lpIterOperType(); // For overflow instructions
5031 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5032 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5036 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5038 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5039 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5040 void VERIFY_lpTestTree();
5042 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5043 GenTreePtr lpIterator(); // the iterator node in the loop test
5044 GenTreePtr lpLimit(); // the limit node in the loop test
5046 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5047 // LPFLG_CONST_LIMIT
5048 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5050 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5051 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5052 // LPFLG_ARRLEN_LIMIT
5054 // Returns "true" iff "*this" contains the blk.
5055 bool lpContains(BasicBlock* blk)
5057 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5059 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5060 // to be equal, but requiring bottoms to be different.)
5061 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5063 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5066 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5067 // bottoms to be different.)
5068 bool lpContains(const LoopDsc& lp2)
5070 return lpContains(lp2.lpFirst, lp2.lpBottom);
5073 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5074 // (allowing firsts to be equal, but requiring bottoms to be different.)
5075 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5077 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5080 // Returns "true" iff "*this" is (properly) contained by "lp2"
5081 // (allowing firsts to be equal, but requiring bottoms to be different.)
5082 bool lpContainedBy(const LoopDsc& lp2)
5084 return lpContains(lp2.lpFirst, lp2.lpBottom);
5087 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5088 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5090 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5092 // Returns "true" iff "*this" is disjoint from "lp2".
5093 bool lpDisjoint(const LoopDsc& lp2)
5095 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5097 // Returns "true" iff the loop is well-formed (see code for defn).
5100 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5101 lpEntry->bbNum <= lpBottom->bbNum &&
5102 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5107 bool fgMightHaveLoop(); // returns true if there are any backedges
5108 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5111 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5112 unsigned char optLoopCount; // number of tracked loops
5115 unsigned optCallCount; // number of calls made in the method
5116 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5117 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5118 unsigned optLoopsCloned; // number of loops cloned in the current method.
5121 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5122 void optPrintLoopInfo(unsigned loopNum,
5124 BasicBlock* lpFirst,
5126 BasicBlock* lpEntry,
5127 BasicBlock* lpBottom,
5128 unsigned char lpExitCnt,
5130 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5131 void optPrintLoopInfo(unsigned lnum);
5132 void optPrintLoopRecording(unsigned lnum);
5134 void optCheckPreds();
5137 void optSetBlockWeights();
5139 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5141 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5143 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5145 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5146 unsigned optIsLoopIncrTree(GenTreePtr incr);
5147 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5148 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5149 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5150 bool optExtractInitTestIncr(BasicBlock* head,
5155 GenTreePtr* ppIncr);
5157 void optRecordLoop(BasicBlock* head,
5163 unsigned char exitCnt);
5165 void optFindNaturalLoops();
5167 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5168 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5169 bool optCanonicalizeLoopNest(unsigned char loopInd);
5171 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5172 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5173 bool optCanonicalizeLoop(unsigned char loopInd);
5175 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5176 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5177 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5178 bool optLoopContains(unsigned l1, unsigned l2);
5180 // Requires "loopInd" to be a valid index into the loop table.
5181 // Updates the loop table by changing loop "loopInd", whose head is required
5182 // to be "from", to be "to". Also performs this transformation for any
5183 // loop nested in "loopInd" that shares the same head as "loopInd".
5184 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5186 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5187 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5188 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5190 // Marks the containsCall information to "lnum" and any parent loops.
5191 void AddContainsCallAllContainingLoops(unsigned lnum);
5192 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5193 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5194 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5195 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5196 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5197 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5199 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5200 // of "from".) Copies the jump destination from "from" to "to".
5201 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5203 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5204 unsigned optLoopDepth(unsigned lnum)
5206 unsigned par = optLoopTable[lnum].lpParent;
5207 if (par == BasicBlock::NOT_IN_LOOP)
5213 return 1 + optLoopDepth(par);
5217 void fgOptWhileLoop(BasicBlock* block);
5219 bool optComputeLoopRep(int constInit,
5222 genTreeOps iterOper,
5224 genTreeOps testOper,
5227 unsigned* iterCount);
5228 #if FEATURE_STACK_FP_X87
5231 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5232 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5233 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5234 #endif // FEATURE_STACK_FP_X87
5237 static fgWalkPreFn optIsVarAssgCB;
5240 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5242 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5244 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5246 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5248 /**************************************************************************
5249 * Optimization conditions
5250 *************************************************************************/
5252 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5253 bool optPentium4(void);
5254 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5255 bool optAvoidIntMult(void);
5260 // The following is the upper limit on how many expressions we'll keep track
5261 // of for the CSE analysis.
5263 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5265 static const int MIN_CSE_COST = 2;
5267 // Keeps tracked cse indices
5268 BitVecTraits* cseTraits;
5272 /* Generic list of nodes - used by the CSE logic */
5280 typedef struct treeLst* treeLstPtr;
5284 treeStmtLst* tslNext;
5285 GenTreePtr tslTree; // tree node
5286 GenTreePtr tslStmt; // statement containing the tree
5287 BasicBlock* tslBlock; // block containing the statement
5290 typedef struct treeStmtLst* treeStmtLstPtr;
5292 // The following logic keeps track of expressions via a simple hash table.
5296 CSEdsc* csdNextInBucket; // used by the hash table
5298 unsigned csdHashValue; // the orginal hashkey
5300 unsigned csdIndex; // 1..optCSECandidateCount
5301 char csdLiveAcrossCall; // 0 or 1
5303 unsigned short csdDefCount; // definition count
5304 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5306 unsigned csdDefWtCnt; // weighted def count
5307 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5309 GenTreePtr csdTree; // treenode containing the 1st occurance
5310 GenTreePtr csdStmt; // stmt containing the 1st occurance
5311 BasicBlock* csdBlock; // block containing the 1st occurance
5313 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5314 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5317 static const size_t s_optCSEhashSize;
5318 CSEdsc** optCSEhash;
5323 CSEdsc* optCSEfindDsc(unsigned index);
5324 void optUnmarkCSE(GenTreePtr tree);
5326 // user defined callback data for the tree walk function optCSE_MaskHelper()
5327 struct optCSE_MaskData
5329 EXPSET_TP CSE_defMask;
5330 EXPSET_TP CSE_useMask;
5333 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5334 static fgWalkPreFn optCSE_MaskHelper;
5336 // This function walks all the node for an given tree
5337 // and return the mask of CSE definitions and uses for the tree
5339 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5341 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5342 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5343 bool optCSE_canSwap(GenTree* tree);
5345 static fgWalkPostFn optPropagateNonCSE;
5346 static fgWalkPreFn optHasNonCSEChild;
5348 static fgWalkPreFn optUnmarkCSEs;
5350 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5351 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5353 void optCleanupCSEs();
5356 void optEnsureClearCSEInfo();
5359 #endif // FEATURE_ANYCSE
5361 #if FEATURE_VALNUM_CSE
5362 /**************************************************************************
5363 * Value Number based CSEs
5364 *************************************************************************/
5367 void optOptimizeValnumCSEs();
5370 void optValnumCSE_Init();
5371 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5372 unsigned optValnumCSE_Locate();
5373 void optValnumCSE_InitDataFlow();
5374 void optValnumCSE_DataFlow();
5375 void optValnumCSE_Availablity();
5376 void optValnumCSE_Heuristic();
5377 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5379 #endif // FEATURE_VALNUM_CSE
5382 bool optDoCSE; // True when we have found a duplicate CSE tree
5383 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5384 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5385 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5386 unsigned optCSEstart; // The first local variable number that is a CSE
5387 unsigned optCSEcount; // The total count of CSE's introduced.
5388 unsigned optCSEweight; // The weight of the current block when we are
5389 // scanning for CSE expressions
5391 bool optIsCSEcandidate(GenTreePtr tree);
5393 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5395 bool lclNumIsTrueCSE(unsigned lclNum) const
5397 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5400 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5402 bool lclNumIsCSE(unsigned lclNum) const
5404 return lvaTable[lclNum].lvIsCSE;
5408 bool optConfigDisableCSE();
5409 bool optConfigDisableCSE2();
5411 void optOptimizeCSEs();
5413 #endif // FEATURE_ANYCSE
5421 unsigned ivaVar; // Variable we are interested in, or -1
5422 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5423 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5424 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5425 callInterf ivaMaskCall; // What kind of calls are there?
5428 static callInterf optCallInterf(GenTreePtr call);
5431 // VN based copy propagation.
5432 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5433 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5434 LclNumToGenTreePtrStack;
5436 // Kill set to track variables with intervening definitions.
5437 VARSET_TP optCopyPropKillSet;
5439 // Copy propagation functions.
5440 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5441 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5442 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5443 bool optIsSsaLocal(GenTreePtr tree);
5444 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5445 void optVnCopyProp();
5447 /**************************************************************************
5448 * Early value propagation
5449 *************************************************************************/
5455 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5459 static unsigned GetHashCode(SSAName ssaNm)
5461 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5464 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5466 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5470 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5471 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5472 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5473 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5474 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5476 unsigned optMethodFlags;
5478 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5479 // No throughput diff was found with backward walk bound between 3-8.
5480 static const int optEarlyPropRecurBound = 5;
5482 enum class optPropKind
5490 bool gtIsVtableRef(GenTreePtr tree);
5491 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5492 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5493 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5494 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5495 bool optEarlyPropRewriteTree(GenTreePtr tree);
5496 bool optDoEarlyPropForBlock(BasicBlock* block);
5497 bool optDoEarlyPropForFunc();
5498 void optEarlyProp();
5499 void optFoldNullCheck(GenTreePtr tree);
5500 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5503 /**************************************************************************
5504 * Value/Assertion propagation
5505 *************************************************************************/
5507 // Data structures for assertion prop
5508 BitVecTraits* apTraits;
5512 enum optAssertionKind
5527 O1K_ARRLEN_OPER_BND,
5528 O1K_ARRLEN_LOOP_BND,
5529 O1K_CONSTANT_LOOP_BND,
5550 optAssertionKind assertionKind;
5553 unsigned lclNum; // assigned to or property of this local var number
5561 struct AssertionDscOp1
5563 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5570 struct AssertionDscOp2
5572 optOp2Kind kind; // a const or copy assignment
5576 ssize_t iconVal; // integer
5577 unsigned iconFlags; // gtFlags
5579 struct Range // integer subrange
5593 bool IsArrLenArithBound()
5595 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5597 bool IsArrLenBound()
5599 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5601 bool IsConstantBound()
5603 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5604 op1.kind == O1K_CONSTANT_LOOP_BND);
5606 bool IsBoundsCheckNoThrow()
5608 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5611 bool IsCopyAssertion()
5613 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5616 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5618 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5619 a1->op2.kind == a2->op2.kind;
5622 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5624 if (kind == OAK_EQUAL)
5626 return kind2 == OAK_NOT_EQUAL;
5628 else if (kind == OAK_NOT_EQUAL)
5630 return kind2 == OAK_EQUAL;
5635 static ssize_t GetLowerBoundForIntegralType(var_types type)
5655 static ssize_t GetUpperBoundForIntegralType(var_types type)
5679 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5681 return (op1.kind == that->op1.kind) &&
5682 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5685 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5687 if (op2.kind != that->op2.kind)
5693 case O2K_IND_CNS_INT:
5695 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5697 case O2K_CONST_LONG:
5698 return (op2.lconVal == that->op2.lconVal);
5700 case O2K_CONST_DOUBLE:
5701 // exact match because of positive and negative zero.
