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 lvAssignOne : 1; // assigned at least once?
273 unsigned char lvAssignTwo : 1; // assigned at least twice?
276 unsigned char lvSpilled : 1; // enregistered variable was spilled
277 #ifndef _TARGET_64BIT_
278 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
279 #endif // !_TARGET_64BIT_
280 #ifdef _TARGET_64BIT_
281 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
284 unsigned char lvKeepType : 1; // Don't change the type of this variable
285 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
287 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
289 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
290 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
292 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
293 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
294 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
295 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
296 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
298 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
299 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
302 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
303 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
304 // with (lvIsRegArg && lvIsHfa())
305 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
306 #endif // FEATURE_HFA
309 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
310 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
311 // Consider cleaning this up so this workaround is not required.
312 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
313 // I.e. there is no longer any reference to the struct directly.
314 // In this case we can simply remove this struct local.
316 #ifndef LEGACY_BACKEND
317 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
318 #endif // !LEGACY_BACKEND
321 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
322 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
323 unsigned char lvSIMDType : 1; // This is a SIMD struct
324 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
325 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
326 #endif // FEATURE_SIMD
327 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
330 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
332 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
333 // Valid on promoted struct local fields.
336 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
337 unsigned char lvFldOffset;
338 unsigned char lvFldOrdinal;
340 #if FEATURE_MULTIREG_ARGS
341 regNumber lvRegNumForSlot(unsigned slotNum)
347 else if (slotNum == 1)
349 return lvOtherArgReg;
353 assert(false && "Invalid slotNum!");
358 #endif // FEATURE_MULTIREG_ARGS
376 bool lvIsHfaRegArg() const
379 return _lvIsHfaRegArg;
385 void lvSetIsHfaRegArg()
388 _lvIsHfaRegArg = true;
392 bool lvHfaTypeIsFloat() const
395 return _lvHfaTypeIsFloat;
401 void lvSetHfaTypeIsFloat(bool value)
404 _lvHfaTypeIsFloat = value;
408 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
409 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
411 unsigned lvHfaSlots() const
414 assert(lvType == TYP_STRUCT);
416 return lvExactSize / sizeof(float);
417 #else // _TARGET_ARM64_
418 if (lvHfaTypeIsFloat())
420 return lvExactSize / sizeof(float);
424 return lvExactSize / sizeof(double);
426 #endif // _TARGET_ARM64_
429 // lvIsMultiRegArgOrRet()
430 // returns true if this is a multireg LclVar struct used in an argument context
431 // or if this is a multireg LclVar struct assigned from a multireg call
432 bool lvIsMultiRegArgOrRet()
434 return lvIsMultiRegArg || lvIsMultiRegRet;
438 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
439 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
440 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
441 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
442 // to non-zero if the variable gets the same register assignment for its entire
444 #if !defined(_TARGET_64BIT_)
445 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
446 #endif // !defined(_TARGET_64BIT_)
448 regNumberSmall _lvArgReg; // The register in which this argument is passed.
450 #if FEATURE_MULTIREG_ARGS
451 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
452 // Note this is defined but not used by ARM32
453 #endif // FEATURE_MULTIREG_ARGS
455 #ifndef LEGACY_BACKEND
457 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
458 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
460 #endif // !LEGACY_BACKEND
463 // The register number is stored in a small format (8 bits), but the getters return and the setters take
464 // a full-size (unsigned) format, to localize the casts here.
466 /////////////////////
468 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
470 regNumber GetRegNum() const
472 return (regNumber)_lvRegNum;
475 void SetRegNum(regNumber reg)
477 _lvRegNum = (regNumberSmall)reg;
478 assert(_lvRegNum == reg);
481 /////////////////////
483 #if defined(_TARGET_64BIT_)
484 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
486 regNumber GetOtherReg() const
488 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
489 // "unreachable code" warnings
493 void SetOtherReg(regNumber reg)
495 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
496 // "unreachable code" warnings
498 #else // !_TARGET_64BIT_
499 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
501 regNumber GetOtherReg() const
503 return (regNumber)_lvOtherReg;
506 void SetOtherReg(regNumber reg)
508 _lvOtherReg = (regNumberSmall)reg;
509 assert(_lvOtherReg == reg);
511 #endif // !_TARGET_64BIT_
513 /////////////////////
515 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
517 regNumber GetArgReg() const
519 return (regNumber)_lvArgReg;
522 void SetArgReg(regNumber reg)
524 _lvArgReg = (regNumberSmall)reg;
525 assert(_lvArgReg == reg);
528 #if FEATURE_MULTIREG_ARGS
529 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
531 regNumber GetOtherArgReg() const
533 return (regNumber)_lvOtherArgReg;
536 void SetOtherArgReg(regNumber reg)
538 _lvOtherArgReg = (regNumberSmall)reg;
539 assert(_lvOtherArgReg == reg);
541 #endif // FEATURE_MULTIREG_ARGS
544 // Is this is a SIMD struct?
545 bool lvIsSIMDType() const
550 // Is this is a SIMD struct which is used for SIMD intrinsic?
551 bool lvIsUsedInSIMDIntrinsic() const
553 return lvUsedInSIMDIntrinsic;
556 // If feature_simd not enabled, return false
557 bool lvIsSIMDType() const
561 bool lvIsUsedInSIMDIntrinsic() const
567 /////////////////////
569 #ifndef LEGACY_BACKEND
570 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
572 regNumber GetArgInitReg() const
574 return (regNumber)_lvArgInitReg;
577 void SetArgInitReg(regNumber reg)
579 _lvArgInitReg = (regNumberSmall)reg;
580 assert(_lvArgInitReg == reg);
583 /////////////////////
585 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
587 regPairNo GetArgInitRegPair() const
589 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
590 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
594 void SetArgInitRegPair(regPairNo regPair)
596 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
597 _lvArgInitRegPair = (regPairNoSmall)regPair;
598 assert(_lvArgInitRegPair == regPair);
601 /////////////////////
603 bool lvIsRegCandidate() const
605 return lvLRACandidate != 0;
608 bool lvIsInReg() const
610 return lvIsRegCandidate() && (lvRegNum != REG_STK);
613 #else // LEGACY_BACKEND
615 bool lvIsRegCandidate() const
617 return lvTracked != 0;
620 bool lvIsInReg() const
622 return lvRegister != 0;
625 #endif // LEGACY_BACKEND
627 regMaskTP lvRegMask() const
629 regMaskTP regMask = RBM_NONE;
630 if (varTypeIsFloating(TypeGet()))
632 if (lvRegNum != REG_STK)
634 regMask = genRegMaskFloat(lvRegNum, TypeGet());
639 if (lvRegNum != REG_STK)
641 regMask = genRegMask(lvRegNum);
644 // For longs we may have two regs
645 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
647 regMask |= genRegMask(lvOtherReg);
653 regMaskSmall lvPrefReg; // set of regs it prefers to live in
655 unsigned short lvVarIndex; // variable tracking index
656 unsigned short lvRefCnt; // unweighted (real) reference count
657 unsigned lvRefCntWtd; // weighted reference count
658 int lvStkOffs; // stack offset of home
659 unsigned lvExactSize; // (exact) size of the type in bytes
661 // Is this a promoted struct?
662 // This method returns true only for structs (including SIMD structs), not for
663 // locals that are split on a 32-bit target.
664 // It is only necessary to use this:
665 // 1) if only structs are wanted, and
666 // 2) if Lowering has already been done.
667 // Otherwise lvPromoted is valid.
668 bool lvPromotedStruct()
670 #if !defined(_TARGET_64BIT_)
671 return (lvPromoted && !varTypeIsLong(lvType));
672 #else // defined(_TARGET_64BIT_)
674 #endif // defined(_TARGET_64BIT_)
677 unsigned lvSize() // Size needed for storage representation. Only used for structs or TYP_BLK.
679 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
680 // where the struct itself is no longer used because all access is via its member fields.
681 // When that happens, the struct is marked as unused and its type has been changed to
682 // TYP_INT (to keep the GC tracking code from looking at it).
683 // See Compiler::raAssignVars() for details. For example:
684 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
685 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
686 // float V03.f1 (offs=0x00) -> V12 tmp7
687 // f8 (last use) (last use) $345
688 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
689 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
691 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
693 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
694 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
695 // this for arguments, which must be passed according the defined ABI.
696 if ((lvType == TYP_SIMD12) && !lvIsParam)
698 assert(lvExactSize == 12);
701 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
703 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
706 unsigned lvSlotNum; // original slot # (if remapped)
708 typeInfo lvVerTypeInfo; // type info needed for verification
710 BYTE* lvGcLayout; // GC layout info for structs
713 GenTreePtr lvKnownDim; // array size if known
717 BlockSet lvRefBlks; // Set of blocks that contain refs
718 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
719 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
721 var_types TypeGet() const
723 return (var_types)lvType;
725 bool lvStackAligned() const
727 assert(lvIsStructField);
728 return ((lvFldOffset % sizeof(void*)) == 0);
730 bool lvNormalizeOnLoad() const
732 return varTypeIsSmall(TypeGet()) &&
733 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
734 (lvIsParam || lvAddrExposed || lvIsStructField);
737 bool lvNormalizeOnStore()
739 return varTypeIsSmall(TypeGet()) &&
740 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
741 !(lvIsParam || lvAddrExposed || lvIsStructField);
744 void lvaResetSortAgainFlag(Compiler* pComp);
745 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
746 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
747 void setPrefReg(regNumber regNum, Compiler* pComp);
748 void addPrefReg(regMaskTP regMask, Compiler* pComp);
749 bool IsFloatRegType() const
751 return isFloatRegType(lvType) || lvIsHfaRegArg();
753 var_types GetHfaType() const
755 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
757 void SetHfaType(var_types type)
759 assert(varTypeIsFloating(type));
760 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
763 #ifndef LEGACY_BACKEND
764 var_types lvaArgType();
767 PerSsaArray lvPerSsaData;
770 // Keep track of the # of SsaNames, for a bounds check.
771 unsigned lvNumSsaNames;
774 // Returns the address of the per-Ssa data for the given ssaNum (which is required
775 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
776 // not an SSA variable).
777 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
779 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
780 assert(SsaConfig::RESERVED_SSA_NUM == 0);
781 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
782 assert(zeroBased < lvNumSsaNames);
783 return &lvPerSsaData.GetRef(zeroBased);
788 void PrintVarReg() const
790 if (isRegPairType(TypeGet()))
792 printf("%s:%s", getRegName(lvOtherReg), // hi32
793 getRegName(lvRegNum)); // lo32
797 printf("%s", getRegName(lvRegNum));
802 }; // class LclVarDsc
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
810 XX The temporary lclVars allocated by the compiler for code generation XX
812 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
813 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
816 /*****************************************************************************
818 * The following keeps track of temporaries allocated in the stack frame
819 * during code-generation (after register allocation). These spill-temps are
820 * only used if we run out of registers while evaluating a tree.
822 * These are different from the more common temps allocated by lvaGrabTemp().
833 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
841 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
845 0); // temps must have a negative number (so they have a different number from all local variables)
846 tdOffs = BAD_TEMP_OFFSET;
850 IMPL_LIMITATION("too many spill temps");
855 bool tdLegalOffset() const
857 return tdOffs != BAD_TEMP_OFFSET;
861 int tdTempOffs() const
863 assert(tdLegalOffset());
866 void tdSetTempOffs(int offs)
869 assert(tdLegalOffset());
871 void tdAdjustTempOffs(int offs)
874 assert(tdLegalOffset());
877 int tdTempNum() const
882 unsigned tdTempSize() const
886 var_types tdTempType() const
892 // interface to hide linearscan implementation from rest of compiler
893 class LinearScanInterface
896 virtual void doLinearScan() = 0;
897 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
900 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
902 // Information about arrays: their element type and size, and the offset of the first element.
903 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
904 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
905 // for example, in value numbering of array index expressions.
908 var_types m_elemType;
909 CORINFO_CLASS_HANDLE m_elemStructType;
911 unsigned m_elemOffset;
913 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
917 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
918 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
923 // This enumeration names the phases into which we divide compilation. The phases should completely
924 // partition a compilation.
927 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
928 #include "compphases.h"
932 extern const char* PhaseNames[];
933 extern const char* PhaseEnums[];
934 extern const LPCWSTR PhaseShortNames[];
936 // The following enum provides a simple 1:1 mapping to CLR API's
937 enum API_ICorJitInfo_Names
939 #define DEF_CLR_API(name) API_##name,
940 #include "ICorJitInfo_API_names.h"
944 //---------------------------------------------------------------
948 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
949 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
950 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
951 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
952 // by "m_timerFailure" being true.
953 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
956 #ifdef FEATURE_JIT_METHOD_PERF
957 // The string names of the phases.
958 static const char* PhaseNames[];
960 static bool PhaseHasChildren[];
961 static int PhaseParent[];
963 unsigned m_byteCodeBytes;
964 unsigned __int64 m_totalCycles;
965 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
966 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
967 #if MEASURE_CLRAPI_CALLS
968 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
969 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
971 // For better documentation, we call EndPhase on
972 // non-leaf phases. We should also call EndPhase on the
973 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
974 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
975 // We add all such "redundant end phase" intervals to this variable below; we print
976 // it out in a report, so we can verify that it is, indeed, very small. If it ever
977 // isn't, this means that we're doing something significant between the end of the last
978 // declared subphase and the end of its parent.
979 unsigned __int64 m_parentPhaseEndSlop;
982 #if MEASURE_CLRAPI_CALLS
983 // The following measures the time spent inside each individual CLR API call.
984 unsigned m_allClrAPIcalls;
985 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
986 unsigned __int64 m_allClrAPIcycles;
987 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
988 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
989 #endif // MEASURE_CLRAPI_CALLS
991 CompTimeInfo(unsigned byteCodeBytes);
995 #ifdef FEATURE_JIT_METHOD_PERF
997 #if MEASURE_CLRAPI_CALLS
998 struct WrapICorJitInfo;
1001 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1002 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1003 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1004 // threads, so it is protected by a lock.
1005 // This class is intended to be used as a singleton type, with only a single instance.
1006 class CompTimeSummaryInfo
1008 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1009 static CritSecObject s_compTimeSummaryLock;
1013 CompTimeInfo m_total;
1014 CompTimeInfo m_maximum;
1016 int m_numFilteredMethods;
1017 CompTimeInfo m_filtered;
1019 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1020 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1021 // If any OS interaction fails, returns 0.0.
1022 double CyclesPerSecond();
1024 // This can use what ever data you want to determine if the value to be added
1025 // belongs in the filtered section (it's always included in the unfiltered section)
1026 bool IncludedInFilteredData(CompTimeInfo& info);
1029 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1030 static CompTimeSummaryInfo s_compTimeSummary;
1032 CompTimeSummaryInfo()
1033 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1037 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1038 // This is thread safe.
1039 void AddInfo(CompTimeInfo& info, bool includePhases);
1041 // Print the summary information to "f".
1042 // This is not thread-safe; assumed to be called by only one thread.
1043 void Print(FILE* f);
1046 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1047 // and when the current phase started. This is intended to be part of a Compilation object. This is
1048 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1052 unsigned __int64 m_start; // Start of the compilation.
1053 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1054 #if MEASURE_CLRAPI_CALLS
1055 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1056 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1057 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1058 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1059 static double s_cyclesPerSec; // Cached for speedier measurements
1062 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1064 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1066 static CritSecObject s_csvLock; // Lock to protect the time log file.
1067 void PrintCsvMethodStats(Compiler* comp);
1070 void* operator new(size_t);
1071 void* operator new[](size_t);
1072 void operator delete(void*);
1073 void operator delete[](void*);
1076 // Initialized the timer instance
1077 JitTimer(unsigned byteCodeSize);
1079 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1081 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1084 static void PrintCsvHeader();
1086 // Ends the current phase (argument is for a redundant check).
1087 void EndPhase(Phases phase);
1089 #if MEASURE_CLRAPI_CALLS
1090 // Start and end a timed CLR API call.
1091 void CLRApiCallEnter(unsigned apix);
1092 void CLRApiCallLeave(unsigned apix);
1093 #endif // MEASURE_CLRAPI_CALLS
1095 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1096 // and adds it to "sum".
1097 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1099 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1100 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1101 // "m_info" to true.
1102 bool GetThreadCycles(unsigned __int64* cycles)
1104 bool res = CycleTimer::GetThreadCyclesS(cycles);
1107 m_info.m_timerFailure = true;
1112 #endif // FEATURE_JIT_METHOD_PERF
1114 //------------------- Function/Funclet info -------------------------------
1115 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1117 FUNC_ROOT, // The main/root function (always id==0)
1118 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1119 FUNC_FILTER, // a funclet associated with an EH filter
1122 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1129 BYTE funFlags; // Currently unused, just here for padding
1130 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1131 // funclet. It is only valid if funKind field indicates this is a
1132 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1134 #if defined(_TARGET_AMD64_)
1136 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1137 emitLocation* startLoc;
1138 emitLocation* endLoc;
1139 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1140 emitLocation* coldEndLoc;
1141 UNWIND_INFO unwindHeader;
1142 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1143 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1144 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1145 unsigned unwindCodeSlot;
1147 #ifdef UNIX_AMD64_ABI
1148 jitstd::vector<CFI_CODE>* cfiCodes;
1149 #endif // UNIX_AMD64_ABI
1151 #elif defined(_TARGET_ARMARCH_)
1153 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1154 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1155 // Note: we only have a pointer here instead of the actual object,
1156 // to save memory in the JIT case (compared to the NGEN case),
1157 // where we don't have any cold section.
1158 // Note 2: we currently don't support hot/cold splitting in functions
1159 // with EH, so uwiCold will be NULL for all funclets.
1161 #endif // _TARGET_ARMARCH_
1163 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1164 // that isn't shared between the main function body and funclets.
1167 struct fgArgTabEntry
1170 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1173 otherRegNum = REG_NA;
1174 isStruct = false; // is this a struct arg
1176 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1178 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1180 // it will point at the actual argument in the gtCallLateArgs list.
1181 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1183 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1185 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1187 unsigned numRegs; // Count of number of registers that this argument uses
1189 // A slot is a pointer sized region in the OutArg area.
1190 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1191 unsigned numSlots; // Count of number of slots that this argument uses
1193 unsigned alignment; // 1 or 2 (slots/registers)
1194 unsigned lateArgInx; // index into gtCallLateArgs list
1195 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1197 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1198 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1199 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1200 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1201 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1202 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1203 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1204 // previous arguments.
1205 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1206 // to be on the stack despite its arg list position.
1208 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1209 bool isStruct : 1; // True if this is a struct arg
1211 regNumber otherRegNum; // The (second) register to use when passing this argument.
1213 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1214 #elif defined(_TARGET_X86_)
1215 __declspec(property(get = getIsStruct)) bool isStruct;
1218 return varTypeIsStruct(node);
1220 #endif // _TARGET_X86_
1223 void SetIsHfaRegArg(bool hfaRegArg)
1225 isHfaRegArg = hfaRegArg;
1228 void SetIsBackFilled(bool backFilled)
1230 isBackFilled = backFilled;
1233 bool IsBackFilled() const
1235 return isBackFilled;
1237 #else // !_TARGET_ARM_
1238 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1240 void SetIsHfaRegArg(bool hfaRegArg)
1244 void SetIsBackFilled(bool backFilled)
1248 bool IsBackFilled() const
1252 #endif // !_TARGET_ARM_
1258 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1260 //-------------------------------------------------------------------------
1262 // The class fgArgInfo is used to handle the arguments
1263 // when morphing a GT_CALL node.
1268 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1269 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1270 unsigned argCount; // Updatable arg count value
1271 unsigned nextSlotNum; // Updatable slot count value
1272 unsigned stkLevel; // Stack depth when we make this call (for x86)
1274 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1275 bool hasRegArgs; // true if we have one or more register arguments
1276 bool hasStackArgs; // true if we have one or more stack arguments
1277 bool argsComplete; // marker for state
1278 bool argsSorted; // marker for state
1279 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1282 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1285 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1286 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1288 fgArgTabEntryPtr AddRegArg(
1289 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1291 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1292 fgArgTabEntryPtr AddRegArg(
1299 const bool isStruct,
1300 const regNumber otherRegNum = REG_NA,
1301 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1302 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1304 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1308 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1310 void RemorphReset();
1311 fgArgTabEntryPtr RemorphRegArg(
1312 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1314 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1316 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1318 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1320 void ArgsComplete();
1324 void EvalArgsToTemps();
1326 void RecordStkLevel(unsigned stkLvl);
1327 unsigned RetrieveStkLevel();
1333 fgArgTabEntryPtr* ArgTable()
1337 unsigned GetNextSlotNum()
1347 return hasStackArgs;
1349 bool AreArgsComplete() const
1351 return argsComplete;
1356 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1357 // We have the ability to mark source expressions with "Test Labels."
1358 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1359 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1361 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1364 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1365 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1366 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1367 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1368 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1371 struct TestLabelAndNum
1376 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1381 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1383 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1386 // This class implements the "IAllocator" interface, so that we can use
1387 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1389 class CompAllocator : public IAllocator
1392 #if MEASURE_MEM_ALLOC
1396 CompAllocator(Compiler* comp, CompMemKind cmk)
1398 #if MEASURE_MEM_ALLOC
1404 inline void* Alloc(size_t sz);
1406 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1408 // For the compiler's no-release allocator, free operations are no-ops.
1415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1416 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1418 XX The big guy. The sections are currently organized as : XX
1420 XX o GenTree and BasicBlock XX
1432 XX o PrologScopeInfo XX
1433 XX o CodeGenerator XX
1438 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1439 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1444 friend class emitter;
1445 friend class UnwindInfo;
1446 friend class UnwindFragmentInfo;
1447 friend class UnwindEpilogInfo;
1448 friend class JitTimer;
1449 friend class LinearScan;
1450 friend class fgArgInfo;
1451 friend class Rationalizer;
1453 friend class Lowering;
1454 friend class CSE_DataFlow;
1455 friend class CSE_Heuristic;
1456 friend class CodeGenInterface;
1457 friend class CodeGen;
1458 friend class LclVarDsc;
1459 friend class TempDsc;
1461 friend class ObjectAllocator;
1463 #ifndef _TARGET_64BIT_
1464 friend class DecomposeLongs;
1465 #endif // !_TARGET_64BIT_
1468 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1469 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1471 XX Misc structs definitions XX
1473 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1474 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1478 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1497 bool dumpIRDataflow;
1498 bool dumpIRBlockHeaders;
1500 LPCWSTR dumpIRPhase;
1501 LPCWSTR dumpIRFormat;
1503 bool shouldUseVerboseTrees();
1504 bool asciiTrees; // If true, dump trees using only ASCII characters
1505 bool shouldDumpASCIITrees();
1506 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1507 bool shouldUseVerboseSsa();
1508 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1509 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1511 const char* VarNameToStr(VarName name)
1516 DWORD expensiveDebugCheckLevel;
1519 #if FEATURE_MULTIREG_RET
1520 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1521 #endif // FEATURE_MULTIREG_RET
1524 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1525 #endif // ARM_SOFTFP
1527 //-------------------------------------------------------------------------
1528 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1529 // HFAs are one to four element structs where each element is the same
1530 // type, either all float or all double. They are treated specially
1531 // in the ARM Procedure Call Standard, specifically, they are passed in
1532 // floating-point registers instead of the general purpose registers.
1535 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1536 bool IsHfa(GenTreePtr tree);
1538 var_types GetHfaType(GenTreePtr tree);
1539 unsigned GetHfaCount(GenTreePtr tree);
1541 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1542 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1544 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1545 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1547 //-------------------------------------------------------------------------
1548 // The following is used for validating format of EH table
1552 typedef struct EHNodeDsc* pEHNodeDsc;
1554 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1555 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1568 EHBlockType ehnBlockType; // kind of EH block
1569 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1570 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1571 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1573 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1574 pEHNodeDsc ehnChild; // leftmost nested block
1576 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1577 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1579 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1580 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1582 inline void ehnSetTryNodeType()
1584 ehnBlockType = TryNode;
1586 inline void ehnSetFilterNodeType()
1588 ehnBlockType = FilterNode;
1590 inline void ehnSetHandlerNodeType()
1592 ehnBlockType = HandlerNode;
1594 inline void ehnSetFinallyNodeType()
1596 ehnBlockType = FinallyNode;
1598 inline void ehnSetFaultNodeType()
1600 ehnBlockType = FaultNode;
1603 inline BOOL ehnIsTryBlock()
1605 return ehnBlockType == TryNode;
1607 inline BOOL ehnIsFilterBlock()
1609 return ehnBlockType == FilterNode;
1611 inline BOOL ehnIsHandlerBlock()
1613 return ehnBlockType == HandlerNode;
1615 inline BOOL ehnIsFinallyBlock()
1617 return ehnBlockType == FinallyNode;
1619 inline BOOL ehnIsFaultBlock()
1621 return ehnBlockType == FaultNode;
1624 // returns true if there is any overlap between the two nodes
1625 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1627 if (node1->ehnStartOffset < node2->ehnStartOffset)
1629 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1633 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1637 // fails with BADCODE if inner is not completely nested inside outer
1638 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1640 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1644 //-------------------------------------------------------------------------
1645 // Exception handling functions
1648 #if !FEATURE_EH_FUNCLETS
1650 bool ehNeedsShadowSPslots()
1652 return (info.compXcptnsCount || opts.compDbgEnC);
1655 // 0 for methods with no EH
1656 // 1 for methods with non-nested EH, or where only the try blocks are nested
1657 // 2 for a method with a catch within a catch
1659 unsigned ehMaxHndNestingCount;
1661 #endif // !FEATURE_EH_FUNCLETS
1663 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1664 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1666 bool bbInCatchHandlerILRange(BasicBlock* blk);
1667 bool bbInFilterILRange(BasicBlock* blk);
1668 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1669 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1670 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1671 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1672 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1674 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1675 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1677 // Returns true if "block" is the start of a try region.