5702 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5704 case O2K_LCLVAR_COPY:
5706 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5707 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5710 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5713 // we will return false
5717 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5723 bool Complementary(AssertionDsc* that, bool vnBased)
5725 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5726 HasSameOp2(that, vnBased);
5729 bool Equals(AssertionDsc* that, bool vnBased)
5731 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5735 typedef unsigned short AssertionIndex;
5738 static fgWalkPreFn optAddCopiesCallback;
5739 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5740 unsigned optAddCopyLclNum;
5741 GenTreePtr optAddCopyAsgnNode;
5743 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5744 bool optAssertionPropagated; // set to true if we modified the trees
5745 bool optAssertionPropagatedCurrentStmt;
5747 GenTreePtr optAssertionPropCurrentTree;
5749 AssertionIndex* optComplementaryAssertionMap;
5750 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5751 // using the value of a local var) for each local var
5752 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5753 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5754 AssertionIndex optMaxAssertionCount;
5757 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5758 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5759 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5760 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5761 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5762 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5764 AssertionIndex GetAssertionCount()
5766 return optAssertionCount;
5768 ASSERT_TP* bbJtrueAssertionOut;
5769 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5770 ValueNumToAssertsMap;
5771 ValueNumToAssertsMap* optValueNumToAsserts;
5773 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5775 // Assertion prop helpers.
5776 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5777 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5778 void optAssertionInit(bool isLocalProp);
5779 void optAssertionTraitsInit(AssertionIndex assertionCount);
5780 #if LOCAL_ASSERTION_PROP
5781 void optAssertionReset(AssertionIndex limit);
5782 void optAssertionRemove(AssertionIndex index);
5785 // Assertion prop data flow functions.
5786 void optAssertionPropMain();
5787 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5788 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5789 ASSERT_TP* optInitAssertionDataflowFlags();
5790 ASSERT_TP* optComputeAssertionGen();
5792 // Assertion Gen functions.
5793 void optAssertionGen(GenTreePtr tree);
5794 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5795 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5796 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5797 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5798 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5799 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5801 // Assertion creation functions.
5802 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5803 AssertionIndex optCreateAssertion(GenTreePtr op1,
5805 optAssertionKind assertionKind,
5806 AssertionDsc* assertion);
5807 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5809 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5810 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5811 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5813 void optPrintVnAssertionMapping();
5815 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5817 // Used for respective assertion propagations.
5818 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5819 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5820 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5821 bool optAssertionIsNonNull(GenTreePtr op,
5822 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5824 // Used for Relop propagation.
5825 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5826 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5827 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5829 // Assertion prop for lcl var functions.
5830 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5831 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5833 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5834 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5835 const GenTreePtr tree,
5836 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5837 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5839 // Assertion propagation functions.
5840 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5841 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5842 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5843 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5844 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5845 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5846 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5847 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5848 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5849 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5850 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5851 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5853 // Implied assertion functions.
5854 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5855 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5856 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5857 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5859 ASSERT_VALRET_TP optNewFullAssertSet();
5860 ASSERT_VALRET_TP optNewEmptyAssertSet();
5863 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5864 void optDebugCheckAssertion(AssertionDsc* assertion);
5865 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5867 void optAddCopies();
5868 #endif // ASSERTION_PROP
5870 /**************************************************************************
5872 *************************************************************************/
5875 struct LoopCloneVisitorInfo
5877 LoopCloneContext* context;
5880 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5881 : context(context), loopNum(loopNum), stmt(nullptr)
5886 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5887 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5888 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5889 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5890 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5891 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5892 void optObtainLoopCloningOpts(LoopCloneContext* context);
5893 bool optIsLoopClonable(unsigned loopInd);
5895 bool optCanCloneLoops();
5898 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5900 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5901 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5902 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5903 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5907 void optInsertLoopCloningStress(BasicBlock* head);
5909 #if COUNT_RANGECHECKS
5910 static unsigned optRangeChkRmv;
5911 static unsigned optRangeChkAll;
5920 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5925 RngChkDsc* rcdNextInBucket; // used by the hash table
5927 unsigned short rcdHashValue; // to make matching faster
5928 unsigned short rcdIndex; // 0..optRngChkCount-1
5930 GenTreePtr rcdTree; // the array index tree
5933 unsigned optRngChkCount;
5934 static const size_t optRngChkHashSize;
5936 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5937 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5939 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5942 bool optLoopsMarked;
5945 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5946 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5950 XX Does the register allocation and puts the remaining lclVars on the stack XX
5952 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5953 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5957 #ifndef LEGACY_BACKEND
5962 #else // LEGACY_BACKEND
5967 #endif // LEGACY_BACKEND
5969 #ifdef LEGACY_BACKEND
5971 void raAssignVars(); // register allocation
5972 #endif // LEGACY_BACKEND
5974 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
5976 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
5978 void raMarkStkVars();
5981 // Some things are used by both LSRA and regpredict allocators.
5983 FrameType rpFrameType;
5984 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
5986 #ifdef LEGACY_BACKEND
5987 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
5989 #endif // LEGACY_BACKEND
5991 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
5993 #if FEATURE_FP_REGALLOC
5994 enum enumConfigRegisterFP
5996 CONFIG_REGISTER_FP_NONE = 0x0,
5997 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
5998 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
5999 CONFIG_REGISTER_FP_FULL = 0x3,
6001 enumConfigRegisterFP raConfigRegisterFP();
6002 #endif // FEATURE_FP_REGALLOC
6005 regMaskTP raConfigRestrictMaskFP();
6008 #ifndef LEGACY_BACKEND
6009 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6010 #else // LEGACY_BACKEND
6011 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6012 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6013 bool raNewBlocks; // True is we added killing blocks for FPU registers
6014 unsigned rpPasses; // Number of passes made by the register predicter
6015 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6016 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6017 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6018 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6019 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6020 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6021 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6022 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6023 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6024 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6025 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6026 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6028 bool rpRegAllocDone; // Set to true after we have completed register allocation
6030 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6032 void raSetupArgMasks(RegState* r);
6034 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6036 void raDumpVarIntf(); // Dump the variable to variable interference graph
6037 void raDumpRegIntf(); // Dump the variable to register interference graph
6039 void raAdjustVarIntf();
6041 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6043 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6045 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6046 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6048 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6050 static fgWalkPreFn rpMarkRegIntf;
6052 regMaskTP rpPredictAddressMode(
6053 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6055 void rpPredictRefAssign(unsigned lclNum);
6057 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6059 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6061 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6063 void rpPredictRegUse(); // Entry point
6065 unsigned raPredictTreeRegUse(GenTreePtr tree);
6066 unsigned raPredictListRegUse(GenTreePtr list);
6068 void raSetRegVarOrder(var_types regType,
6069 regNumber* customVarOrder,
6070 unsigned* customVarOrderSize,
6072 regMaskTP avoidReg);
6074 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6075 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6076 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6077 void raAddToStkPredict(unsigned val)
6079 unsigned newStkPredict = rpStkPredict + val;
6080 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6081 rpStkPredict = UINT_MAX - 1;
6083 rpStkPredict = newStkPredict;
6087 #if !FEATURE_FP_REGALLOC
6088 void raDispFPlifeInfo();
6092 regMaskTP genReturnRegForTree(GenTreePtr tree);
6093 #endif // LEGACY_BACKEND
6095 /* raIsVarargsStackArg is called by raMaskStkVars and by
6096 lvaSortByRefCount. It identifies the special case
6097 where a varargs function has a parameter passed on the
6098 stack, other than the special varargs handle. Such parameters
6099 require special treatment, because they cannot be tracked
6100 by the GC (their offsets in the stack are not known
6104 bool raIsVarargsStackArg(unsigned lclNum)
6108 LclVarDsc* varDsc = &lvaTable[lclNum];
6110 assert(varDsc->lvIsParam);
6112 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6114 #else // _TARGET_X86_
6118 #endif // _TARGET_X86_
6121 #ifdef LEGACY_BACKEND
6122 // Records the current prediction, if it's better than any previous recorded prediction.
6123 void rpRecordPrediction();
6124 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6125 void rpUseRecordedPredictionIfBetter();
6127 // Data members used in the methods above.
6128 unsigned rpBestRecordedStkPredict;
6129 struct VarRegPrediction
6131 bool m_isEnregistered;
6132 regNumberSmall m_regNum;
6133 regNumberSmall m_otherReg;
6135 VarRegPrediction* rpBestRecordedPrediction;
6136 #endif // LEGACY_BACKEND
6139 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6140 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6144 XX Get to the class and method info from the Execution Engine given XX
6145 XX tokens for the class and method XX
6147 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6148 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6152 /* These are the different addressing modes used to access a local var.
6153 * The JIT has to report the location of the locals back to the EE
6154 * for debugging purposes.
6160 VLT_REG_BYREF, // this type is currently only used for value types on X64
6163 VLT_STK_BYREF, // this type is currently only used for value types on X64
6177 siVarLocType vlType;
6180 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6182 // VLT_REG_BYREF -- the specified register contains the address of the variable
6190 // VLT_STK -- Any 32 bit value which is on the stack
6191 // eg. [ESP+0x20], or [EBP-0x28]
6192 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6193 // eg. mov EAX, [ESP+0x20]; [EAX]
6197 regNumber vlsBaseReg;
6198 NATIVE_OFFSET vlsOffset;
6201 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6210 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6211 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6219 regNumber vlrssBaseReg;
6220 NATIVE_OFFSET vlrssOffset;
6224 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6225 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6231 regNumber vlsrsBaseReg;
6232 NATIVE_OFFSET vlsrsOffset;
6238 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6239 // eg 2 DWords at [ESP+0x10]
6243 regNumber vls2BaseReg;
6244 NATIVE_OFFSET vls2Offset;
6247 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6248 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6255 // VLT_FIXED_VA -- fixed argument of a varargs function.
6256 // The argument location depends on the size of the variable
6257 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6258 // location of the first arg. This argument can then be accessed
6259 // relative to the position of the first arg
6263 unsigned vlfvOffset;
6270 void* rpValue; // pointer to the in-process
6271 // location of the value.
6277 bool vlIsInReg(regNumber reg);
6278 bool vlIsOnStk(regNumber reg, signed offset);
6281 /*************************************************************************/
6286 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6287 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6288 CORINFO_CALLINFO_FLAGS flags,
6289 CORINFO_CALL_INFO* pResult);
6290 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6292 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6293 CORINFO_ACCESS_FLAGS flags,
6294 CORINFO_FIELD_INFO* pResult);
6298 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6300 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6302 bool IsSuperPMIException(unsigned code)
6304 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6306 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6307 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6308 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6309 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6310 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6311 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6312 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6313 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6317 case EXCEPTIONCODE_DebugBreakorAV:
6318 case EXCEPTIONCODE_MC:
6319 case EXCEPTIONCODE_LWM:
6320 case EXCEPTIONCODE_SASM:
6321 case EXCEPTIONCODE_SSYM:
6322 case EXCEPTIONCODE_CALLUTILS:
6323 case EXCEPTIONCODE_TYPEUTILS:
6324 case EXCEPTIONCODE_ASSERT:
6331 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6332 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6334 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6335 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6338 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6339 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6340 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6342 // VOM info, method sigs
6344 void eeGetSig(unsigned sigTok,
6345 CORINFO_MODULE_HANDLE scope,
6346 CORINFO_CONTEXT_HANDLE context,
6347 CORINFO_SIG_INFO* retSig);
6349 void eeGetCallSiteSig(unsigned sigTok,
6350 CORINFO_MODULE_HANDLE scope,
6351 CORINFO_CONTEXT_HANDLE context,
6352 CORINFO_SIG_INFO* retSig);
6354 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6356 // Method entry-points, instrs
6358 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6360 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6362 CORINFO_EE_INFO eeInfo;
6363 bool eeInfoInitialized;
6365 CORINFO_EE_INFO* eeGetEEInfo();
6367 // Gets the offset of a SDArray's first element
6368 unsigned eeGetArrayDataOffset(var_types type);
6369 // Gets the offset of a MDArray's first element
6370 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6372 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6374 // Returns the page size for the target machine as reported by the EE.