1678 bool bbIsTryBeg(BasicBlock* block);
1680 // Returns true if "block" is the start of a handler or filter region.
1681 bool bbIsHandlerBeg(BasicBlock* block);
1683 // Returns true iff "block" is where control flows if an exception is raised in the
1684 // try region, and sets "*regionIndex" to the index of the try for the handler.
1685 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1686 // block of the filter, but not for the filter's handler.
1687 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1689 bool ehHasCallableHandlers();
1691 // Return the EH descriptor for the given region index.
1692 EHblkDsc* ehGetDsc(unsigned regionIndex);
1694 // Return the EH index given a region descriptor.
1695 unsigned ehGetIndex(EHblkDsc* ehDsc);
1697 // Return the EH descriptor index of the enclosing try, for the given region index.
1698 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1700 // Return the EH descriptor index of the enclosing handler, for the given region index.
1701 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1703 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1704 // block is not in a 'try' region).
1705 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1707 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1708 // if this block is not in a filter or handler region).
1709 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1711 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1712 // nullptr if this block's exceptions propagate to caller).
1713 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1715 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1716 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1717 bool ehIsBlockEHLast(BasicBlock* block);
1719 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1721 // Return the region index of the most nested EH region this block is in.
1722 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1724 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1725 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1727 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1728 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1729 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1730 // (It can never be a filter.)
1731 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1733 // A block has been deleted. Update the EH table appropriately.
1734 void ehUpdateForDeletedBlock(BasicBlock* block);
1736 // Determine whether a block can be deleted while preserving the EH normalization rules.
1737 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1739 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1740 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1742 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1743 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1744 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1745 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1746 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1747 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1748 // lives in a filter.)
1749 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1751 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1752 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1753 // (nullptr if the last block is the last block in the program).
1754 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1755 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1758 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1759 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1760 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1763 #if FEATURE_EH_FUNCLETS
1764 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1765 // if there is a filter that protects a region with a nested EH clause (such as a
1766 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1767 // genFuncletProlog() for more details. However, the VM seems to use it for more
1768 // purposes, maybe including debugging. Until we are sure otherwise, always create
1769 // a PSPSym for functions with any EH.
1770 bool ehNeedsPSPSym() const
1774 #else // _TARGET_X86_
1775 return compHndBBtabCount > 0;
1776 #endif // _TARGET_X86_
1779 bool ehAnyFunclets(); // Are there any funclets in this function?
1780 unsigned ehFuncletCount(); // Return the count of funclets in the function
1782 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1783 #else // !FEATURE_EH_FUNCLETS
1784 bool ehAnyFunclets()
1788 unsigned ehFuncletCount()
1793 unsigned bbThrowIndex(BasicBlock* blk)
1795 return blk->bbTryIndex;
1796 } // Get the index to use as the cache key for sharing throw blocks
1797 #endif // !FEATURE_EH_FUNCLETS
1799 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1800 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1801 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1802 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1803 // convenient to also consider it a predecessor.)
1804 flowList* BlockPredsWithEH(BasicBlock* blk);
1806 // This table is useful for memoization of the method above.
1807 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1809 BlockToFlowListMap* m_blockToEHPreds;
1810 BlockToFlowListMap* GetBlockToEHPreds()
1812 if (m_blockToEHPreds == nullptr)
1814 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1816 return m_blockToEHPreds;
1819 void* ehEmitCookie(BasicBlock* block);
1820 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1822 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1824 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1826 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1828 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1830 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1832 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1834 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1836 void fgAllocEHTable();
1838 void fgRemoveEHTableEntry(unsigned XTnum);
1840 #if FEATURE_EH_FUNCLETS
1842 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1844 #endif // FEATURE_EH_FUNCLETS
1848 #endif // !FEATURE_EH
1850 void fgSortEHTable();
1852 // Causes the EH table to obey some well-formedness conditions, by inserting
1853 // empty BB's when necessary:
1854 // * No block is both the first block of a handler and the first block of a try.
1855 // * No block is the first block of multiple 'try' regions.
1856 // * No block is the last block of multiple EH regions.
1857 void fgNormalizeEH();
1858 bool fgNormalizeEHCase1();
1859 bool fgNormalizeEHCase2();
1860 bool fgNormalizeEHCase3();
1863 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1864 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1865 void fgVerifyHandlerTab();
1866 void fgDispHandlerTab();
1869 bool fgNeedToSortEHTable;
1871 void verInitEHTree(unsigned numEHClauses);
1872 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1873 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1874 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1875 void verCheckNestingLevel(EHNodeDsc* initRoot);
1878 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1879 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1881 XX GenTree and BasicBlock XX
1883 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1885 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1886 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1889 // Functions to create nodes
1890 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1893 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1895 // For binary opers.
1896 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1898 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1900 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1901 var_types type = TYP_I_IMPL,
1902 GenTreePtr op1 = nullptr,
1903 GenTreePtr op2 = nullptr);
1905 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1907 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1909 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1911 GenTreePtr gtNewJmpTableNode();
1912 GenTreePtr gtNewIconHandleNode(
1913 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1915 unsigned gtTokenToIconFlags(unsigned token);
1917 GenTreePtr gtNewIconEmbHndNode(void* value,
1920 unsigned handle1 = 0,
1921 void* handle2 = nullptr,
1922 void* compileTimeHandle = nullptr);
1924 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1925 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1926 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1927 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1929 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1931 GenTreePtr gtNewLconNode(__int64 value);
1933 GenTreePtr gtNewDconNode(double value);
1935 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1937 GenTreePtr gtNewZeroConNode(var_types type);
1939 GenTreePtr gtNewOneConNode(var_types type);
1942 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1943 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1946 GenTreeBlk* gtNewBlkOpNode(
1947 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1949 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1952 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1955 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1956 void gtSetObjGcInfo(GenTreeObj* objNode);
1957 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1958 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1960 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1962 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1964 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1965 CORINFO_METHOD_HANDLE handle,
1967 GenTreeArgList* args,
1968 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1970 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1972 GenTreeArgList* args,
1973 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1975 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1978 GenTreeArgList* args = nullptr);
1980 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1983 GenTreeSIMD* gtNewSIMDNode(
1984 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1985 GenTreeSIMD* gtNewSIMDNode(var_types type,
1988 SIMDIntrinsicID simdIntrinsicID,
1993 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1994 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
1995 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
1997 GenTreePtr gtNewCodeRef(BasicBlock* block);
1999 GenTreePtr gtNewFieldRef(
2000 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2002 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2004 GenTreeArgList* gtNewArgList(GenTreePtr op);
2005 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2006 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2008 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2009 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2010 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2011 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2013 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2015 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2017 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2018 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2019 CORINFO_ACCESS_FLAGS access,
2020 CORINFO_FIELD_INFO* pFieldInfo,
2022 CORINFO_CLASS_HANDLE structType,
2025 GenTreePtr gtNewNothingNode();
2027 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2029 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2031 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2033 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2035 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2037 //------------------------------------------------------------------------
2038 // Other GenTree functions
2040 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2042 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2043 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2044 // IntCnses with value `deepVarVal`.
2045 GenTreePtr gtCloneExpr(
2046 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2048 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2049 // `varNum` to int constants with value `varVal`.
2050 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2052 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2055 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2057 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2059 // Returns "true" iff the complexity (not formally defined, but first interpretation
2060 // is #of nodes in subtree) of "tree" is greater than "limit".
2061 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2062 // before they have been set.)
2063 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2065 bool gtCompareTree(GenTree* op1, GenTree* op2);
2067 GenTreePtr gtReverseCond(GenTree* tree);
2069 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2071 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2073 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2075 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2078 unsigned gtHashValue(GenTree* tree);
2080 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2083 void gtPrepareCost(GenTree* tree);
2084 bool gtIsLikelyRegVar(GenTree* tree);
2086 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2088 // Returns true iff the secondNode can be swapped with firstNode.
2089 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2091 unsigned gtSetEvalOrder(GenTree* tree);
2093 #if FEATURE_STACK_FP_X87
2095 void gtComputeFPlvls(GenTreePtr tree);
2096 #endif // FEATURE_STACK_FP_X87
2098 void gtSetStmtInfo(GenTree* stmt);
2100 // Returns "true" iff "node" has any of the side effects in "flags".
2101 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2103 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2104 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2106 // Appends 'expr' in front of 'list'
2107 // 'list' will typically start off as 'nullptr'
2108 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2109 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2111 void gtExtractSideEffList(GenTreePtr expr,
2113 unsigned flags = GTF_SIDE_EFFECT,
2114 bool ignoreRoot = false);
2116 GenTreePtr gtGetThisArg(GenTreePtr call);
2118 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2119 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2120 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2121 // the given "fldHnd", is such an object pointer.
2122 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2124 // Return true if call is a recursive call; return false otherwise.
2125 bool gtIsRecursiveCall(GenTreeCall* call)
2127 return (call->gtCallMethHnd == info.compMethodHnd);
2130 //-------------------------------------------------------------------------
2132 GenTreePtr gtFoldExpr(GenTreePtr tree);
2135 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2136 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2137 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2138 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2139 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2140 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2141 // optimizations for now.
2142 __attribute__((optnone))
2144 gtFoldExprConst(GenTreePtr tree);
2145 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2146 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2148 //-------------------------------------------------------------------------
2149 // Get the handle, if any.
2150 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2151 // Get the handle, and assert if not found.
2152 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2154 //-------------------------------------------------------------------------
2155 // Functions to display the trees
2158 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2160 void gtDispVN(GenTreePtr tree);
2161 void gtDispConst(GenTreePtr tree);
2162 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2163 void gtDispNodeName(GenTreePtr tree);
2164 void gtDispRegVal(GenTreePtr tree);
2176 void gtDispChild(GenTreePtr child,
2177 IndentStack* indentStack,
2179 __in_opt const char* msg = nullptr,
2180 bool topOnly = false);
2181 void gtDispTree(GenTreePtr tree,
2182 IndentStack* indentStack = nullptr,
2183 __in_opt const char* msg = nullptr,
2184 bool topOnly = false,
2185 bool isLIR = false);
2186 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2187 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2188 char* gtGetLclVarName(unsigned lclNum);
2189 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2190 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2191 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2192 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2193 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2194 void gtDispFieldSeq(FieldSeqNode* pfsn);
2196 void gtDispRange(LIR::ReadOnlyRange const& range);
2198 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2200 void gtDispLIRNode(GenTree* node);
2212 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2213 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2216 static fgWalkPreFn gtAssertColonCond;
2218 static fgWalkPreFn gtMarkColonCond;
2219 static fgWalkPreFn gtClearColonCond;
2221 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2222 bool gtHasCatchArg(GenTreePtr tree);
2223 bool gtHasUnmanagedCall(GenTreePtr tree);
2225 typedef ArrayStack<GenTree*> GenTreeStack;
2227 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2228 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2230 //=========================================================================
2231 // BasicBlock functions
2233 // This is a debug flag we will use to assert when creating block during codegen
2234 // as this interferes with procedure splitting. If you know what you're doing, set
2235 // it to true before creating the block. (DEBUG only)
2236 bool fgSafeBasicBlockCreation;
2239 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2242 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2243 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2247 XX The variables to be used by the code generator. XX
2249 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2250 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2254 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2255 // be placed in the stack frame and it's fields must be laid out sequentially.
2257 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2258 // a local variable that can be enregistered or placed in the stack frame.
2259 // The fields do not need to be laid out sequentially
2261 enum lvaPromotionType
2263 PROMOTION_TYPE_NONE, // The struct local is not promoted
2264 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2265 // and its field locals are independent of its parent struct local.
2266 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2267 // but its field locals depend on its parent struct local.
2270 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2271 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2273 /*****************************************************************************/
2275 enum FrameLayoutState
2278 INITIAL_FRAME_LAYOUT,
2279 PRE_REGALLOC_FRAME_LAYOUT,
2280 REGALLOC_FRAME_LAYOUT,
2281 TENTATIVE_FRAME_LAYOUT,
2286 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2287 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2288 bool lvaSortAgain; // true: We need to sort the lvaTable
2289 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2290 unsigned lvaCount; // total number of locals
2292 unsigned lvaRefCount; // total number of references to locals
2293 LclVarDsc* lvaTable; // variable descriptor table
2294 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2296 LclVarDsc** lvaRefSorted; // table sorted by refcount
2298 unsigned short lvaTrackedCount; // actual # of locals being tracked
2299 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2301 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2302 // Only for AMD64 System V cache the first caller stack homed argument.
2303 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2304 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2307 VARSET_TP lvaTrackedVars; // set of tracked variables
2309 #ifndef _TARGET_64BIT_
2310 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2312 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2314 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2315 // It that changes, this changes. VarSets from different epochs
2316 // cannot be meaningfully combined.
2318 unsigned GetCurLVEpoch()
2323 // reverse map of tracked number to var number
2324 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2326 #ifdef LEGACY_BACKEND
2327 // variable interference graph
2328 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2331 // variable preference graph
2332 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2336 // # of procs compiled a with double-aligned stack
2337 static unsigned s_lvaDoubleAlignedProcsCount;
2341 // Getters and setters for address-exposed and do-not-enregister local var properties.
2342 bool lvaVarAddrExposed(unsigned varNum);
2343 void lvaSetVarAddrExposed(unsigned varNum);
2344 bool lvaVarDoNotEnregister(unsigned varNum);
2346 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2347 enum DoNotEnregisterReason
2352 DNER_VMNeedsStackAddr,
2353 DNER_LiveInOutOfHandler,
2354 DNER_LiveAcrossUnmanagedCall,
2355 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2356 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2357 #ifdef JIT32_GCENCODER
2362 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2364 unsigned lvaVarargsHandleArg;
2366 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2368 #endif // _TARGET_X86_
2370 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2371 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2372 #if FEATURE_FIXED_OUT_ARGS
2373 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2375 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2376 // that tracks whether the lock has been taken
2378 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2379 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2380 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2382 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2383 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2385 #if FEATURE_FIXED_OUT_ARGS
2386 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2387 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2388 #endif // FEATURE_FIXED_OUT_ARGS
2391 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2392 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2393 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2394 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2395 // this variable to be this scratch word whenever struct promotion occurs.
2396 unsigned lvaPromotedStructAssemblyScratchVar;
2397 #endif // _TARGET_ARM_
2400 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2401 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2404 bool lvaGenericsContextUsed;
2406 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2407 // CORINFO_GENERICS_CTXT_FROM_THIS?
2408 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2410 //-------------------------------------------------------------------------
2411 // All these frame offsets are inter-related and must be kept in sync
2413 #if !FEATURE_EH_FUNCLETS
2414 // This is used for the callable handlers
2415 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2416 #endif // FEATURE_EH_FUNCLETS
2418 unsigned lvaCachedGenericContextArgOffs;
2419 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2422 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2424 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2426 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2427 // after the reg predict we will use a computed maxTmpSize
2428 // which is based upon the number of spill temps predicted by reg predict
2429 // All this is necessary because if we under-estimate the size of the spill
2430 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2432 // Pre codegen max spill temp size.
2433 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2435 //-------------------------------------------------------------------------
2437 unsigned lvaGetMaxSpillTempSize();
2439 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2440 #endif // _TARGET_ARM_
2441 void lvaAssignFrameOffsets(FrameLayoutState curState);
2442 void lvaFixVirtualFrameOffsets();
2444 #ifndef LEGACY_BACKEND
2445 void lvaUpdateArgsWithInitialReg();
2446 #endif // !LEGACY_BACKEND
2448 void lvaAssignVirtualFrameOffsetsToArgs();
2449 #ifdef UNIX_AMD64_ABI
2450 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2451 #else // !UNIX_AMD64_ABI
2452 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2453 #endif // !UNIX_AMD64_ABI
2454 void lvaAssignVirtualFrameOffsetsToLocals();
2455 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2456 #ifdef _TARGET_AMD64_
2457 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2458 bool lvaIsCalleeSavedIntRegCountEven();
2460 void lvaAlignFrame();
2461 void lvaAssignFrameOffsetsToPromotedStructs();
2462 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2465 void lvaDumpRegLocation(unsigned lclNum);
2466 void lvaDumpFrameLocation(unsigned lclNum);
2467 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2468 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2469 // layout state defined by lvaDoneFrameLayout
2472 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2473 // to avoid bugs from borderline cases.
2474 #define MAX_FrameSize 0x3FFFFFFF
2475 void lvaIncrementFrameSize(unsigned size);
2477 unsigned lvaFrameSize(FrameLayoutState curState);
2479 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2480 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2482 // Returns the caller-SP-relative offset for the local variable "varNum."
2483 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2485 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2486 int lvaGetSPRelativeOffset(unsigned varNum);
2488 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2489 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2491 //------------------------ For splitting types ----------------------------
2493 void lvaInitTypeRef();
2495 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2496 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2497 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2498 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2499 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2500 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2502 void lvaInitVarDsc(LclVarDsc* varDsc,
2504 CorInfoType corInfoType,
2505 CORINFO_CLASS_HANDLE typeHnd,
2506 CORINFO_ARG_LIST_HANDLE varList,
2507 CORINFO_SIG_INFO* varSig);
2509 static unsigned lvaTypeRefMask(var_types type);
2511 var_types lvaGetActualType(unsigned lclNum);
2512 var_types lvaGetRealType(unsigned lclNum);
2514 //-------------------------------------------------------------------------
2518 unsigned lvaLclSize(unsigned varNum);
2519 unsigned lvaLclExactSize(unsigned varNum);
2521 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2523 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2524 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2525 // the return result.
2526 bool lvaLclVarRefsAccum(
2527 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2529 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2530 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2531 // and (destructively) unions "trkedVars" into "*result".
2532 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2534 ALLVARSET_VALARG_TP allVars,
2535 VARSET_VALARG_TP trkdVars);
2537 bool lvaHaveManyLocals() const;
2539 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2540 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2541 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2544 void lvaSortByRefCount();
2545 void lvaDumpRefCounts();
2547 void lvaMarkLocalVars(BasicBlock* block);
2549 void lvaMarkLocalVars(); // Local variable ref-counting
2551 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2553 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2555 static fgWalkPreFn lvaIncRefCntsCB;
2556 void lvaIncRefCnts(GenTreePtr tree);
2558 static fgWalkPreFn lvaDecRefCntsCB;
2559 void lvaDecRefCnts(GenTreePtr tree);
2560 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2561 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2562 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2565 struct lvaStressLclFldArgs
2567 Compiler* m_pCompiler;
2571 static fgWalkPreFn lvaStressLclFldCB;
2572 void lvaStressLclFld();
2574 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2575 void lvaDispVarSet(VARSET_VALARG_TP set);
2580 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2582 int lvaFrameAddress(int varNum, bool* pFPbased);
2585 bool lvaIsParameter(unsigned varNum);
2586 bool lvaIsRegArgument(unsigned varNum);
2587 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2588 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2589 // that writes to arg0
2591 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2592 // (this is an overload of lvIsTemp because there are no temp parameters).
2593 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2594 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2595 bool lvaIsImplicitByRefLocal(unsigned varNum)
2597 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2598 LclVarDsc* varDsc = &(lvaTable[varNum]);
2599 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2601 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2604 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2608 // Returns true if this local var is a multireg struct
2609 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2611 // If the class is a TYP_STRUCT, get/set a class handle describing it
2613 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2614 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2616 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2618 // Info about struct fields
2619 struct lvaStructFieldInfo
2621 CORINFO_FIELD_HANDLE fldHnd;
2622 unsigned char fldOffset;
2623 unsigned char fldOrdinal;
2626 CORINFO_CLASS_HANDLE fldTypeHnd;
2629 // Info about struct to be promoted.
2630 struct lvaStructPromotionInfo
2632 CORINFO_CLASS_HANDLE typeHnd;
2634 bool requiresScratchVar;
2637 unsigned char fieldCnt;
2638 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2640 lvaStructPromotionInfo()
2641 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2646 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2647 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2648 lvaStructPromotionInfo* StructPromotionInfo,
2650 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2651 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2652 #if !defined(_TARGET_64BIT_)
2653 void lvaPromoteLongVars();
2654 #endif // !defined(_TARGET_64BIT_)
2655 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2656 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2657 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2658 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2659 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2660 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2661 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2663 BYTE* lvaGetGcLayout(unsigned varNum);
2664 bool lvaTypeIsGC(unsigned varNum);
2665 unsigned lvaGSSecurityCookie; // LclVar number
2666 bool lvaTempsHaveLargerOffsetThanVars();
2668 unsigned lvaSecurityObject; // variable representing the security object on the stack
2669 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2671 #if FEATURE_EH_FUNCLETS
2672 unsigned lvaPSPSym; // variable representing the PSPSym
2675 InlineInfo* impInlineInfo;
2676 InlineStrategy* m_inlineStrategy;
2678 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2679 Compiler* impInlineRoot();
2681 #if defined(DEBUG) || defined(INLINE_DATA)
2682 unsigned __int64 getInlineCycleCount()
2684 return m_compCycles;
2686 #endif // defined(DEBUG) || defined(INLINE_DATA)
2688 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2689 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2691 //=========================================================================
2693 //=========================================================================
2696 //---------------- Local variable ref-counting ----------------------------
2699 BasicBlock* lvaMarkRefsCurBlock;
2700 GenTreePtr lvaMarkRefsCurStmt;
2702 BasicBlock::weight_t lvaMarkRefsWeight;
2704 static fgWalkPreFn lvaMarkLclRefsCallback;
2705 void lvaMarkLclRefs(GenTreePtr tree);
2707 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2708 PerSsaArray lvHeapPerSsaData;
2709 unsigned lvHeapNumSsaNames;
2712 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2713 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2714 // not an SSA variable).
2715 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2717 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2718 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2720 assert(ssaNum < lvHeapNumSsaNames);
2721 return &lvHeapPerSsaData.GetRef(ssaNum);
2725 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2726 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2730 XX Imports the given method and converts it to semantic trees XX
2732 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2739 void impImport(BasicBlock* method);
2741 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2742 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2743 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2744 CORINFO_CLASS_HANDLE impGetStringClass();
2745 CORINFO_CLASS_HANDLE impGetObjectClass();
2747 //=========================================================================
2749 //=========================================================================
2752 //-------------------- Stack manipulation ---------------------------------
2754 unsigned impStkSize; // Size of the full stack
2756 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2758 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2760 struct SavedStack // used to save/restore stack contents.