6375 inline size_t eeGetPageSize()
6377 #if COR_JIT_EE_VERSION > 460
6378 return eeGetEEInfo()->osPageSize;
6379 #else // COR_JIT_EE_VERSION <= 460
6380 return CORINFO_PAGE_SIZE;
6381 #endif // COR_JIT_EE_VERSION > 460
6384 // Returns the frame size at which we will generate a loop to probe the stack.
6385 inline size_t getVeryLargeFrameSize()
6388 // The looping probe code is 40 bytes, whereas the straight-line probing for
6389 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6390 // or greater, to generate smaller code.
6391 return 2 * eeGetPageSize();
6393 return 3 * eeGetPageSize();
6397 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6399 #if COR_JIT_EE_VERSION > 460
6400 return eeGetEEInfo()->targetAbi == abi;
6402 return CORINFO_DESKTOP_ABI == abi;
6406 inline bool generateCFIUnwindCodes()
6408 #ifdef UNIX_AMD64_ABI
6409 return IsTargetAbi(CORINFO_CORERT_ABI);
6417 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6419 // Debugging support - Line number info
6421 void eeGetStmtOffsets();
6423 unsigned eeBoundariesCount;
6425 struct boundariesDsc
6427 UNATIVE_OFFSET nativeIP;
6429 unsigned sourceReason;
6430 } * eeBoundaries; // Boundaries to report to EE
6431 void eeSetLIcount(unsigned count);
6432 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6436 static void eeDispILOffs(IL_OFFSET offs);
6437 static void eeDispLineInfo(const boundariesDsc* line);
6438 void eeDispLineInfos();
6441 // Debugging support - Local var info
6445 unsigned eeVarsCount;
6447 struct VarResultInfo
6449 UNATIVE_OFFSET startOffset;
6450 UNATIVE_OFFSET endOffset;
6454 void eeSetLVcount(unsigned count);
6455 void eeSetLVinfo(unsigned which,
6456 UNATIVE_OFFSET startOffs,
6457 UNATIVE_OFFSET length,
6462 const siVarLoc& loc);
6466 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6467 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6470 // ICorJitInfo wrappers
6472 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6474 void eeAllocUnwindInfo(BYTE* pHotCode,
6480 CorJitFuncKind funcKind);
6482 void eeSetEHcount(unsigned cEH);
6484 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6486 WORD eeGetRelocTypeHint(void* target);
6488 // ICorStaticInfo wrapper functions
6490 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6492 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6494 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6497 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6498 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6499 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6500 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6502 template <typename ParamType>
6503 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6505 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6508 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6510 // Utility functions
6512 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6515 const wchar_t* eeGetCPString(size_t stringHandle);
6518 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6520 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6521 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6523 static fgWalkPreFn CountSharedStaticHelper;
6524 static bool IsSharedStaticHelper(GenTreePtr tree);
6525 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6527 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6528 // returns true/false if 'field' is a Jit Data offset
6529 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6530 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6531 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6533 /*****************************************************************************/
6538 enum TEMP_USAGE_TYPE
6544 static var_types tmpNormalizeType(var_types type);
6545 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6546 void tmpRlsTemp(TempDsc* temp);
6547 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6550 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6551 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6555 bool tmpAllFree() const;
6558 #ifndef LEGACY_BACKEND
6559 void tmpPreAllocateTemps(var_types type, unsigned count);
6560 #endif // !LEGACY_BACKEND
6563 #ifdef LEGACY_BACKEND
6564 unsigned tmpIntSpillMax; // number of int-sized spill temps
6565 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6566 #endif // LEGACY_BACKEND
6568 unsigned tmpCount; // Number of temps
6569 unsigned tmpSize; // Size of all the temps
6572 // Used by RegSet::rsSpillChk()
6573 unsigned tmpGetCount; // Temps which haven't been released yet
6576 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6578 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6579 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6582 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6583 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6587 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6588 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6592 CodeGenInterface* codeGen;
6594 // The following holds information about instr offsets in terms of generated code.
6598 IPmappingDsc* ipmdNext; // next line# record
6599 IL_OFFSETX ipmdILoffsx; // the instr offset
6600 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6601 bool ipmdIsLabel; // Can this code be a branch label?
6604 // Record the instr offset mapping to the generated code
6606 IPmappingDsc* genIPmappingList;
6607 IPmappingDsc* genIPmappingLast;
6609 // Managed RetVal - A side hash table meant to record the mapping from a
6610 // GT_CALL node to its IL offset. This info is used to emit sequence points
6611 // that can be used by debugger to determine the native offset at which the
6612 // managed RetVal will be available.
6614 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6615 // favor of a side table for two reasons: 1) We need IL offset for only those
6616 // GT_CALL nodes (created during importation) that correspond to an IL call and
6617 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6618 // structure and IL offset is needed only when generating debuggable code. Therefore
6619 // it is desirable to avoid memory size penalty in retail scenarios.
6620 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6621 CallSiteILOffsetTable;
6622 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6624 unsigned genReturnLocal; // Local number for the return value when applicable.
6625 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6627 // The following properties are part of CodeGenContext. Getters are provided here for
6628 // convenience and backward compatibility, but the properties can only be set by invoking
6629 // the setter on CodeGenContext directly.
6631 __declspec(property(get = getEmitter)) emitter* genEmitter;
6632 emitter* getEmitter()
6634 return codeGen->getEmitter();
6637 const bool isFramePointerUsed()
6639 return codeGen->isFramePointerUsed();
6642 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6643 bool getInterruptible()
6645 return codeGen->genInterruptible;
6647 void setInterruptible(bool value)
6649 codeGen->setInterruptible(value);
6653 const bool genDoubleAlign()
6655 return codeGen->doDoubleAlign();
6657 DWORD getCanDoubleAlign();
6658 bool shouldDoubleAlign(unsigned refCntStk,
6660 unsigned refCntWtdReg,
6661 unsigned refCntStkParam,
6662 unsigned refCntWtdStkDbl);
6663 #endif // DOUBLE_ALIGN
6665 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6666 bool getFullPtrRegMap()
6668 return codeGen->genFullPtrRegMap;
6670 void setFullPtrRegMap(bool value)
6672 codeGen->setFullPtrRegMap(value);
6675 // Things that MAY belong either in CodeGen or CodeGenContext
6677 #if FEATURE_EH_FUNCLETS
6678 FuncInfoDsc* compFuncInfos;
6679 unsigned short compCurrFuncIdx;
6680 unsigned short compFuncInfoCount;
6682 unsigned short compFuncCount()
6684 assert(fgFuncletsCreated);
6685 return compFuncInfoCount;
6688 #else // !FEATURE_EH_FUNCLETS
6690 // This is a no-op when there are no funclets!
6691 void genUpdateCurrentFunclet(BasicBlock* block)
6696 FuncInfoDsc compFuncInfoRoot;
6698 static const unsigned compCurrFuncIdx = 0;
6700 unsigned short compFuncCount()
6705 #endif // !FEATURE_EH_FUNCLETS
6707 FuncInfoDsc* funCurrentFunc();
6708 void funSetCurrentFunc(unsigned funcIdx);
6709 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6710 unsigned int funGetFuncIdx(BasicBlock* block);
6714 VARSET_TP compCurLife; // current live variables
6715 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6717 template <bool ForCodeGen>
6718 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6720 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6722 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6725 template <bool ForCodeGen>
6726 void compUpdateLife(GenTreePtr tree);
6728 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6729 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6730 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6731 template <bool ForCodeGen>
6732 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6734 template <bool ForCodeGen>
6735 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6737 // Gets a register mask that represent the kill set for a helper call since
6738 // not all JIT Helper calls follow the standard ABI on the target architecture.
6739 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6741 // Gets a register mask that represent the kill set for a NoGC helper call.
6742 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6745 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6746 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6747 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6748 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6749 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6750 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6751 #endif // _TARGET_ARM_
6753 // 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
6755 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6757 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6758 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6759 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6760 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6761 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6762 // for the tracked var indices of the field vars, as in a live var set).
6763 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6765 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6767 if (m_promotedStructDeathVars == nullptr)
6769 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6771 return m_promotedStructDeathVars;
6775 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6776 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6780 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6781 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6784 #if !defined(__GNUC__)
6785 #pragma region Unwind information
6790 // Infrastructure functions: start/stop/reserve/emit.
6793 void unwindBegProlog();
6794 void unwindEndProlog();
6795 void unwindBegEpilog();
6796 void unwindEndEpilog();
6797 void unwindReserve();
6798 void unwindEmit(void* pHotCode, void* pColdCode);
6801 // Specific unwind information functions: called by code generation to indicate a particular
6802 // prolog or epilog unwindable instruction has been generated.
6805 void unwindPush(regNumber reg);
6806 void unwindAllocStack(unsigned size);
6807 void unwindSetFrameReg(regNumber reg, unsigned offset);
6808 void unwindSaveReg(regNumber reg, unsigned offset);
6810 #if defined(_TARGET_ARM_)
6811 void unwindPushMaskInt(regMaskTP mask);
6812 void unwindPushMaskFloat(regMaskTP mask);
6813 void unwindPopMaskInt(regMaskTP mask);
6814 void unwindPopMaskFloat(regMaskTP mask);
6815 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6816 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6817 // called via unwindPadding().
6818 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6819 // instruction and the current location.
6820 #endif // _TARGET_ARM_
6822 #if defined(_TARGET_ARM64_)
6824 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6825 // instruction and the current location.
6826 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6827 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6828 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6829 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6830 void unwindSaveNext(); // unwind code: save_next
6831 void unwindReturn(regNumber reg); // ret lr
6832 #endif // defined(_TARGET_ARM64_)
6835 // Private "helper" functions for the unwind implementation.
6839 #if FEATURE_EH_FUNCLETS
6840 void unwindGetFuncLocations(FuncInfoDsc* func,
6841 bool getHotSectionData,
6842 /* OUT */ emitLocation** ppStartLoc,
6843 /* OUT */ emitLocation** ppEndLoc);
6844 #endif // FEATURE_EH_FUNCLETS
6846 void unwindReserveFunc(FuncInfoDsc* func);
6847 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6849 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6851 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6852 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6854 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6856 #if defined(_TARGET_AMD64_)
6858 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6860 void unwindBegPrologWindows();
6861 void unwindPushWindows(regNumber reg);
6862 void unwindAllocStackWindows(unsigned size);
6863 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6864 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6866 #ifdef UNIX_AMD64_ABI
6867 void unwindBegPrologCFI();
6868 void unwindPushCFI(regNumber reg);
6869 void unwindAllocStackCFI(unsigned size);
6870 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6871 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6872 int mapRegNumToDwarfReg(regNumber reg);
6873 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6874 #endif // UNIX_AMD64_ABI
6875 #elif defined(_TARGET_ARM_)
6877 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6878 void unwindPushPopMaskFloat(regMaskTP mask);
6879 void unwindSplit(FuncInfoDsc* func);
6881 #endif // _TARGET_ARM_
6883 #if !defined(__GNUC__)
6884 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6888 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6889 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6893 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6894 XX that contains the distinguished, well-known SIMD type definitions). XX
6896 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6897 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6900 // Get highest available instruction set for floating point codegen
6901 InstructionSet getFloatingPointInstructionSet()
6903 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6906 return InstructionSet_AVX;
6911 return InstructionSet_SSE3_4;
6915 assert(canUseSSE2());
6916 return InstructionSet_SSE2;
6918 assert(!"getFPInstructionSet() is not implemented for target arch");
6920 return InstructionSet_NONE;
6924 // Get highest available instruction set for SIMD codegen
6925 InstructionSet getSIMDInstructionSet()
6927 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6928 return getFloatingPointInstructionSet();
6930 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6932 return InstructionSet_NONE;
6938 // Should we support SIMD intrinsics?
6941 // Have we identified any SIMD types?
6942 // This is currently used by struct promotion to avoid getting type information for a struct
6943 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6945 bool _usesSIMDTypes;
6946 bool usesSIMDTypes()
6948 return _usesSIMDTypes;
6950 void setUsesSIMDTypes(bool value)
6952 _usesSIMDTypes = value;
6955 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6956 // that require indexed access to the individual fields of the vector, which is not well supported
6957 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6958 unsigned lvaSIMDInitTempVarNum;
6961 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6962 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6963 CORINFO_CLASS_HANDLE SIMDIntHandle;
6964 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6965 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6966 CORINFO_CLASS_HANDLE SIMDShortHandle;
6967 CORINFO_CLASS_HANDLE SIMDByteHandle;
6968 CORINFO_CLASS_HANDLE SIMDLongHandle;
6969 CORINFO_CLASS_HANDLE SIMDUIntHandle;
6970 CORINFO_CLASS_HANDLE SIMDULongHandle;
6971 CORINFO_CLASS_HANDLE SIMDVector2Handle;
6972 CORINFO_CLASS_HANDLE SIMDVector3Handle;
6973 CORINFO_CLASS_HANDLE SIMDVector4Handle;
6974 CORINFO_CLASS_HANDLE SIMDVectorHandle;
6976 // Get the handle for a SIMD type.
6977 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
6979 if (simdBaseType == TYP_FLOAT)
6984 return SIMDVector2Handle;
6986 return SIMDVector3Handle;
6988 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
6990 return SIMDVector4Handle;
6999 assert(simdType == getSIMDVectorType());
7000 switch (simdBaseType)
7003 return SIMDFloatHandle;
7005 return SIMDDoubleHandle;
7007 return SIMDIntHandle;
7009 return SIMDUShortHandle;
7011 return SIMDUShortHandle;
7013 return SIMDUByteHandle;
7015 return SIMDShortHandle;
7017 return SIMDByteHandle;
7019 return SIMDLongHandle;
7021 return SIMDUIntHandle;
7023 return SIMDULongHandle;
7025 assert(!"Didn't find a class handle for simdType");
7027 return NO_CLASS_HANDLE;
7031 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7032 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7033 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7035 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7036 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7037 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7038 bool isSIMDTypeLocal(GenTree* tree)
7040 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7043 // Returns true if the type of the tree is a byref of TYP_SIMD
7044 bool isAddrOfSIMDType(GenTree* tree)
7046 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7048 switch (tree->OperGet())
7051 return varTypeIsSIMD(tree->gtGetOp1());
7053 case GT_LCL_VAR_ADDR:
7054 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7057 return isSIMDTypeLocal(tree);
7064 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7066 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7067 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7068 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7071 // Returns base type of a TYP_SIMD local.
7072 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7073 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7075 if (isSIMDTypeLocal(tree))
7077 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7083 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7085 return info.compCompHnd->isInSIMDModule(clsHnd);
7088 bool isSIMDClass(typeInfo* pTypeInfo)
7090 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7093 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7094 // if it is not a SIMD type or is an unsupported base type.
7095 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7097 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7099 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7102 // Get SIMD Intrinsic info given the method handle.
7103 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7104 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7105 CORINFO_METHOD_HANDLE methodHnd,
7106 CORINFO_SIG_INFO* sig,
7109 var_types* baseType,
7110 unsigned* sizeBytes);
7112 // Pops and returns GenTree node from importers type stack.
7113 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7114 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7116 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7117 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7119 // Creates a GT_SIMD tree for Select operation
7120 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7122 unsigned simdVectorSize,
7127 // Creates a GT_SIMD tree for Min/Max operation
7128 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7129 CORINFO_CLASS_HANDLE typeHnd,
7131 unsigned simdVectorSize,
7135 // Transforms operands and returns the SIMD intrinsic to be applied on
7136 // transformed operands to obtain given relop result.
7137 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7138 CORINFO_CLASS_HANDLE typeHnd,
7139 unsigned simdVectorSize,
7140 var_types* baseType,
7144 // Creates a GT_SIMD tree for Abs intrinsic.
7145 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7147 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7148 // Transforms operands and returns the SIMD intrinsic to be applied on
7149 // transformed operands to obtain == comparison result.
7150 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7151 unsigned simdVectorSize,
7155 // Transforms operands and returns the SIMD intrinsic to be applied on
7156 // transformed operands to obtain > comparison result.
7157 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7158 unsigned simdVectorSize,
7162 // Transforms operands and returns the SIMD intrinsic to be applied on
7163 // transformed operands to obtain >= comparison result.
7164 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7165 unsigned simdVectorSize,
7169 // Transforms operands and returns the SIMD intrinsic to be applied on
7170 // transformed operands to obtain >= comparison result in case of int32
7171 // and small int base type vectors.
7172 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7173 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7174 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7176 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7177 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7178 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7179 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7180 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7182 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7183 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7184 GenTreePtr newobjThis,
7185 CORINFO_CLASS_HANDLE clsHnd,
7186 CORINFO_METHOD_HANDLE method,
7187 CORINFO_SIG_INFO* sig,
7190 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7192 // Whether SIMD vector occupies part of SIMD register.
7193 // SSE2: vector2f/3f are considered sub register SIMD types.
7194 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7195 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7197 unsigned sizeBytes = 0;
7198 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7199 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7202 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7204 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7207 // Get the type for the hardware SIMD vector.
7208 // This is the maximum SIMD type supported for this target.
7209 var_types getSIMDVectorType()
7211 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7218 assert(canUseSSE2());
7222 assert(!"getSIMDVectorType() unimplemented on target arch");
7227 // Get the size of the SIMD type in bytes
7228 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7230 unsigned sizeBytes = 0;
7231 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7235 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7236 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7238 // Get the the number of elements of basetype of SIMD vector given by its type handle
7239 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7241 // Get preferred alignment of SIMD type.
7242 int getSIMDTypeAlignment(var_types simdType);
7244 // Get the number of bytes in a SIMD Vector for the current compilation.
7245 unsigned getSIMDVectorRegisterByteLength()
7247 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7250 return YMM_REGSIZE_BYTES;
7254 assert(canUseSSE2());
7255 return XMM_REGSIZE_BYTES;
7258 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7263 // The minimum and maximum possible number of bytes in a SIMD vector.
7264 unsigned int maxSIMDStructBytes()
7266 return getSIMDVectorRegisterByteLength();
7268 unsigned int minSIMDStructBytes()
7270 return emitTypeSize(TYP_SIMD8);
7273 #ifdef FEATURE_AVX_SUPPORT
7274 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7275 static const unsigned maxPossibleSIMDStructBytes = 32;
7276 #else // !FEATURE_AVX_SUPPORT
7277 static const unsigned maxPossibleSIMDStructBytes = 16;
7278 #endif // !FEATURE_AVX_SUPPORT
7280 // Returns the codegen type for a given SIMD size.
7281 var_types getSIMDTypeForSize(unsigned size)
7283 var_types simdType = TYP_UNDEF;
7286 simdType = TYP_SIMD8;
7288 else if (size == 12)
7290 simdType = TYP_SIMD12;
7292 else if (size == 16)
7294 simdType = TYP_SIMD16;
7296 #ifdef FEATURE_AVX_SUPPORT
7297 else if (size == 32)
7299 simdType = TYP_SIMD32;
7301 #endif // FEATURE_AVX_SUPPORT
7304 noway_assert(!"Unexpected size for SIMD type");
7309 unsigned getSIMDInitTempVarNum()
7311 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7313 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7314 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7316 return lvaSIMDInitTempVarNum;
7319 #endif // FEATURE_SIMD
7322 //------------------------------------------------------------------------
7323 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7325 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7326 // candidate for enregistration.
7328 unsigned largestEnregisterableStructSize()
7331 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7332 if (vectorRegSize > TARGET_POINTER_SIZE)
7334 return vectorRegSize;
7337 #endif // FEATURE_SIMD
7339 return TARGET_POINTER_SIZE;
7344 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7345 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7346 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7348 // Is this var is of type simd struct?
7349 bool lclVarIsSIMDType(unsigned varNum)
7351 LclVarDsc* varDsc = lvaTable + varNum;
7352 return varDsc->lvIsSIMDType();
7355 // Is this Local node a SIMD local?
7356 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7358 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7361 // Returns true if the TYP_SIMD locals on stack are aligned at their
7362 // preferred byte boundary specified by getSIMDTypeAlignment().
7364 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7365 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7366 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7367 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7368 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7369 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7370 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7373 bool isSIMDTypeLocalAligned(unsigned varNum)
7375 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7376 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7379 int off = lvaFrameAddress(varNum, &ebpBased);
7380 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7381 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7382 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7385 #endif // FEATURE_SIMD
7390 // Whether SSE2 is available
7391 bool canUseSSE2() const
7393 #ifdef _TARGET_XARCH_
7394 return opts.compCanUseSSE2;
7400 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7401 bool CanUseSSE3_4() const
7403 #ifdef _TARGET_XARCH_
7404 return opts.compCanUseSSE3_4;
7410 bool canUseAVX() const
7412 #ifdef FEATURE_AVX_SUPPORT
7413 return opts.compCanUseAVX;
7420 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7421 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7425 XX Generic info about the compilation and the method being compiled. XX
7426 XX It is responsible for driving the other phases. XX
7427 XX It is also responsible for all the memory management. XX
7429 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7430 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7434 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7436 InlineResult* compInlineResult; // The result of importing the inlinee method.
7438 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7439 bool compJmpOpUsed; // Does the method do a JMP
7440 bool compLongUsed; // Does the method use TYP_LONG
7441 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7442 bool compTailCallUsed; // Does the method do a tailcall
7443 bool compLocallocUsed; // Does the method use localloc.
7444 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7445 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7446 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7448 // NOTE: These values are only reliable after
7449 // the importing is completely finished.