2762 unsigned ssDepth; // number of values on stack
2763 StackEntry* ssTrees; // saved tree values
2766 bool impIsPrimitive(CorInfoType type);
2767 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2769 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2770 void impPushOnStackNoType(GenTreePtr tree);
2772 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2773 void impPushNullObjRefOnStack();
2774 StackEntry impPopStack();
2775 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2776 GenTreePtr impPopStack(typeInfo& ti);
2777 StackEntry& impStackTop(unsigned n = 0);
2779 void impSaveStackState(SavedStack* savePtr, bool copy);
2780 void impRestoreStackState(SavedStack* savePtr);
2782 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2783 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2784 CORINFO_CALL_INFO* pCallInfo);
2786 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2788 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2790 bool impCanPInvokeInline();
2791 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2792 void impCheckForPInvokeCall(
2793 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2794 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2795 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2797 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2798 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2799 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2801 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2802 CORINFO_METHOD_HANDLE calleeMethodHnd,
2803 CORINFO_CLASS_HANDLE delegateTypeHnd);
2805 var_types impImportCall(OPCODE opcode,
2806 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2807 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2809 GenTreePtr newobjThis,
2811 CORINFO_CALL_INFO* callInfo,
2812 IL_OFFSET rawILOffset);
2814 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2816 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2818 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2821 var_types impImportJitTestLabelMark(int numArgs);
2824 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2826 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2828 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2829 CORINFO_ACCESS_FLAGS access,
2830 CORINFO_FIELD_INFO* pFieldInfo,
2833 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2835 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2837 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2839 void impImportLeave(BasicBlock* block);
2840 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2841 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2842 CORINFO_CLASS_HANDLE clsHnd,
2843 CORINFO_METHOD_HANDLE method,
2844 CORINFO_SIG_INFO* sig,
2848 CorInfoIntrinsics* pIntrinsicID);
2849 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2850 CORINFO_SIG_INFO* sig,
2853 CorInfoIntrinsics intrinsicID);
2854 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2856 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2858 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2859 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2860 CORINFO_THIS_TRANSFORM transform);
2862 //----------------- Manipulating the trees and stmts ----------------------
2864 GenTreePtr impTreeList; // Trees for the BB being imported
2865 GenTreePtr impTreeLast; // The last tree for the current BB
2869 CHECK_SPILL_ALL = -1,
2870 CHECK_SPILL_NONE = -2
2874 void impBeginTreeList();
2875 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2876 void impEndTreeList(BasicBlock* block);
2877 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2878 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2879 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2880 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2881 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2882 void impAssignTempGen(unsigned tmp,
2885 GenTreePtr* pAfterStmt = nullptr,
2886 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2887 BasicBlock* block = nullptr);
2888 void impAssignTempGen(unsigned tmpNum,
2890 CORINFO_CLASS_HANDLE structHnd,
2892 GenTreePtr* pAfterStmt = nullptr,
2893 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2894 BasicBlock* block = nullptr);
2895 GenTreePtr impCloneExpr(GenTreePtr tree,
2897 CORINFO_CLASS_HANDLE structHnd,
2899 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2900 GenTreePtr impAssignStruct(GenTreePtr dest,
2902 CORINFO_CLASS_HANDLE structHnd,
2904 GenTreePtr* pAfterStmt = nullptr,
2905 BasicBlock* block = nullptr);
2906 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2908 CORINFO_CLASS_HANDLE structHnd,
2910 GenTreePtr* pAfterStmt = nullptr,
2911 BasicBlock* block = nullptr);
2913 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2914 CORINFO_CLASS_HANDLE structHnd,
2918 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2919 BYTE* gcLayout = nullptr,
2920 unsigned* numGCVars = nullptr,
2921 var_types* simdBaseType = nullptr);
2923 GenTreePtr impNormStructVal(GenTreePtr structVal,
2924 CORINFO_CLASS_HANDLE structHnd,
2926 bool forceNormalization = false);
2928 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2929 BOOL* pRuntimeLookup = nullptr,
2930 BOOL mustRestoreHandle = FALSE,
2931 BOOL importParent = FALSE);
2933 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2934 BOOL* pRuntimeLookup = nullptr,
2935 BOOL mustRestoreHandle = FALSE)
2937 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2940 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2941 CORINFO_LOOKUP* pLookup,
2943 void* compileTimeHandle);
2945 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2947 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2948 CORINFO_LOOKUP* pLookup,
2949 void* compileTimeHandle);
2951 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2953 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2954 CorInfoHelpFunc helper,
2956 GenTreeArgList* arg = nullptr,
2957 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
2959 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
2961 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2964 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
2965 CORINFO_CLASS_HANDLE typeClass,
2969 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
2970 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
2971 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
2972 static bool IsMathIntrinsic(GenTreePtr tree);
2975 //----------------- Importing the method ----------------------------------
2977 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
2980 unsigned impCurOpcOffs;
2981 const char* impCurOpcName;
2982 bool impNestedStackSpill;
2984 // For displaying instrs with generated native code (-n:B)
2985 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
2986 void impNoteLastILoffs();
2989 /* IL offset of the stmt currently being imported. It gets set to
2990 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
2991 updated at IL offsets for which we have to report mapping info.
2992 It also includes flag bits, so use jitGetILoffs()
2993 to get the actual IL offset value.
2996 IL_OFFSETX impCurStmtOffs;
2997 void impCurStmtOffsSet(IL_OFFSET offs);
2999 void impNoteBranchOffs();
3001 unsigned impInitBlockLineInfo();
3003 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3004 bool impIsThis(GenTreePtr obj);
3005 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3006 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3007 bool impIsAnySTLOC(OPCODE opcode)
3009 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3010 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3013 GenTreeArgList* impPopList(unsigned count,
3015 CORINFO_SIG_INFO* sig,
3016 GenTreeArgList* prefixTree = nullptr);
3018 GenTreeArgList* impPopRevList(unsigned count,
3020 CORINFO_SIG_INFO* sig,
3021 unsigned skipReverseCount = 0);
3024 * Get current IL offset with stack-empty info incoporated
3026 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3028 //---------------- Spilling the importer stack ----------------------------
3034 SavedStack pdSavedStack;
3035 ThisInitState pdThisPtrInit;
3038 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3039 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3041 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3042 ExpandArray<BYTE> impPendingBlockMembers;
3044 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3045 // Operates on the map in the top-level ancestor.
3046 BYTE impGetPendingBlockMember(BasicBlock* blk)
3048 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3051 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3052 // Operates on the map in the top-level ancestor.
3053 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3055 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3058 bool impCanReimport;
3060 bool impSpillStackEntry(unsigned level,
3064 bool bAssertOnRecursion,
3069 void impSpillStackEnsure(bool spillLeaves = false);
3070 void impEvalSideEffects();
3071 void impSpillSpecialSideEff();
3072 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3073 void impSpillValueClasses();
3074 void impSpillEvalStack();
3075 static fgWalkPreFn impFindValueClasses;
3076 void impSpillLclRefs(ssize_t lclNum);
3078 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3080 void impImportBlockCode(BasicBlock* block);
3082 void impReimportMarkBlock(BasicBlock* block);
3083 void impReimportMarkSuccessors(BasicBlock* block);
3085 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3087 void impImportBlockPending(BasicBlock* block);
3089 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3090 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3091 // for the block, but instead, just re-uses the block's existing EntryState.
3092 void impReimportBlockPending(BasicBlock* block);
3094 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3096 void impImportBlock(BasicBlock* block);
3098 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3099 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3100 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3101 // the variables that will be used -- and for all the predecessors of those successors, and the
3102 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3103 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3104 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3105 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3106 // of local variable numbers, so we represent them with the base local variable number), returns that.
3107 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3108 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3109 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3110 // on which kind of member of the clique the block is).
3111 unsigned impGetSpillTmpBase(BasicBlock* block);
3113 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3114 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3115 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3116 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3117 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3118 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3119 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3120 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3121 // successors receive a native int. Similarly float and double are unified to double.
3122 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3123 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3124 // predecessors, so they insert an upcast if needed).
3125 void impReimportSpillClique(BasicBlock* block);
3127 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3128 // block, and represent the predecessor and successor members of the clique currently being computed.
3129 // *** Access to these will need to be locked in a parallel compiler.
3130 ExpandArray<BYTE> impSpillCliquePredMembers;
3131 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3139 // Abstract class for receiving a callback while walking a spill clique
3140 class SpillCliqueWalker
3143 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3146 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3147 class SetSpillTempsBase : public SpillCliqueWalker
3152 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3155 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3158 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3159 class ReimportSpillClique : public SpillCliqueWalker
3164 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3167 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3170 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3171 // predecessor or successor within the spill clique
3172 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3174 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3175 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3176 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3177 void impRetypeEntryStateTemps(BasicBlock* blk);
3179 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3180 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3182 void impPushVar(GenTree* op, typeInfo tiRetVal);
3183 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3184 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3186 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3188 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3189 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3190 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3193 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3196 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3197 struct BlockListNode
3200 BlockListNode* m_next;
3201 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3204 void* operator new(size_t sz, Compiler* comp);
3206 BlockListNode* impBlockListNodeFreeList;
3208 BlockListNode* AllocBlockListNode();
3209 void FreeBlockListNode(BlockListNode* node);
3211 bool impIsValueType(typeInfo* pTypeInfo);
3212 var_types mangleVarArgsType(var_types type);
3215 regNumber getCallArgIntRegister(regNumber floatReg);
3216 regNumber getCallArgFloatRegister(regNumber intReg);
3217 #endif // FEATURE_VARARG
3220 static unsigned jitTotalMethodCompiled;
3224 static LONG jitNestingLevel;
3227 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3229 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3231 // STATIC inlining decision based on the IL code.
3232 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3233 CORINFO_METHOD_INFO* methInfo,
3235 InlineResult* inlineResult);
3237 void impCheckCanInline(GenTreePtr call,
3238 CORINFO_METHOD_HANDLE fncHandle,
3240 CORINFO_CONTEXT_HANDLE exactContextHnd,
3241 InlineCandidateInfo** ppInlineCandidateInfo,
3242 InlineResult* inlineResult);
3244 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3245 GenTreePtr curArgVal,
3247 InlineResult* inlineResult);
3249 void impInlineInitVars(InlineInfo* pInlineInfo);
3251 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3253 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3255 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3257 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3258 GenTreePtr variableBeingDereferenced,
3259 InlArgInfo* inlArgInfo);
3261 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3263 bool impTailCallRetTypeCompatible(var_types callerRetType,
3264 CORINFO_CLASS_HANDLE callerRetTypeClass,
3265 var_types calleeRetType,
3266 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3268 bool impIsTailCallILPattern(bool tailPrefixed,
3270 const BYTE* codeAddrOfNextOpcode,
3271 const BYTE* codeEnd,
3273 bool* IsCallPopRet = nullptr);
3275 bool impIsImplicitTailCallCandidate(
3276 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3279 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3280 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3284 XX Info about the basic-blocks, their contents and the flow analysis XX
3286 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3287 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3291 BasicBlock* fgFirstBB; // Beginning of the basic block list
3292 BasicBlock* fgLastBB; // End of the basic block list
3293 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3294 #if FEATURE_EH_FUNCLETS
3295 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3297 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3299 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3300 unsigned fgEdgeCount; // # of control flow edges between the BBs
3301 unsigned fgBBcount; // # of BBs in the method
3303 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3305 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3306 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3307 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3308 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3310 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3311 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3312 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3313 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3314 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3315 // index). The arrays are of size fgBBNumMax + 1.
3316 unsigned* fgDomTreePreOrder;
3317 unsigned* fgDomTreePostOrder;
3319 bool fgBBVarSetsInited;
3321 // Allocate array like T* a = new T[fgBBNumMax + 1];
3322 // Using helper so we don't keep forgetting +1.
3323 template <typename T>
3324 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3326 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3329 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3330 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3331 // cannot be meaningfully combined. Note that new blocks can be created with higher
3332 // block numbers without changing the basic block epoch. These blocks *cannot*
3333 // participate in a block set until the blocks are all renumbered, causing the epoch
3334 // to change. This is useful if continuing to use previous block sets is valuable.
3335 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3336 unsigned fgCurBBEpoch;
3338 unsigned GetCurBasicBlockEpoch()
3340 return fgCurBBEpoch;
3343 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3344 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3345 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3346 unsigned fgCurBBEpochSize;
3348 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3349 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3350 unsigned fgBBSetCountInSizeTUnits;
3352 void NewBasicBlockEpoch()
3354 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3356 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3358 fgCurBBEpochSize = fgBBNumMax + 1;
3359 fgBBSetCountInSizeTUnits =
3360 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3363 // All BlockSet objects are now invalid!
3364 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3365 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3369 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3370 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3371 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3372 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3374 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3375 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3376 // array of size_t bitsets), then print that out.
3377 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3384 void EnsureBasicBlockEpoch()
3386 if (fgCurBBEpochSize != fgBBNumMax + 1)
3388 NewBasicBlockEpoch();
3392 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3393 void fgEnsureFirstBBisScratch();
3394 bool fgFirstBBisScratch();
3395 bool fgBBisScratch(BasicBlock* block);
3397 void fgExtendEHRegionBefore(BasicBlock* block);
3398 void fgExtendEHRegionAfter(BasicBlock* block);
3400 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3402 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3404 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3407 BasicBlock* nearBlk,
3408 bool putInFilter = false,
3409 bool runRarely = false,
3410 bool insertAtEnd = false);
3412 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3414 bool runRarely = false,
3415 bool insertAtEnd = false);
3417 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3419 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3420 BasicBlock* afterBlk,
3421 unsigned xcptnIndex,
3422 bool putInTryRegion);
3424 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3425 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3426 void fgUnlinkBlock(BasicBlock* block);
3428 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3429 bool fgMultipleNots;
3432 bool fgModified; // True if the flow graph has been modified recently
3433 bool fgComputePredsDone; // Have we computed the bbPreds list
3434 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3435 bool fgDomsComputed; // Have we computed the dominator sets?