7451 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7452 // we can iterate over these efficiently.
7454 #if CPU_USES_BLOCK_MOVE
7455 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7459 // State information - which phases have completed?
7460 // These are kept together for easy discoverability
7462 bool bRangeAllowStress;
7463 bool compCodeGenDone;
7464 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7465 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7466 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7467 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7470 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7471 bool fgLocalVarLivenessChanged;
7473 bool compStackProbePrologDone;
7475 #ifndef LEGACY_BACKEND
7477 #endif // !LEGACY_BACKEND
7478 bool compRationalIRForm;
7480 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7482 bool compGeneratingProlog;
7483 bool compGeneratingEpilog;
7484 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7485 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7486 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7487 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7488 bool getNeedsGSSecurityCookie() const
7490 return compNeedsGSSecurityCookie;
7492 void setNeedsGSSecurityCookie()
7494 compNeedsGSSecurityCookie = true;
7497 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7498 // frame layout calculations, this is the level we are currently
7501 //---------------------------- JITing options -----------------------------
7514 JitFlags* jitFlags; // all flags passed from the EE
7515 unsigned compFlags; // method attributes
7517 codeOptimize compCodeOpt; // what type of code optimizations
7521 #ifdef _TARGET_XARCH_
7522 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7523 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7525 #ifdef FEATURE_AVX_SUPPORT
7526 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7527 #endif // FEATURE_AVX_SUPPORT
7528 #endif // _TARGET_XARCH_
7530 // optimize maximally and/or favor speed over size?
7532 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7533 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7534 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7535 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7536 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7538 // Maximun number of locals before turning off the inlining
7539 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7542 unsigned instrCount;
7543 unsigned lvRefCount;
7544 bool compMinOptsIsSet;
7546 bool compMinOptsIsUsed;
7548 inline bool MinOpts()
7550 assert(compMinOptsIsSet);
7551 compMinOptsIsUsed = true;
7554 inline bool IsMinOptsSet()
7556 return compMinOptsIsSet;
7559 inline bool MinOpts()
7563 inline bool IsMinOptsSet()
7565 return compMinOptsIsSet;
7568 inline void SetMinOpts(bool val)
7570 assert(!compMinOptsIsUsed);
7571 assert(!compMinOptsIsSet || (compMinOpts == val));
7573 compMinOptsIsSet = true;
7576 // true if the CLFLG_* for an optimization is set.
7577 inline bool OptEnabled(unsigned optFlag)
7579 return !!(compFlags & optFlag);
7582 #ifdef FEATURE_READYTORUN_COMPILER
7583 inline bool IsReadyToRun()
7585 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7588 inline bool IsReadyToRun()
7594 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7595 // PInvoke transitions inline (e.g. when targeting CoreRT).
7596 inline bool ShouldUsePInvokeHelpers()
7598 #if COR_JIT_EE_VERSION > 460
7599 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7605 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7607 inline bool IsReversePInvoke()
7609 #if COR_JIT_EE_VERSION > 460
7610 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7616 // true if we must generate code compatible with JIT32 quirks
7617 inline bool IsJit32Compat()
7619 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7620 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7626 // true if we must generate code compatible with Jit64 quirks
7627 inline bool IsJit64Compat()
7629 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7630 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7631 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7638 bool compScopeInfo; // Generate the LocalVar info ?
7639 bool compDbgCode; // Generate debugger-friendly code?
7640 bool compDbgInfo; // Gather debugging info?
7643 #ifdef PROFILING_SUPPORTED
7644 bool compNoPInvokeInlineCB;
7646 static const bool compNoPInvokeInlineCB;
7650 bool compGcChecks; // Check arguments and return values to ensure they are sane
7651 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7652 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7656 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7657 // to be allocated on the stack.
7658 // It will be set to true in the following cases:
7659 // 1. When the method being compiled has a declarative security
7660 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7661 // This is also the case when we inject a prolog and epilog in the method.
7663 // 2. When the method being compiled has imperative security (i.e. the method
7664 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7666 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7668 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7669 // which gets reported as a GC root to stackwalker.
7670 // (See also ICodeManager::GetAddrOfSecurityObject.)
7677 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7678 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7682 #ifdef UNIX_AMD64_ABI
7683 // This flag is indicating if there is a need to align the frame.
7684 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7685 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7686 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7687 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7688 // there are calls and making sure the frame alignment logic is executed.
7689 bool compNeedToAlignFrame;
7690 #endif // UNIX_AMD64_ABI
7692 bool compProcedureSplitting; // Separate cold code from hot code
7694 bool genFPorder; // Preserve FP order (operations are non-commutative)
7695 bool genFPopt; // Can we do frame-pointer-omission optimization?
7696 bool altJit; // True if we are an altjit and are compiling this method
7699 bool optRepeat; // Repeat optimizer phases k times
7700 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7701 bool dspCode; // Display native code generated
7702 bool dspEHTable; // Display the EH table reported to the VM
7703 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7704 bool dspEmit; // Display emitter output
7705 bool dspLines; // Display source-code lines intermixed with native code output
7706 bool dmpHex; // Display raw bytes in hex of native code output
7707 bool varNames; // Display variables names in native code output
7708 bool disAsm; // Display native code as it is generated
7709 bool disAsmSpilled; // Display native code when any register spilling occurs
7710 bool disDiffable; // Makes the Disassembly code 'diff-able'
7711 bool disAsm2; // Display native code after it is generated using external disassembler
7712 bool dspOrder; // Display names of each of the methods that we ngen/jit
7713 bool dspUnwind; // Display the unwind info output
7714 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7715 bool compLongAddress; // Force using large pseudo instructions for long address
7716 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7717 bool dspGCtbls; // Display the GC tables
7721 bool doLateDisasm; // Run the late disassembler
7722 #endif // LATE_DISASM
7724 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7725 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7726 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7727 static const bool dspGCtbls = true;
7730 // We need stack probes to guarantee that we won't trigger a stack overflow
7731 // when calling unmanaged code until they get a chance to set up a frame, because
7732 // the EE will have no idea where it is.
7734 // We will only be doing this currently for hosted environments. Unfortunately
7735 // we need to take care of stubs, so potentially, we will have to do the probes
7736 // for any call. We have a plan for not needing for stubs though
7737 bool compNeedStackProbes;
7739 #ifdef PROFILING_SUPPORTED
7740 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7741 // This option helps make the JIT behave as if it is running under a profiler.
7742 bool compJitELTHookEnabled;
7743 #endif // PROFILING_SUPPORTED
7745 #if FEATURE_TAILCALL_OPT
7746 // Whether opportunistic or implicit tail call optimization is enabled.
7747 bool compTailCallOpt;
7748 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7749 bool compTailCallLoopOpt;
7753 static const bool compUseSoftFP = true;
7754 #else // !ARM_SOFTFP
7755 static const bool compUseSoftFP = false;
7758 GCPollType compGCPollType;
7762 static bool s_pAltJitExcludeAssembliesListInitialized;
7763 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7768 template <typename T>
7771 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7774 template <typename T>
7777 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7780 static int dspTreeID(GenTree* tree)
7782 return tree->gtTreeID;
7784 static void printTreeID(GenTree* tree)
7786 if (tree == nullptr)
7792 printf("[%06d]", dspTreeID(tree));
7799 #define STRESS_MODES \
7803 /* "Variations" stress areas which we try to mix up with each other. */ \
7804 /* These should not be exhaustively used as they might */ \
7805 /* hide/trivialize other areas */ \
7807 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7808 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7809 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7810 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7811 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7812 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7813 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7814 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7815 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7816 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7817 STRESS_MODE(NULL_OBJECT_CHECK) \
7818 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7819 STRESS_MODE(RANDOM_INLINE) \
7821 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7823 /* "Check" stress areas that can be exhaustively used if we */ \
7824 /* dont care about performance at all */ \
7826 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7827 STRESS_MODE(CHK_FLOW_UPDATE) \
7828 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7830 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7834 #define STRESS_MODE(mode) STRESS_##mode,
7841 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7842 BYTE compActiveStressModes[STRESS_COUNT];
7845 #define MAX_STRESS_WEIGHT 100
7847 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7851 bool compInlineStress()
7853 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7856 bool compRandomInlineStress()
7858 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7863 bool compTailCallStress()
7866 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7872 codeOptimize compCodeOpt()
7875 // Switching between size & speed has measurable throughput impact
7876 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7877 // DEBUG, but should generate identical code between CHK & RET builds,
7878 // so that's not acceptable.
7879 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7880 // Investigate the cause of the throughput regression.
7882 return opts.compCodeOpt;
7884 return BLENDED_CODE;
7888 //--------------------- Info about the procedure --------------------------
7892 COMP_HANDLE compCompHnd;
7893 CORINFO_MODULE_HANDLE compScopeHnd;
7894 CORINFO_CLASS_HANDLE compClassHnd;
7895 CORINFO_METHOD_HANDLE compMethodHnd;
7896 CORINFO_METHOD_INFO* compMethodInfo;
7898 BOOL hasCircularClassConstraints;
7899 BOOL hasCircularMethodConstraints;
7901 #if defined(DEBUG) || defined(LATE_DISASM)
7902 const char* compMethodName;
7903 const char* compClassName;
7904 const char* compFullName;
7905 #endif // defined(DEBUG) || defined(LATE_DISASM)
7907 #if defined(DEBUG) || defined(INLINE_DATA)
7908 // Method hash is logcally const, but computed
7910 mutable unsigned compMethodHashPrivate;
7911 unsigned compMethodHash() const;
7912 #endif // defined(DEBUG) || defined(INLINE_DATA)
7914 #ifdef PSEUDORANDOM_NOP_INSERTION
7915 // things for pseudorandom nop insertion
7916 unsigned compChecksum;
7920 // The following holds the FLG_xxxx flags for the method we're compiling.
7923 // The following holds the class attributes for the method we're compiling.
7924 unsigned compClassAttr;
7926 const BYTE* compCode;
7927 IL_OFFSET compILCodeSize; // The IL code size
7928 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7929 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7930 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7931 // (2) the code is hot/cold split, and we issued less code than we expected
7932 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7934 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7935 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7936 bool compIsContextful : 1; // contextful method
7937 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7938 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7939 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7940 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7941 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7943 var_types compRetType; // Return type of the method as declared in IL
7944 var_types compRetNativeType; // Normalized return type as per target arch ABI
7945 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7946 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7947 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7948 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7949 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7950 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7951 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7952 unsigned compMaxStack;
7953 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7954 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7956 unsigned compCallUnmanaged; // count of unmanaged calls
7957 unsigned compLvFrameListRoot; // lclNum for the Frame root
7958 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7959 // You should generally use compHndBBtabCount instead: it is the
7960 // current number of EH clauses (after additions like synchronized
7961 // methods and funclets, and removals like unreachable code deletion).
7963 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7964 // and the VM expects that, or the JIT is a "self-host" compiler
7965 // (e.g., x86 hosted targeting x86) and the VM expects that.
7967 /* The following holds IL scope information about local variables.