3437 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3438 bool fgHasPostfix; // any postfix ++/-- found?
3439 unsigned fgIncrCount; // number of increment nodes found
3441 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3445 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3446 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3449 bool fgRemoveRestOfBlock; // true if we know that we will throw
3450 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3452 // There are two modes for ordering of the trees.
3453 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3454 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3455 // by traversing the tree according to the order of the operands.
3456 // - In FGOrderLinear, the dominant ordering is the linear order.
3463 FlowGraphOrder fgOrder;
3465 // The following are boolean flags that keep track of the state of internal data structures
3467 bool fgStmtListThreaded;
3468 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3469 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3470 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3471 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3472 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3473 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3474 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3475 // This is derived from the profile data
3476 // or is BB_UNITY_WEIGHT when we don't have profile data
3478 #if FEATURE_EH_FUNCLETS
3479 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3480 #endif // FEATURE_EH_FUNCLETS
3482 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3483 // since fgMorphTree can be called from several places
3484 bool fgExpandInline; // indicates that we are creating tree for the inliner
3486 bool impBoxTempInUse; // the temp below is valid and available
3487 unsigned impBoxTemp; // a temporary that is used for boxing
3490 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3491 // and we are trying to compile again in a "safer", minopts mode?
3495 unsigned impInlinedCodeSize;
3498 //-------------------------------------------------------------------------
3506 void fgRemoveEmptyFinally();
3508 void fgCloneFinally();
3510 GenTreePtr fgGetCritSectOfStaticMethod();
3512 #if !defined(_TARGET_X86_)
3514 void fgAddSyncMethodEnterExit();
3516 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3518 void fgConvertSyncReturnToLeave(BasicBlock* block);
3520 #endif // !_TARGET_X86_
3522 void fgAddReversePInvokeEnterExit();
3524 bool fgMoreThanOneReturnBlock();
3526 // The number of separate return points in the method.
3527 unsigned fgReturnCount;
3529 void fgAddInternal();
3531 bool fgFoldConditional(BasicBlock* block);
3533 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3534 void fgMorphBlocks();
3536 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3538 void fgCheckArgCnt();
3539 void fgSetOptions();
3542 static fgWalkPreFn fgAssertNoQmark;
3543 void fgPreExpandQmarkChecks(GenTreePtr expr);
3544 void fgPostExpandQmarkChecks();
3545 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3548 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3550 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3551 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3552 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3553 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3554 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3556 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3557 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3558 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3559 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3561 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3562 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3563 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3564 void fgExpandQmarkNodes();
3568 // Do "simple lowering." This functionality is (conceptually) part of "general"
3569 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3570 void fgSimpleLowering();
3572 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3574 GenTreePtr fgInitThisClass();
3576 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3578 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3580 void fgLocalVarLiveness();
3582 void fgLocalVarLivenessInit();
3584 #ifdef LEGACY_BACKEND
3585 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3587 void fgPerNodeLocalVarLiveness(GenTree* node);
3589 void fgPerBlockLocalVarLiveness();
3591 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3593 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3595 // This is used in the liveness computation, as a temporary. When we use the
3596 // arbitrary-length VarSet representation, it is better not to allocate a new one
3598 VARSET_TP fgMarkIntfUnionVS;
3600 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3602 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3604 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3606 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3608 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3610 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3612 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3613 GenTreePtr startNode,
3615 VARSET_VALARG_TP volatileVars,
3616 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3618 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3620 bool fgRemoveDeadStore(GenTree** pTree,
3624 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3626 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3628 // For updating liveset during traversal AFTER fgComputeLife has completed
3629 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3630 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3632 // Returns the set of live variables after endTree,
3633 // assuming that liveSet is the set of live variables BEFORE tree.
3634 // Requires that fgComputeLife has completed, and that tree is in the same
3635 // statement as endTree, and that it comes before endTree in execution order
3637 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3639 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3640 while (tree != nullptr && tree != endTree->gtNext)
3642 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3643 tree = tree->gtNext;
3645 assert(tree == endTree->gtNext);
3649 void fgInterBlockLocalVarLiveness();
3651 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3652 // "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
3653 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3654 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3655 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3656 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3657 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3659 if (m_opAsgnVarDefSsaNums == nullptr)
3661 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3663 return m_opAsgnVarDefSsaNums;
3666 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3667 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3668 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3670 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3672 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3673 // Except: assumes that lcl is a def, and if it is
3674 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3675 // rather than the "use" SSA number recorded in the tree "lcl".
3676 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3678 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3679 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3680 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3681 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3682 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3684 // (byref addrS1 = &s1,
3685 // *(addrS1 * offsetof(f0)) = s2f0,
3687 // *(addrS1 * offsetof(fn)) = s2fn)
3689 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3690 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3691 // give it SSA names and value numbers?
3693 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3694 // end with an instance of the structure below, whose fields are described in the declaration.
3695 struct IndirectAssignmentAnnotation
3697 unsigned m_lclNum; // The local num that is being indirectly assigned.
3698 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3699 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3700 // be the singleton field sequence "g". The individual assignments would
3701 // further append the fields of "s.g" to that.
3702 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3703 // structure has a single field).
3704 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3705 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3708 IndirectAssignmentAnnotation(unsigned lclNum,
3709 FieldSeqNode* fldSeq,
3711 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3712 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3713 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3717 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3718 NodeToIndirAssignMap;
3719 NodeToIndirAssignMap* m_indirAssignMap;
3720 NodeToIndirAssignMap* GetIndirAssignMap()
3722 if (m_indirAssignMap == nullptr)
3724 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3725 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3726 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3728 return m_indirAssignMap;
3731 // Performs SSA conversion.
3734 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3735 void fgResetForSsa();
3737 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3739 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3740 inline bool fgExcludeFromSsa(unsigned lclNum);
3742 // The value numbers for this compilation.
3743 ValueNumStore* vnStore;
3746 ValueNumStore* GetValueNumStore()
3751 // Do value numbering (assign a value number to each
3753 void fgValueNumber();
3755 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3756 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3757 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3758 // match the element type of the array or fldSeq. When this type doesn't match
3759 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3761 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3764 FieldSeqNode* fldSeq,
3768 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3769 // has been parsed to yield the other input arguments. If evaluation of the address
3770 // can raise exceptions, those should be captured in the exception set "excVN."
3771 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3772 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3773 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3774 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3775 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3777 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3778 CORINFO_CLASS_HANDLE elemTypeEq,
3782 FieldSeqNode* fldSeq);
3784 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3785 // by evaluating the array index expression "tree". Returns the value number resulting from
3786 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3787 // "GT_IND" that does the dereference, and it is given the returned value number.
3788 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3790 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3792 // Utility functions for fgValueNumber.
3794 // Perform value-numbering for the trees in "blk".
3795 void fgValueNumberBlock(BasicBlock* blk);
3797 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3798 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3799 // assumed for the heap at the start "entryBlk".
3800 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3802 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3803 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3805 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3806 // value in that SSA #.
3807 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3809 // The input 'tree' is a leaf node that is a constant
3810 // Assign the proper value number to the tree
3811 void fgValueNumberTreeConst(GenTreePtr tree);
3813 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3814 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3816 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3818 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3820 // Does value-numbering for a block assignment.
3821 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3823 // Does value-numbering for a cast tree.
3824 void fgValueNumberCastTree(GenTreePtr tree);
3826 // Does value-numbering for an intrinsic tree.
3827 void fgValueNumberIntrinsic(GenTreePtr tree);
3829 // Does value-numbering for a call. We interpret some helper calls.
3830 void fgValueNumberCall(GenTreeCall* call);
3832 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3833 void fgUpdateArgListVNs(GenTreeArgList* args);
3835 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3836 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3838 // Requires "helpCall" to be a helper call. Assigns it a value number;
3839 // we understand the semantics of some of the calls. Returns "true" if
3840 // the call may modify the heap (we assume arbitrary memory side effects if so).
3841 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3843 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3844 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3846 // This is the current value number for the "Heap" implicit variable while
3847 // doing value numbering. This is the value number under the "liberal" interpretation
3848 // of heap values; the "conservative" interpretation needs no VN, since every access of
3849 // the heap yields an unknown value.
3850 ValueNum fgCurHeapVN;
3852 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3853 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3854 // is 1, and the rest is an encoding of "elemTyp".
3855 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3857 if (elemStructType != nullptr)
3859 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3860 varTypeIsIntegral(elemTyp));
3861 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3862 return elemStructType;
3866 elemTyp = varTypeUnsignedToSigned(elemTyp);
3867 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3870 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3871 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3872 // the struct type of the element).
3873 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3875 size_t clsHndVal = size_t(clsHnd);
3876 if (clsHndVal & 0x1)
3878 return var_types(clsHndVal >> 1);
3886 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3887 var_types getJitGCType(BYTE gcType);
3889 enum structPassingKind
3891 SPK_Unknown, // Invalid value, never returned
3892 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3893 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3894 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3895 // parameters registers are used, then the stack will be used)
3896 // for X86 passed on the stack, for ARM32 passed in registers
3897 // or the stack or split between registers and the stack.
3898 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3900 }; // The struct is passed/returned by reference to a copy/buffer.
3902 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3903 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3904 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3905 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3907 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3909 // Get the type that is used to pass values of the given struct type.
3910 // If you have already retrieved the struct size then pass it as the optional third argument
3912 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3913 structPassingKind* wbPassStruct,
3914 unsigned structSize = 0);
3916 // Get the type that is used to return values of the given struct type.
3917 // If you have already retrieved the struct size then pass it as the optional third argument
3919 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3920 structPassingKind* wbPassStruct = nullptr,
3921 unsigned structSize = 0);
3924 // Print a representation of "vnp" or "vn" on standard output.
3925 // If "level" is non-zero, we also print out a partial expansion of the value.
3926 void vnpPrint(ValueNumPair vnp, unsigned level);
3927 void vnPrint(ValueNum vn, unsigned level);
3930 // Dominator computation member functions
3931 // Not exposed outside Compiler
3933 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3935 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3937 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3938 // flow graph. We first assume the fields bbIDom on each
3939 // basic block are invalid. This computation is needed later
3940 // by fgBuildDomTree to build the dominance tree structure.
3941 // Based on: A Simple, Fast Dominance Algorithm
3942 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3944 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3945 // Note: this is relatively slow compared to calling fgDominate(),
3946 // especially if dealing with a single block versus block check.
3948 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3950 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3952 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3954 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3956 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3958 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3959 // processed in topological sort, this function takes care of that.
3961 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3963 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3964 // Returns this as a set.
3966 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3967 // root nodes. Returns this as a set.
3970 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
3973 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
3974 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
3977 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
3978 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
3979 // && postOrder(A) >= postOrder(B) making the computation O(1).
3980 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
3982 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
3984 void fgUpdateChangedFlowGraph();
3987 // Compute the predecessors of the blocks in the control flow graph.
3988 void fgComputePreds();
3990 // Remove all predecessor information.
3991 void fgRemovePreds();
3993 // Compute the cheap flow graph predecessors lists. This is used in some early phases
3994 // before the full predecessors lists are computed.
3995 void fgComputeCheapPreds();
3998 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4000 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4010 // Initialize the per-block variable sets (used for liveness analysis).
4011 void fgInitBlockVarSets();
4013 // true if we've gone through and created GC Poll calls.
4014 bool fgGCPollsCreated;
4015 void fgMarkGCPollBlocks();
4016 void fgCreateGCPolls();
4017 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4019 // Requires that "block" is a block that returns from
4020 // a finally. Returns the number of successors (jump targets of
4021 // of blocks in the covered "try" that did a "LEAVE".)
4022 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4024 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4025 // a finally. Returns its "i"th successor (jump targets of
4026 // of blocks in the covered "try" that did a "LEAVE".)
4027 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4028 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4031 // Factor out common portions of the impls of the methods above.
4032 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4035 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4036 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4037 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4038 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4039 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4040 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4041 // we leave the entry associated with the block, but it will no longer be accessed.)
4042 struct SwitchUniqueSuccSet
4044 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4045 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4048 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4049 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4050 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4051 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4054 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4055 BlockToSwitchDescMap;
4058 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4059 // iteration over only the distinct successors.
4060 BlockToSwitchDescMap* m_switchDescMap;
4063 BlockToSwitchDescMap* GetSwitchDescMap()
4065 if (m_switchDescMap == nullptr)
4067 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4069 return m_switchDescMap;
4072 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4073 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4074 // we don't accidentally look up and return the wrong switch data.
4075 void InvalidateUniqueSwitchSuccMap()
4077 m_switchDescMap = nullptr;
4080 // Requires "switchBlock" to be a block that ends in a switch. Returns
4081 // the corresponding SwitchUniqueSuccSet.
4082 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4084 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4085 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4086 // remove it from "this", and ensure that "to" is a member.
4087 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4089 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4090 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4092 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4094 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4096 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4098 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4100 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4102 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4104 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4106 void fgRemoveBlockAsPred(BasicBlock* block);
4108 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4110 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4112 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4114 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4116 flowList* fgAddRefPred(BasicBlock* block,
4117 BasicBlock* blockPred,
4118 flowList* oldEdge = nullptr,
4119 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4122 void fgFindBasicBlocks();
4124 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4126 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4128 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4129 bool putInTryRegion,
4130 BasicBlock* startBlk,
4132 BasicBlock* nearBlk,
4133 BasicBlock* jumpBlk,
4136 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4138 void fgRemoveEmptyBlocks();
4140 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4142 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4144 void fgCreateLoopPreHeader(unsigned lnum);
4146 void fgUnreachableBlock(BasicBlock* block);
4148 void fgRemoveConditionalJump(BasicBlock* block);
4150 BasicBlock* fgLastBBInMainFunction();
4152 BasicBlock* fgEndBBAfterMainFunction();
4154 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4156 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4158 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4160 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4162 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4164 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4166 bool fgRenumberBlocks();
4168 bool fgExpandRarelyRunBlocks();
4170 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4172 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4174 enum FG_RELOCATE_TYPE
4176 FG_RELOCATE_TRY, // relocate the 'try' region
4177 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4179 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4181 #if FEATURE_EH_FUNCLETS
4182 #if defined(_TARGET_ARM_)
4183 void fgClearFinallyTargetBit(BasicBlock* block);
4184 #endif // defined(_TARGET_ARM_)
4185 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4186 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4187 void fgInsertFuncletPrologBlock(BasicBlock* block);
4188 void fgCreateFuncletPrologBlocks();
4189 void fgCreateFunclets();
4190 #else // !FEATURE_EH_FUNCLETS
4191 bool fgRelocateEHRegions();
4192 #endif // !FEATURE_EH_FUNCLETS
4194 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4196 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4198 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4200 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4202 bool fgOptimizeEmptyBlock(BasicBlock* block);
4204 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4206 bool fgOptimizeBranch(BasicBlock* bJump);
4208 bool fgOptimizeSwitchBranches(BasicBlock* block);
4210 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4212 bool fgOptimizeSwitchJumps();
4214 void fgPrintEdgeWeights();
4216 void fgComputeEdgeWeights();
4218 void fgReorderBlocks();
4220 void fgDetermineFirstColdBlock();
4222 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4224 bool fgUpdateFlowGraph(bool doTailDup = false);
4226 void fgFindOperOrder();
4228 // method that returns if you should split here
4229 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4231 void fgSetBlockOrder();
4233 void fgRemoveReturnBlock(BasicBlock* block);
4235 /* Helper code that has been factored out */
4236 inline void fgConvertBBToThrowBB(BasicBlock* block);
4238 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4239 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4240 GenTreePtr fgMakeTmpArgNode(
4241 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4243 // The following check for loops that don't execute calls
4244 bool fgLoopCallMarked;
4246 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4247 void fgLoopCallMark();
4249 void fgMarkLoopHead(BasicBlock* block);
4251 unsigned fgGetCodeEstimate(BasicBlock* block);
4254 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4255 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4256 bool fgDumpFlowGraph(Phases phase);
4258 #endif // DUMP_FLOWGRAPHS
4263 void fgDispBBLiveness(BasicBlock* block);
4264 void fgDispBBLiveness();
4265 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4266 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4267 void fgDispBasicBlocks(bool dumpTrees = false);
4268 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4269 void fgDumpBlock(BasicBlock* block);
4270 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4272 static fgWalkPreFn fgStress64RsltMulCB;
4273 void fgStress64RsltMul();
4274 void fgDebugCheckUpdate();
4275 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4276 void fgDebugCheckBlockLinks();
4277 void fgDebugCheckLinks(bool morphTrees = false);
4278 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4279 void fgDebugCheckFlags(GenTreePtr tree);
4280 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4281 void fgDebugCheckTryFinallyExits();
4284 #ifdef LEGACY_BACKEND
4285 static void fgOrderBlockOps(GenTreePtr tree,
4289 GenTreePtr* opsPtr, // OUT
4290 regMaskTP* regsPtr); // OUT
4291 #endif // LEGACY_BACKEND
4293 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4294 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4296 inline bool fgIsInlining()
4298 return fgExpandInline;
4301 void fgTraverseRPO();
4303 //--------------------- Walking the trees in the IR -----------------------
4308 fgWalkPreFn* wtprVisitorFn;
4309 fgWalkPostFn* wtpoVisitorFn;
4310 void* pCallbackData; // user-provided data
4311 bool wtprLclsOnly; // whether to only visit lclvar nodes
4312 GenTreePtr parent; // parent of current node, provided to callback
4313 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4315 bool printModified; // callback can use this
4319 template <bool computeStack>
4320 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4322 // general purpose tree-walker that is capable of doing pre- and post- order
4323 // callbacks at the same time
4324 template <bool doPreOrder, bool doPostOrder>
4325 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4327 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4328 fgWalkPreFn* visitor,
4329 void* pCallBackData = nullptr,
4330 bool lclVarsOnly = false,
4331 bool computeStack = false);
4333 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4334 fgWalkPreFn* preVisitor,
4335 fgWalkPostFn* postVisitor,
4336 void* pCallBackData = nullptr);
4338 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4342 template <bool computeStack>
4343 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4345 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4346 fgWalkPostFn* visitor,
4347 void* pCallBackData = nullptr,
4348 bool computeStack = false);
4350 // An fgWalkPreFn that looks for expressions that have inline throws in
4351 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4352 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4353 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4354 // properly propagated to parent trees). It returns WALK_CONTINUE
4356 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4357 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4358 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4360 /**************************************************************************
4362 *************************************************************************/
4365 friend class SsaBuilder;
4366 friend struct ValueNumberState;
4368 //--------------------- Detect the basic blocks ---------------------------
4370 BasicBlock** fgBBs; // Table of pointers to the BBs
4372 void fgInitBBLookup();
4373 BasicBlock* fgLookupBB(unsigned addr);
4375 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4377 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4379 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4381 void fgLinkBasicBlocks();
4383 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4385 void fgCheckBasicBlockControlFlow();
4387 void fgControlFlowPermitted(BasicBlock* blkSrc,
4388 BasicBlock* blkDest,
4389 BOOL IsLeave = false /* is the src a leave block */);
4391 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4393 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4395 void fgAdjustForAddressExposedOrWrittenThis();
4397 bool fgProfileData_ILSizeMismatch;
4398 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4399 ULONG fgProfileBufferCount;
4400 ULONG fgNumProfileRuns;
4402 unsigned fgStressBBProf()
4405 unsigned result = JitConfig.JitStressBBProf();
4408 if (compStressCompile(STRESS_BB_PROFILE, 15))
4419 bool fgHaveProfileData();
4420 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4422 bool fgIsUsingProfileWeights()
4424 return (fgHaveProfileData() || fgStressBBProf());
4426 void fgInstrumentMethod();
4428 //-------- Insert a statement at the start or end of a basic block --------
4432 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4436 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4438 public: // Used by linear scan register allocation
4439 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4442 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4443 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4445 public: // Used by linear scan register allocation
4446 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4449 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4451 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4453 // Create a new temporary variable to hold the result of *ppTree,
4454 // and transform the graph accordingly.
4455 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4456 GenTree* fgMakeMultiUse(GenTree** ppTree);
4459 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4460 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4461 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4463 //-------- Determine the order in which the trees will be evaluated -------
4465 unsigned fgTreeSeqNum;
4466 GenTree* fgTreeSeqLst;
4467 GenTree* fgTreeSeqBeg;
4469 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4470 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4471 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4472 void fgSetStmtSeq(GenTree* tree);
4473 void fgSetBlockOrder(BasicBlock* block);
4475 //------------------------- Morphing --------------------------------------
4477 unsigned fgPtrArgCntCur;
4478 unsigned fgPtrArgCntMax;
4479 hashBv* fgOutgoingArgTemps;
4480 hashBv* fgCurrentlyInUseArgTemps;
4482 bool compCanEncodePtrArgCntMax();
4484 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4487 void fgMoveOpsLeft(GenTreePtr tree);
4490 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4492 bool fgIsThrow(GenTreePtr tree);
4494 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4495 bool fgIsBlockCold(BasicBlock* block);
4497 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4499 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4501 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4503 bool fgMorphRelopToQmark(GenTreePtr tree);
4505 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4506 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4507 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4508 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4509 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4510 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4511 // small; hence the other fields of MorphAddrContext.
4512 enum MorphAddrContextKind
4517 struct MorphAddrContext
4519 MorphAddrContextKind m_kind;
4520 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4521 // top-level indirection and here have been constants.
4522 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4523 // In that case, is the sum of those constant offsets.
4525 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4530 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4531 static MorphAddrContext s_CopyBlockMAC;
4534 GenTreePtr fgCopySIMDNode(GenTreeSIMD* simdNode);
4535 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4536 var_types* baseTypeOut,
4538 unsigned* simdSizeOut,
4539 bool ignoreUsedInSIMDIntrinsic = false);
4540 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4541 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4542 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4543 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4545 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4546 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4547 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4549 #endif // FEATURE_SIMD
4550 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4551 GenTreePtr fgMorphCast(GenTreePtr tree);
4552 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4553 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4555 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4558 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4559 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4561 void fgFixupStructReturn(GenTreePtr call);
4562 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4563 bool fgAddrCouldBeNull(GenTreePtr addr);
4564 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4565 bool fgCanFastTailCall(GenTreeCall* call);
4566 void fgMorphTailCall(GenTreeCall* call);
4567 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4568 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4569 fgArgTabEntryPtr argTabEntry,
4571 IL_OFFSETX callILOffset,
4572 GenTreePtr tmpAssignmentInsertionPoint,
4573 GenTreePtr paramAssignmentInsertionPoint);
4574 static int fgEstimateCallStackSize(GenTreeCall* call);
4575 GenTreePtr fgMorphCall(GenTreeCall* call);
4576 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4577 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4579 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4580 static fgWalkPreFn fgFindNonInlineCandidate;
4582 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4583 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4584 void fgAssignSetVarDef(GenTreePtr tree);
4585 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4586 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4587 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4588 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4589 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4590 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4591 void fgMorphUnsafeBlk(GenTreeObj* obj);
4592 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4593 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4594 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4595 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4596 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4597 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4598 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4600 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4601 GenTreePtr fgMorphConst(GenTreePtr tree);
4604 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4607 #if LOCAL_ASSERTION_PROP
4608 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4609 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4611 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4613 GenTreeStmt* fgMorphStmt;
4615 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4616 // used when morphing big offset.
4618 //----------------------- Liveness analysis -------------------------------
4620 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4621 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4623 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4624 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4625 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4627 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4629 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4630 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4632 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4633 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4635 void fgExtendDbgScopes();
4636 void fgExtendDbgLifetimes();
4639 void fgDispDebugScopes();
4642 //-------------------------------------------------------------------------
4644 // The following keeps track of any code we've added for things like array
4645 // range checking or explicit calls to enable GC, and so on.
4650 AddCodeDsc* acdNext;
4651 BasicBlock* acdDstBlk; // block to which we jump
4653 SpecialCodeKind acdKind; // what kind of a special block is this?
4654 unsigned short acdStkLvl;
4658 static unsigned acdHelper(SpecialCodeKind codeKind);
4660 AddCodeDsc* fgAddCodeList;
4662 bool fgRngChkThrowAdded;
4663 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4665 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4667 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4670 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4673 bool fgIsCodeAdded();
4675 bool fgIsThrowHlpBlk(BasicBlock* block);
4676 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4678 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4680 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4681 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4682 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4683 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4684 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4686 #if FEATURE_MULTIREG_RET
4687 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4688 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4689 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4690 #endif // FEATURE_MULTIREG_RET
4692 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4695 static fgWalkPreFn fgDebugCheckInlineCandidates;
4698 void fgPromoteStructs();
4699 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4700 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4701 void fgMarkImplicitByRefArgs();
4702 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4703 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4704 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4705 void fgMarkAddressExposedLocals();
4706 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4708 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4710 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4712 // The given local variable, required to be a struct variable, is being assigned via
4713 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4714 // the variable is not enregistered, and is therefore not promoted independently.