7970 unsigned compVarScopesCount;
7971 VarScopeDsc* compVarScopes;
7973 /* The following holds information about instr offsets for
7974 * which we need to report IP-mappings
7977 IL_OFFSET* compStmtOffsets; // sorted
7978 unsigned compStmtOffsetsCount;
7979 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
7981 #define CPU_X86 0x0100 // The generic X86 CPU
7982 #define CPU_X86_PENTIUM_4 0x0110
7984 #define CPU_X64 0x0200 // The generic x64 CPU
7985 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
7986 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
7988 #define CPU_ARM 0x0300 // The generic ARM CPU
7990 unsigned genCPU; // What CPU are we running on
7993 // Returns true if the method being compiled returns a non-void and non-struct value.
7994 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
7995 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
7996 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
7997 // Methods returning such structs are considered to return non-struct return value and
7998 // this method returns true in that case.
7999 bool compMethodReturnsNativeScalarType()
8001 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8004 // Returns true if the method being compiled returns RetBuf addr as its return value
8005 bool compMethodReturnsRetBufAddr()
8007 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8008 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8010 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8011 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8012 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8013 // methods with hidden RetBufArg.
8015 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8016 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8017 // returning the address of RetBuf.
8019 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8020 // to be returned in RAX.
8021 CLANG_FORMAT_COMMENT_ANCHOR;
8023 #ifdef _TARGET_AMD64_
8024 return (info.compRetBuffArg != BAD_VAR_NUM);
8025 #else // !_TARGET_AMD64_
8026 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8027 #endif // !_TARGET_AMD64_
8030 // Returns true if the method returns a value in more than one return register
8031 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8032 // TODO-ARM64: Does this apply for ARM64 too?
8033 bool compMethodReturnsMultiRegRetType()
8035 #if FEATURE_MULTIREG_RET
8036 #if defined(_TARGET_X86_)
8037 // On x86 only 64-bit longs are returned in multiple registers
8038 return varTypeIsLong(info.compRetNativeType);
8039 #else // targets: X64-UNIX, ARM64 or ARM32
8040 // On all other targets that support multireg return values:
8041 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8042 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8043 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8044 #endif // TARGET_XXX
8046 #else // not FEATURE_MULTIREG_RET
8048 // For this architecture there are no multireg returns
8051 #endif // FEATURE_MULTIREG_RET
8054 #if FEATURE_MULTIREG_ARGS
8055 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8056 // return the gcPtr layout for the pointers sized fields
8057 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8058 #endif // FEATURE_MULTIREG_ARGS
8060 // Returns true if the method being compiled returns a value
8061 bool compMethodHasRetVal()
8063 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8064 compMethodReturnsMultiRegRetType();
8069 void compDispLocalVars();
8073 //-------------------------- Global Compiler Data ------------------------------------
8076 static unsigned s_compMethodsCount; // to produce unique label names
8077 unsigned compGenTreeID;
8080 BasicBlock* compCurBB; // the current basic block in process
8081 GenTreePtr compCurStmt; // the current statement in process
8083 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8086 // The following is used to create the 'method JIT info' block.
8087 size_t compInfoBlkSize;
8088 BYTE* compInfoBlkAddr;
8090 EHblkDsc* compHndBBtab; // array of EH data
8091 unsigned compHndBBtabCount; // element count of used elements in EH data array
8092 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8094 #if defined(_TARGET_X86_)
8096 //-------------------------------------------------------------------------
8097 // Tracking of region covered by the monitor in synchronized methods
8098 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8099 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8101 #endif // !_TARGET_X86_
8103 Phases previousCompletedPhase; // the most recently completed phase
8105 //-------------------------------------------------------------------------
8106 // The following keeps track of how many bytes of local frame space we've
8107 // grabbed so far in the current function, and how many argument bytes we
8108 // need to pop when we return.
8111 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8113 // Count of callee-saved regs we pushed in the prolog.
8114 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8115 // In case of Amd64 this doesn't include float regs saved on stack.
8116 unsigned compCalleeRegsPushed;
8118 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8119 // Mask of callee saved float regs on stack.
8120 regMaskTP compCalleeFPRegsSavedMask;
8122 #ifdef _TARGET_AMD64_
8123 // Quirk for VS debug-launch scenario to work:
8124 // Bytes of padding between save-reg area and locals.
8125 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8126 unsigned compVSQuirkStackPaddingNeeded;
8127 bool compQuirkForPPPflag;
8130 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8132 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8133 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8134 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8136 //-------------------------------------------------------------------------
8138 static void compStartup(); // One-time initialization
8139 static void compShutdown(); // One-time finalization
8141 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8144 static void compDisplayStaticSizes(FILE* fout);
8146 //------------ Some utility functions --------------
8148 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8149 void** ppIndirection); /* OUT */
8151 // Several JIT/EE interface functions return a CorInfoType, and also return a
8152 // class handle as an out parameter if the type is a value class. Returns the
8153 // size of the type these describe.
8154 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8157 // Components used by the compiler may write unit test suites, and
8158 // have them run within this method. They will be run only once per process, and only
8159 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8160 // These should fail by asserting.
8161 void compDoComponentUnitTestsOnce();
8164 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8165 CORINFO_MODULE_HANDLE classPtr,
8166 COMP_HANDLE compHnd,
8167 CORINFO_METHOD_INFO* methodInfo,
8168 void** methodCodePtr,
8169 ULONG* methodCodeSize,
8170 JitFlags* compileFlags);
8171 void compCompileFinish();
8172 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8173 COMP_HANDLE compHnd,
8174 CORINFO_METHOD_INFO* methodInfo,
8175 void** methodCodePtr,
8176 ULONG* methodCodeSize,
8177 JitFlags* compileFlags,
8178 CorInfoInstantiationVerification instVerInfo);
8180 ArenaAllocator* compGetAllocator();
8182 #if MEASURE_MEM_ALLOC
8184 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8188 unsigned allocCnt; // # of allocs
8189 UINT64 allocSz; // total size of those alloc.
8190 UINT64 allocSzMax; // Maximum single allocation.
8191 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8192 UINT64 nraTotalSizeAlloc;
8193 UINT64 nraTotalSizeUsed;
8195 static const char* s_CompMemKindNames[]; // Names of the kinds.
8197 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8199 for (int i = 0; i < CMK_Count; i++)
8201 allocSzByKind[i] = 0;
8204 MemStats(const MemStats& ms)
8205 : allocCnt(ms.allocCnt)
8206 , allocSz(ms.allocSz)
8207 , allocSzMax(ms.allocSzMax)
8208 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8209 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8211 for (int i = 0; i < CMK_Count; i++)
8213 allocSzByKind[i] = ms.allocSzByKind[i];
8217 // Until we have ubiquitous constructors.
8220 this->MemStats::MemStats();
8223 void AddAlloc(size_t sz, CompMemKind cmk)
8227 if (sz > allocSzMax)
8231 allocSzByKind[cmk] += sz;
8234 void Print(FILE* f); // Print these stats to f.
8235 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8237 MemStats genMemStats;
8239 struct AggregateMemStats : public MemStats
8243 AggregateMemStats() : MemStats(), nMethods(0)
8247 void Add(const MemStats& ms)
8250 allocCnt += ms.allocCnt;
8251 allocSz += ms.allocSz;
8252 allocSzMax = max(allocSzMax, ms.allocSzMax);
8253 for (int i = 0; i < CMK_Count; i++)
8255 allocSzByKind[i] += ms.allocSzByKind[i];
8257 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8258 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8261 void Print(FILE* f); // Print these stats to jitstdout.
8264 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8265 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8266 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8268 #endif // MEASURE_MEM_ALLOC
8270 #if LOOP_HOIST_STATS
8271 unsigned m_loopsConsidered;
8272 bool m_curLoopHasHoistedExpression;
8273 unsigned m_loopsWithHoistedExpressions;
8274 unsigned m_totalHoistedExpressions;
8276 void AddLoopHoistStats();
8277 void PrintPerMethodLoopHoistStats();
8279 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8280 static unsigned s_loopsConsidered;
8281 static unsigned s_loopsWithHoistedExpressions;
8282 static unsigned s_totalHoistedExpressions;
8284 static void PrintAggregateLoopHoistStats(FILE* f);
8285 #endif // LOOP_HOIST_STATS
8287 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8288 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8289 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8290 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8291 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8292 void compFreeMem(void*);
8294 bool compIsForImportOnly();
8295 bool compIsForInlining();
8296 bool compDonotInline();
8299 const char* compLocalVarName(unsigned varNum, unsigned offs);
8300 VarName compVarName(regNumber reg, bool isFloatReg = false);
8301 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8302 const char* compRegPairName(regPairNo regPair);
8303 const char* compRegNameForSize(regNumber reg, size_t size);
8304 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8305 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8306 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8309 //-------------------------------------------------------------------------
8311 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8313 struct VarScopeMapInfo
8315 VarScopeListNode* head;
8316 VarScopeListNode* tail;
8317 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8319 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8326 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8327 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8329 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8330 VarNumToScopeDscMap;
8332 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8333 VarNumToScopeDscMap* compVarScopeMap;
8335 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8337 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8339 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8341 void compInitVarScopeMap();
8343 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8344 // enter scope, sorted by instr offset
8345 unsigned compNextEnterScope;
8347 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8348 // go out of scope, sorted by instr offset
8349 unsigned compNextExitScope;
8351 void compInitScopeLists();
8353 void compResetScopeLists();
8355 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8357 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8359 void compProcessScopesUntil(unsigned offset,
8361 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8362 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8365 void compDispScopeLists();
8368 bool compIsProfilerHookNeeded();
8370 //-------------------------------------------------------------------------
8371 /* Statistical Data Gathering */
8373 void compJitStats(); // call this function and enable
8374 // various ifdef's below for statistical data
8377 void compCallArgStats();
8378 static void compDispCallArgStats(FILE* fout);
8381 //-------------------------------------------------------------------------
8388 ArenaAllocator* compAllocator;
8391 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8392 // suitable for use by utilcode collection types.
8393 IAllocator* compAsIAllocator;
8395 #if MEASURE_MEM_ALLOC
8396 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8397 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8398 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8400 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8402 #endif // MEASURE_MEM_ALLOC
8404 void compFunctionTraceStart();
8405 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8408 size_t compMaxUncheckedOffsetForNullObject;
8410 void compInitOptions(JitFlags* compileFlags);
8412 void compSetProcessor();
8413 void compInitDebuggingInfo();
8414 void compSetOptimizationLevel();
8415 #ifdef _TARGET_ARMARCH_
8416 bool compRsvdRegCheck(FrameLayoutState curState);
8418 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8420 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8421 void ResetOptAnnotations();
8423 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8424 void RecomputeLoopInfo();
8426 #ifdef PROFILING_SUPPORTED
8427 // Data required for generating profiler Enter/Leave/TailCall hooks
8429 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8430 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8431 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8434 #ifdef _TARGET_AMD64_
8435 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8438 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8439 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8441 IAllocator* getAllocator()
8443 return compAsIAllocator;
8446 #if MEASURE_MEM_ALLOC
8447 IAllocator* getAllocatorBitset()
8449 return compAsIAllocatorBitset;
8451 IAllocator* getAllocatorGC()
8453 return compAsIAllocatorGC;
8455 IAllocator* getAllocatorLoopHoist()
8457 return compAsIAllocatorLoopHoist;
8459 #else // !MEASURE_MEM_ALLOC
8460 IAllocator* getAllocatorBitset()
8462 return compAsIAllocator;
8464 IAllocator* getAllocatorGC()
8466 return compAsIAllocator;
8468 IAllocator* getAllocatorLoopHoist()
8470 return compAsIAllocator;
8472 #endif // !MEASURE_MEM_ALLOC
8475 IAllocator* getAllocatorDebugOnly()
8477 #if MEASURE_MEM_ALLOC
8478 return compAsIAllocatorDebugOnly;
8479 #else // !MEASURE_MEM_ALLOC
8480 return compAsIAllocator;
8481 #endif // !MEASURE_MEM_ALLOC
8486 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8487 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8491 XX Checks for type compatibility and merges types XX
8493 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8494 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8498 // Set to TRUE if verification cannot be skipped for this method
8499 // If we detect unverifiable code, we will lazily check
8500 // canSkipMethodVerification() to see if verification is REALLY needed.
8501 BOOL tiVerificationNeeded;
8503 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8504 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8505 BOOL tiIsVerifiableCode;
8507 // Set to TRUE if runtime callout is needed for this method
8508 BOOL tiRuntimeCalloutNeeded;
8510 // Set to TRUE if security prolog/epilog callout is needed for this method
8511 // Note: This flag is different than compNeedSecurityCheck.
8512 // compNeedSecurityCheck means whether or not a security object needs
8513 // to be allocated on the stack, which is currently true for EnC as well.
8514 // tiSecurityCalloutNeeded means whether or not security callouts need
8515 // to be inserted in the jitted code.
8516 BOOL tiSecurityCalloutNeeded;
8518 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8519 // This support is necessary to suport attributes that are not described in
8520 // for example, signatures. For example, the permanent home byref (byref that
8521 // points to the gc heap), isn't a property of method signatures, therefore,
8522 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8523 // but when deciding if we need to reimport a block, we need to take these
8525 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8527 // Returns TRUE if child is equal to or a subtype of parent.
8528 // normalisedForStack indicates that both types are normalised for the stack
8529 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8531 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8532 // *pDest is modified to represent the merged type. Sets "*changed" to true
8533 // if this changes "*pDest".
8534 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8536 // Set pDest from the primitive value type.
8537 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8539 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8542 // <BUGNUM> VSW 471305
8543 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8544 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8545 // We use a "short" as we need to push/pop this scope.
8547 short compRegSetCheckLevel;
8551 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8554 XX IL verification stuff XX
8557 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8558 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8562 // The following is used to track liveness of local variables, initialization
8563 // of valueclass constructors, and type safe use of IL instructions.
8565 // dynamic state info needed for verification
8566 EntryState verCurrentState;
8568 // this ptr of object type .ctors are considered intited only after
8569 // the base class ctor is called, or an alternate ctor is called.
8570 // An uninited this ptr can be used to access fields, but cannot
8571 // be used to call a member function.
8572 BOOL verTrackObjCtorInitState;
8574 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8576 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8577 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8578 void verInitCurrentState();
8579 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8581 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8582 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8583 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8585 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8586 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8587 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8588 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8589 typeInfo verMakeTypeInfo(CorInfoType ciType,
8590 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8591 BOOL verIsSDArray(typeInfo ti);
8592 typeInfo verGetArrayElemType(typeInfo ti);
8594 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8595 BOOL verNeedsVerification();
8596 BOOL verIsByRefLike(const typeInfo& ti);
8597 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8599 // generic type variables range over types that satisfy IsBoxable
8600 BOOL verIsBoxable(const typeInfo& ti);
8602 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8603 DEBUGARG(unsigned line));
8604 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8605 DEBUGARG(unsigned line));
8606 bool verCheckTailCallConstraint(OPCODE opcode,
8607 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8608 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8609 // on a type parameter?
8610 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8611 // return false to the caller.
8612 // If false, it will throw.
8614 bool verIsBoxedValueType(typeInfo ti);
8616 void verVerifyCall(OPCODE opcode,
8617 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8618 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8620 bool readonlyCall, // is this a "readonly." call?
8621 const BYTE* delegateCreateStart,
8622 const BYTE* codeAddr,
8623 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8625 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8627 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8628 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8629 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8630 const CORINFO_FIELD_INFO& fieldInfo,
8631 const typeInfo* tiThis,
8633 BOOL allowPlainStructAsThis = FALSE);
8634 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8635 void verVerifyThisPtrInitialised();
8636 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8638 // Register allocator
8639 void raInitStackFP();
8640 void raEnregisterVarsPrePassStackFP();
8641 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8642 void raEnregisterVarsPostPassStackFP();
8643 void raGenerateFPRefCounts();
8644 void raEnregisterVarsStackFP();
8645 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8647 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8648 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8650 // returns true if enregistering v1 would save more mem accesses than v2
8651 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8654 void raDumpHeightsStackFP();
8655 void raDumpVariableRegIntfFloat();
8658 #if FEATURE_STACK_FP_X87
8660 // Currently, we use FP transition blocks in only 2 situations:
8662 // -conditional jump on longs where FP stack differs with target: it's not strictly
8663 // necessary, but its low frequency and the code would get complicated if we try to
8664 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8665 // minimize the way we generate the jump code.
8666 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8667 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8669 // However, transition blocks have 2 problems
8671 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8672 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8673 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8674 // in the right place without preordering them), this causes us to have to generate the transition
8675 // blocks in the cold area if we want procedure splitting.
8678 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8679 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8680 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8681 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8682 // a big change in the exception.
8684 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8685 // optimizations. For these 2 cases:
8687 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8688 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8689 // a switch statement.
8691 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8692 // current procedure splitting and exception code have.
8693 bool compMayHaveTransitionBlocks;
8695 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8697 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8699 unsigned raCntStkStackFP;
8700 unsigned raCntWtdStkDblStackFP;
8701 unsigned raCntStkParamDblStackFP;
8703 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8704 // TODO: Do we want to put this in LclVarDsc?
8705 unsigned raPayloadStackFP[lclMAX_TRACKED];
8706 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8708 // Useful for debugging
8709 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8711 #endif // FEATURE_STACK_FP_X87
8714 // One line log function. Default level is 0. Increasing it gives you
8715 // more log information
8717 // levels are currently unused: #define JITDUMP(level,...) ();
8718 void JitLogEE(unsigned level, const char* fmt, ...);
8720 bool compDebugBreak;
8722 bool compJitHaltMethod();
8727 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8728 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8730 XX GS Security checks for unsafe buffers XX
8732 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8736 struct ShadowParamVarInfo
8738 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8739 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8741 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8743 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8744 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8745 // slots and update all trees to refer to shadow slots is done immediately after
8746 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8747 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8748 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8749 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8750 // creating a shadow slot even though this routine returns true.
8752 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8753 // required. There are two cases under which a reg arg could potentially be used from its
8755 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8756 // b) LSRA spills it
8758 // Possible solution to address case (a)
8759 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8760 // in this routine. Note that live out of exception handler is something we may not be
8761 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8762 // Therefore, for methods with exception handling and need GS cookie check we might have
8763 // to take conservative approach.
8765 // Possible solution to address case (b)
8766 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8767 // create a new spill temp if the method needs GS cookie check.
8768 return varDsc->lvIsParam;
8769 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8770 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8777 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8782 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8783 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8784 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8786 void gsGSChecksInitCookie(); // Grabs cookie variable
8787 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8788 bool gsFindVulnerableParams(); // Shadow param analysis code
8789 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8791 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8792 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8794 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8795 // This can be overwritten by setting complus_JITInlineSize env variable.
8797 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8800 #ifdef FEATURE_JIT_METHOD_PERF
8801 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8802 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8804 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8805 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8807 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8809 #if MEASURE_CLRAPI_CALLS
8810 // Thin wrappers that call into JitTimer (if present).
8811 inline void CLRApiCallEnter(unsigned apix);
8812 inline void CLRApiCallLeave(unsigned apix);
8815 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8816 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8821 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8822 // These variables are associated with maintaining SQM data about compile time.
8823 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8824 // in the current compilation.
8825 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8826 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8827 // the inlining phase in the current compilation.
8828 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8830 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8831 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8832 // type-loading and class initialization).
8833 void RecordStateAtEndOfInlining();
8834 // Assumes being called at the end of compilation. Update the SQM state.
8835 void RecordStateAtEndOfCompilation();
8837 #ifdef FEATURE_CLRSQM
8838 // Does anything SQM related necessary at process shutdown time.
8839 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8840 #endif // FEATURE_CLRSQM
8843 #if FUNC_INFO_LOGGING
8844 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8845 // filename to write it to.
8846 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8847 #endif // FUNC_INFO_LOGGING
8849 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8851 // Is the compilation in a full trust context?
8852 bool compIsFullTrust();
8854 #ifndef FEATURE_TRACELOGGING
8855 // Should we actually fire the noway assert body and the exception handler?
8856 bool compShouldThrowOnNoway();
8857 #else // FEATURE_TRACELOGGING
8858 // Should we actually fire the noway assert body and the exception handler?
8859 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8861 // Telemetry instance to use per method compilation.
8862 JitTelemetry compJitTelemetry;
8864 // Get common parameters that have to be logged with most telemetry data.
8865 void compGetTelemetryDefaults(const char** assemblyName,
8866 const char** scopeName,
8867 const char** methodName,
8868 unsigned* methodHash);
8869 #endif // !FEATURE_TRACELOGGING
8873 NodeToTestDataMap* m_nodeTestData;
8875 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8876 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8877 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8878 // Current kept in this.
8880 NodeToTestDataMap* GetNodeTestData()
8882 Compiler* compRoot = impInlineRoot();
8883 if (compRoot->m_nodeTestData == nullptr)
8885 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8887 return compRoot->m_nodeTestData;
8890 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8892 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8893 // currently occur in the AST graph.
8894 NodeToIntMap* FindReachableNodesInNodeTestData();
8896 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8897 // test data, associate that data with "to".
8898 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8900 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8901 // have annotations, attach similar annotations to the corresponding nodes in "to".
8902 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8904 // These are the methods that test that the various conditions implied by the
8905 // test attributes are satisfied.
8906 void JitTestCheckSSA(); // SSA builder tests.
8907 void JitTestCheckVN(); // Value numbering tests.
8910 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8912 FieldSeqStore* m_fieldSeqStore;
8914 FieldSeqStore* GetFieldSeqStore()
8916 Compiler* compRoot = impInlineRoot();
8917 if (compRoot->m_fieldSeqStore == nullptr)
8919 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8920 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8921 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8923 return compRoot->m_fieldSeqStore;
8926 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8928 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8929 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8930 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8931 // attach the field sequence directly to the address node.
8932 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8934 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8936 // Don't need to worry about inlining here
8937 if (m_zeroOffsetFieldMap == nullptr)
8939 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8941 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8942 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8944 return m_zeroOffsetFieldMap;
8947 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8948 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8949 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8950 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8951 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8952 // record the the field sequence using the ZeroOffsetFieldMap described above.
8954 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8955 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8956 // CoreRT. Such case is handled same as the default case.
8957 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8959 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8961 NodeToArrayInfoMap* m_arrayInfoMap;
8963 NodeToArrayInfoMap* GetArrayInfoMap()
8965 Compiler* compRoot = impInlineRoot();
8966 if (compRoot->m_arrayInfoMap == nullptr)
8968 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8969 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8970 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
8972 return compRoot->m_arrayInfoMap;
8975 NodeToUnsignedMap* m_heapSsaMap;
8977 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
8978 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
8979 // all the possible heap states are possible initial states of the corresponding catch block(s).)
8980 NodeToUnsignedMap* GetHeapSsaMap()
8982 Compiler* compRoot = impInlineRoot();
8983 if (compRoot->m_heapSsaMap == nullptr)
8985 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8986 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8987 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
8989 return compRoot->m_heapSsaMap;
8992 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
8993 CORINFO_CLASS_HANDLE m_refAnyClass;
8994 CORINFO_FIELD_HANDLE GetRefanyDataField()
8996 if (m_refAnyClass == nullptr)
8998 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9000 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9002 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9004 if (m_refAnyClass == nullptr)
9006 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9008 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9012 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9014 #if ALLVARSET_COUNTOPS
9015 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9018 static HelperCallProperties s_helperCallProperties;
9020 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9021 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9022 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9024 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9027 unsigned __int8* offset0,
9028 unsigned __int8* offset1);
9029 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9030 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9032 void fgMorphMultiregStructArgs(GenTreeCall* call);
9033 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9035 }; // end of class Compiler
9037 // Inline methods of CompAllocator.
9038 void* CompAllocator::Alloc(size_t sz)
9040 #if MEASURE_MEM_ALLOC
9041 return m_comp->compGetMem(sz, m_cmk);
9043 return m_comp->compGetMem(sz);
9047 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9049 #if MEASURE_MEM_ALLOC
9050 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9052 return m_comp->compGetMemArray(elems, elemSize);
9056 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9057 inline LclVarDsc::LclVarDsc(Compiler* comp)
9058 : // Initialize the ArgRegs to REG_STK.
9059 // The morph will do the right thing to change
9060 // to the right register if passed in register.
9063 #if FEATURE_MULTIREG_ARGS
9064 _lvOtherArgReg(REG_STK)
9066 #endif // FEATURE_MULTIREG_ARGS
9068 lvRefBlks(BlockSetOps::UninitVal())
9070 #endif // ASSERTION_PROP
9071 lvPerSsaData(comp->getAllocator())
9076 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9077 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9079 XX Miscellaneous Compiler stuff XX
9081 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9082 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9085 // Values used to mark the types a stack slot is used for
9087 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9088 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9089 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9090 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9091 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9092 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9093 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9094 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9096 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9098 /*****************************************************************************
9100 * Variables to keep track of total code amounts.
9105 extern size_t grossVMsize;
9106 extern size_t grossNCsize;
9107 extern size_t totalNCsize;
9109 extern unsigned genMethodICnt;
9110 extern unsigned genMethodNCnt;
9111 extern size_t gcHeaderISize;
9112 extern size_t gcPtrMapISize;
9113 extern size_t gcHeaderNSize;
9114 extern size_t gcPtrMapNSize;
9116 #endif // DISPLAY_SIZES
9118 /*****************************************************************************
9120 * Variables to keep track of basic block counts (more data on 1 BB methods)
9123 #if COUNT_BASIC_BLOCKS
9124 extern Histogram bbCntTable;
9125 extern Histogram bbOneBBSizeTable;
9128 /*****************************************************************************
9130 * Used by optFindNaturalLoops to gather statistical information such as
9131 * - total number of natural loops
9132 * - number of loops with 1, 2, ... exit conditions
9133 * - number of loops that have an iterator (for like)
9134 * - number of loops that have a constant iterator
9139 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9140 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9141 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9142 extern unsigned totalLoopCount; // counts the total number of natural loops
9143 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9144 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9145 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9146 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9148 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9149 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9150 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9151 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9152 extern Histogram loopCountTable; // Histogram of loop counts
9153 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9155 #endif // COUNT_LOOPS
9157 /*****************************************************************************
9158 * variables to keep track of how many iterations we go in a dataflow pass
9163 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9164 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9166 #endif // DATAFLOW_ITER
9168 #if MEASURE_BLOCK_SIZE
9169 extern size_t genFlowNodeSize;
9170 extern size_t genFlowNodeCnt;
9171 #endif // MEASURE_BLOCK_SIZE
9173 #if MEASURE_NODE_SIZE
9174 struct NodeSizeStats
9179 genTreeNodeSize = 0;
9180 genTreeNodeActualSize = 0;
9183 size_t genTreeNodeCnt;
9184 size_t genTreeNodeSize; // The size we allocate
9185 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9186 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9187 // a smaller node to a larger one. TODO-Cleanup: add stats on
9188 // SetOper()/ChangeOper() usage to quanitfy this.
9190 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9191 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9192 extern Histogram genTreeNcntHist;
9193 extern Histogram genTreeNsizHist;
9194 #endif // MEASURE_NODE_SIZE
9196 /*****************************************************************************
9197 * Count fatal errors (including noway_asserts).
9201 extern unsigned fatal_badCode;
9202 extern unsigned fatal_noWay;
9203 extern unsigned fatal_NOMEM;
9204 extern unsigned fatal_noWayAssertBody;
9206 extern unsigned fatal_noWayAssertBodyArgs;
9208 extern unsigned fatal_NYI;
9209 #endif // MEASURE_FATAL
9211 /*****************************************************************************
9215 #ifdef _TARGET_XARCH_
9217 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9218 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9219 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9221 const instruction INS_AND = INS_and;
9222 const instruction INS_OR = INS_or;
9223 const instruction INS_XOR = INS_xor;
9224 const instruction INS_NEG = INS_neg;
9225 const instruction INS_TEST = INS_test;
9226 const instruction INS_MUL = INS_imul;
9227 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9228 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9229 const instruction INS_BREAKPOINT = INS_int3;
9230 const instruction INS_ADDC = INS_adc;
9231 const instruction INS_SUBC = INS_sbb;
9232 const instruction INS_NOT = INS_not;
9238 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9239 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9240 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9242 const instruction INS_AND = INS_and;
9243 const instruction INS_OR = INS_orr;
9244 const instruction INS_XOR = INS_eor;
9245 const instruction INS_NEG = INS_rsb;
9246 const instruction INS_TEST = INS_tst;
9247 const instruction INS_MUL = INS_mul;
9248 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9249 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9250 const instruction INS_BREAKPOINT = INS_bkpt;
9251 const instruction INS_ADDC = INS_adc;
9252 const instruction INS_SUBC = INS_sbc;
9253 const instruction INS_NOT = INS_mvn;
9257 #ifdef _TARGET_ARM64_
9259 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9260 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9261 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9263 const instruction INS_AND = INS_and;
9264 const instruction INS_OR = INS_orr;
9265 const instruction INS_XOR = INS_eor;
9266 const instruction INS_NEG = INS_neg;
9267 const instruction INS_TEST = INS_tst;
9268 const instruction INS_MUL = INS_mul;
9269 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9270 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9271 const instruction INS_BREAKPOINT = INS_bkpt;
9272 const instruction INS_ADDC = INS_adc;
9273 const instruction INS_SUBC = INS_sbc;
9274 const instruction INS_NOT = INS_mvn;
9278 /*****************************************************************************/
9280 extern const BYTE genTypeSizes[];
9281 extern const BYTE genTypeAlignments[];
9282 extern const BYTE genTypeStSzs[];
9283 extern const BYTE genActualTypes[];
9285 /*****************************************************************************/
9287 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9288 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9291 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9292 #elif defined(_TARGET_ARM64_)
9293 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9296 /*****************************************************************************/
9298 #define REG_CORRUPT regNumber(REG_NA + 1)
9299 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9300 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9302 /*****************************************************************************/
9304 extern BasicBlock dummyBB;
9306 /*****************************************************************************/
9307 /*****************************************************************************/
9309 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9310 // nodes of a statement in execution order.
9311 // __stmt: a GT_STMT type GenTree*
9312 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9314 #define foreach_treenode_execution_order(__node, __stmt) \
9315 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9317 // foreach_block: An iterator over all blocks in the function.
9318 // __compiler: the Compiler* object
9319 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9321 #define foreach_block(__compiler, __block) \
9322 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9324 /*****************************************************************************/
9325 /*****************************************************************************/
9329 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9331 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9332 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9334 XX Debugging helpers XX
9336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9337 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9340 /*****************************************************************************/
9341 /* The following functions are intended to be called from the debugger, to dump
9342 * various data structures. The can be used in the debugger Watch or Quick Watch
9343 * windows. They are designed to be short to type and take as few arguments as
9344 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9345 * See the function definition comment for more details.
9348 void cBlock(Compiler* comp, BasicBlock* block);
9349 void cBlocks(Compiler* comp);
9350 void cBlocksV(Compiler* comp);
9351 void cTree(Compiler* comp, GenTree* tree);
9352 void cTrees(Compiler* comp);
9353 void cEH(Compiler* comp);
9354 void cVar(Compiler* comp, unsigned lclNum);
9355 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9356 void cVars(Compiler* comp);
9357 void cVarsFinal(Compiler* comp);
9358 void cBlockPreds(Compiler* comp, BasicBlock* block);
9359 void cReach(Compiler* comp);
9360 void cDoms(Compiler* comp);
9361 void cLiveness(Compiler* comp);
9362 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9364 void cFuncIR(Compiler* comp);
9365 void cBlockIR(Compiler* comp, BasicBlock* block);
9366 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9367 void cTreeIR(Compiler* comp, GenTree* tree);
9368 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9369 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9370 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9371 int cOperandIR(Compiler* comp, GenTree* operand);
9372 int cLeafIR(Compiler* comp, GenTree* tree);
9373 int cIndirIR(Compiler* comp, GenTree* tree);
9374 int cListIR(Compiler* comp, GenTree* list);
9375 int cSsaNumIR(Compiler* comp, GenTree* tree);
9376 int cValNumIR(Compiler* comp, GenTree* tree);
9377 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9379 void dBlock(BasicBlock* block);
9382 void dTree(GenTree* tree);
9385 void dVar(unsigned lclNum);
9386 void dVarDsc(LclVarDsc* varDsc);
9389 void dBlockPreds(BasicBlock* block);
9393 void dCVarSet(VARSET_VALARG_TP vars);
9395 void dVarSet(VARSET_VALARG_TP vars);
9396 void dRegMask(regMaskTP mask);
9399 void dBlockIR(BasicBlock* block);
9400 void dTreeIR(GenTree* tree);
9401 void dLoopIR(Compiler::LoopDsc* loop);
9402 void dLoopNumIR(unsigned loopNum);
9403 int dTabStopIR(int curr, int tabstop);
9404 int dTreeTypeIR(GenTree* tree);
9405 int dTreeKindsIR(GenTree* tree);
9406 int dTreeFlagsIR(GenTree* tree);
9407 int dOperandIR(GenTree* operand);
9408 int dLeafIR(GenTree* tree);
9409 int dIndirIR(GenTree* tree);
9410 int dListIR(GenTree* list);
9411 int dSsaNumIR(GenTree* tree);
9412 int dValNumIR(GenTree* tree);
9413 int dDependsIR(GenTree* comma);
9416 GenTree* dFindTree(GenTree* tree, unsigned id);
9417 GenTree* dFindTree(unsigned id);
9418 GenTreeStmt* dFindStmt(unsigned id);
9419 BasicBlock* dFindBlock(unsigned bbNum);
9423 #include "compiler.hpp" // All the shared inline functions
9425 /*****************************************************************************/
9426 #endif //_COMPILER_H_
9427 /*****************************************************************************/