4715 void fgLclFldAssign(unsigned lclNum);
4717 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4718 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4719 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4720 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4723 bool fgPrintInlinedMethods;
4726 bool fgIsBigOffset(size_t offset);
4728 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4729 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4730 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4731 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4732 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4735 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4736 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4740 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4741 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4748 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4751 void optRemoveRangeCheck(
4752 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4753 bool optIsRangeCheckRemovable(GenTreePtr tree);
4756 static fgWalkPreFn optValidRangeCheckIndex;
4757 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4760 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4762 /**************************************************************************
4764 *************************************************************************/
4767 // Do hoisting for all loops.
4768 void optHoistLoopCode();
4770 // To represent sets of VN's that have already been hoisted in outer loops.
4771 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4772 typedef VNToBoolMap VNSet;
4774 struct LoopHoistContext
4777 // The set of variables hoisted in the current loop (or nullptr if there are none).
4778 VNSet* m_pHoistedInCurLoop;
4781 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4782 VNSet m_hoistedInParentLoops;
4783 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4784 // Previous decisions on loop-invariance of value numbers in the current loop.
4785 VNToBoolMap m_curLoopVnInvariantCache;
4787 VNSet* GetHoistedInCurLoop(Compiler* comp)
4789 if (m_pHoistedInCurLoop == nullptr)
4791 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4793 return m_pHoistedInCurLoop;
4796 VNSet* ExtractHoistedInCurLoop()
4798 VNSet* res = m_pHoistedInCurLoop;
4799 m_pHoistedInCurLoop = nullptr;
4803 LoopHoistContext(Compiler* comp)
4804 : m_pHoistedInCurLoop(nullptr)
4805 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4806 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4811 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4812 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4813 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4814 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4816 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4817 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4818 // "m_hoistedInParentLoops".
4820 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4822 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4823 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4824 // expressions to "hoistInLoop".
4825 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4827 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4828 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4830 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4831 // that are invariant in loop "lnum" (an index into the optLoopTable)
4832 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4833 // expressions to "hoistInLoop".
4834 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4835 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4836 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4837 bool optHoistLoopExprsForTree(GenTreePtr tree,
4839 LoopHoistContext* hoistCtxt,
4840 bool* firstBlockAndBeforeSideEffect,
4843 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4844 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4846 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4847 // Constants and init values are always loop invariant.
4848 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4849 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4851 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4852 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4853 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4854 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4855 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4857 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4858 // in the loop table.
4859 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4861 // Records the set of "side effects" of all loops: fields (object instance and static)
4862 // written to, and SZ-array element type equivalence classes updated.
4863 void optComputeLoopSideEffects();
4866 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4867 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4868 // static) written to, and SZ-array element type equivalence classes updated.
4869 void optComputeLoopNestSideEffects(unsigned lnum);
4871 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4872 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4874 // Hoist the expression "expr" out of loop "lnum".
4875 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4878 void optOptimizeBools();
4881 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4883 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4886 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4888 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4889 // the loop into a "do-while" loop
4890 // Also finds all natural loops and records them in the loop table
4892 // Optionally clone loops in the loop table.
4893 void optCloneLoops();
4895 // Clone loop "loopInd" in the loop table.
4896 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4898 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4899 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4900 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4902 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4904 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4907 // This enumeration describes what is killed by a call.
4911 CALLINT_NONE, // no interference (most helpers)
4912 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4913 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4914 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4915 CALLINT_ALL, // kills everything (normal method call)
4919 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4920 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4921 // in bbNext order; we use comparisons on the bbNum to decide order.)
4922 // The blocks that define the body are
4923 // first <= top <= entry <= bottom .
4924 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4925 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4926 // Compiler::optFindNaturalLoops().
4929 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4930 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4931 // loop, but not the outer loop.)
4932 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4934 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4935 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4936 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4938 callInterf lpAsgCall; // "callInterf" for calls in the loop
4939 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4940 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4942 unsigned short lpFlags; // Mask of the LPFLG_* constants
4944 unsigned char lpExitCnt; // number of exits from the loop
4946 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4947 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4948 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4949 // (Actually, an "immediately" nested loop --
4950 // no other child of this loop is a parent of lpChild.)
4951 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4952 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4953 // by following "lpChild" then "lpSibling" links.
4955 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4956 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4958 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4959 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4960 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4962 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4963 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4965 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4966 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4967 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4968 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4970 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
4971 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
4972 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
4974 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
4975 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
4976 // type are assigned to.
4978 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
4979 // If this is set, the fields below may not be accurate (since they become irrelevant.)
4980 bool lpContainsCall; // True if executing the loop body *may* execute a call
4982 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
4983 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
4985 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
4987 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
4988 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
4990 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
4992 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
4993 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
4995 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
4996 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
4998 JitSimplerHashBehavior>
5000 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5001 // instance fields modified
5004 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5005 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5007 JitSimplerHashBehavior>
5009 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5010 // arrays of that type are modified
5013 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5014 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5016 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5017 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5018 // (shifted left, with a low-order bit set to distinguish.)
5019 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5020 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5022 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5024 GenTreePtr lpIterTree; // The "i <op>= const" tree
5025 unsigned lpIterVar(); // iterator variable #
5026 int lpIterConst(); // the constant with which the iterator is incremented
5027 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5028 void VERIFY_lpIterTree();
5030 var_types lpIterOperType(); // For overflow instructions
5033 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5034 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5038 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5040 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5041 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5042 void VERIFY_lpTestTree();
5044 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5045 GenTreePtr lpIterator(); // the iterator node in the loop test
5046 GenTreePtr lpLimit(); // the limit node in the loop test
5048 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5049 // LPFLG_CONST_LIMIT
5050 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5052 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5053 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5054 // LPFLG_ARRLEN_LIMIT
5056 // Returns "true" iff "*this" contains the blk.
5057 bool lpContains(BasicBlock* blk)
5059 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5061 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5062 // to be equal, but requiring bottoms to be different.)
5063 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5065 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5068 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5069 // bottoms to be different.)
5070 bool lpContains(const LoopDsc& lp2)
5072 return lpContains(lp2.lpFirst, lp2.lpBottom);
5075 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5076 // (allowing firsts to be equal, but requiring bottoms to be different.)
5077 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5079 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5082 // Returns "true" iff "*this" is (properly) contained by "lp2"
5083 // (allowing firsts to be equal, but requiring bottoms to be different.)
5084 bool lpContainedBy(const LoopDsc& lp2)
5086 return lpContains(lp2.lpFirst, lp2.lpBottom);
5089 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5090 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5092 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5094 // Returns "true" iff "*this" is disjoint from "lp2".
5095 bool lpDisjoint(const LoopDsc& lp2)
5097 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5099 // Returns "true" iff the loop is well-formed (see code for defn).
5102 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5103 lpEntry->bbNum <= lpBottom->bbNum &&
5104 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5109 bool fgMightHaveLoop(); // returns true if there are any backedges
5110 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5113 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5114 unsigned char optLoopCount; // number of tracked loops
5117 unsigned optCallCount; // number of calls made in the method
5118 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5119 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5120 unsigned optLoopsCloned; // number of loops cloned in the current method.
5123 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5124 void optPrintLoopInfo(unsigned loopNum,
5126 BasicBlock* lpFirst,
5128 BasicBlock* lpEntry,
5129 BasicBlock* lpBottom,
5130 unsigned char lpExitCnt,
5132 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5133 void optPrintLoopInfo(unsigned lnum);
5134 void optPrintLoopRecording(unsigned lnum);
5136 void optCheckPreds();
5139 void optSetBlockWeights();
5141 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5143 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5145 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5147 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5148 unsigned optIsLoopIncrTree(GenTreePtr incr);
5149 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5150 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5151 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5152 bool optExtractInitTestIncr(BasicBlock* head,
5157 GenTreePtr* ppIncr);
5159 void optRecordLoop(BasicBlock* head,
5165 unsigned char exitCnt);
5167 void optFindNaturalLoops();
5169 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5170 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5171 bool optCanonicalizeLoopNest(unsigned char loopInd);
5173 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5174 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5175 bool optCanonicalizeLoop(unsigned char loopInd);
5177 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5178 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5179 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5180 bool optLoopContains(unsigned l1, unsigned l2);
5182 // Requires "loopInd" to be a valid index into the loop table.
5183 // Updates the loop table by changing loop "loopInd", whose head is required
5184 // to be "from", to be "to". Also performs this transformation for any
5185 // loop nested in "loopInd" that shares the same head as "loopInd".
5186 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5188 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5189 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5190 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5192 // Marks the containsCall information to "lnum" and any parent loops.
5193 void AddContainsCallAllContainingLoops(unsigned lnum);
5194 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5195 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5196 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5197 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5198 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5199 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5201 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5202 // of "from".) Copies the jump destination from "from" to "to".
5203 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5205 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5206 unsigned optLoopDepth(unsigned lnum)
5208 unsigned par = optLoopTable[lnum].lpParent;
5209 if (par == BasicBlock::NOT_IN_LOOP)
5215 return 1 + optLoopDepth(par);
5219 void fgOptWhileLoop(BasicBlock* block);
5221 bool optComputeLoopRep(int constInit,
5224 genTreeOps iterOper,
5226 genTreeOps testOper,
5229 unsigned* iterCount);
5230 #if FEATURE_STACK_FP_X87
5233 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5234 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5235 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5236 #endif // FEATURE_STACK_FP_X87
5239 static fgWalkPreFn optIsVarAssgCB;
5242 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5244 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5246 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5248 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5250 /**************************************************************************
5251 * Optimization conditions
5252 *************************************************************************/
5254 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5255 bool optPentium4(void);
5256 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5257 bool optAvoidIntMult(void);
5262 // The following is the upper limit on how many expressions we'll keep track
5263 // of for the CSE analysis.
5265 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5267 static const int MIN_CSE_COST = 2;
5269 // Keeps tracked cse indices
5270 BitVecTraits* cseTraits;
5274 /* Generic list of nodes - used by the CSE logic */
5282 typedef struct treeLst* treeLstPtr;
5286 treeStmtLst* tslNext;
5287 GenTreePtr tslTree; // tree node
5288 GenTreePtr tslStmt; // statement containing the tree
5289 BasicBlock* tslBlock; // block containing the statement
5292 typedef struct treeStmtLst* treeStmtLstPtr;
5294 // The following logic keeps track of expressions via a simple hash table.
5298 CSEdsc* csdNextInBucket; // used by the hash table
5300 unsigned csdHashValue; // the orginal hashkey
5302 unsigned csdIndex; // 1..optCSECandidateCount
5303 char csdLiveAcrossCall; // 0 or 1
5305 unsigned short csdDefCount; // definition count
5306 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5308 unsigned csdDefWtCnt; // weighted def count
5309 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5311 GenTreePtr csdTree; // treenode containing the 1st occurance
5312 GenTreePtr csdStmt; // stmt containing the 1st occurance
5313 BasicBlock* csdBlock; // block containing the 1st occurance
5315 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5316 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5319 static const size_t s_optCSEhashSize;
5320 CSEdsc** optCSEhash;
5325 CSEdsc* optCSEfindDsc(unsigned index);
5326 void optUnmarkCSE(GenTreePtr tree);
5328 // user defined callback data for the tree walk function optCSE_MaskHelper()
5329 struct optCSE_MaskData
5331 EXPSET_TP CSE_defMask;
5332 EXPSET_TP CSE_useMask;
5335 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5336 static fgWalkPreFn optCSE_MaskHelper;
5338 // This function walks all the node for an given tree
5339 // and return the mask of CSE definitions and uses for the tree
5341 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5343 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5344 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5345 bool optCSE_canSwap(GenTree* tree);
5347 static fgWalkPostFn optPropagateNonCSE;
5348 static fgWalkPreFn optHasNonCSEChild;
5350 static fgWalkPreFn optUnmarkCSEs;
5352 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5353 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5355 void optCleanupCSEs();
5358 void optEnsureClearCSEInfo();
5361 #endif // FEATURE_ANYCSE
5363 #if FEATURE_VALNUM_CSE
5364 /**************************************************************************
5365 * Value Number based CSEs
5366 *************************************************************************/
5369 void optOptimizeValnumCSEs();
5372 void optValnumCSE_Init();
5373 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5374 unsigned optValnumCSE_Locate();
5375 void optValnumCSE_InitDataFlow();
5376 void optValnumCSE_DataFlow();
5377 void optValnumCSE_Availablity();
5378 void optValnumCSE_Heuristic();
5379 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5381 #endif // FEATURE_VALNUM_CSE
5384 bool optDoCSE; // True when we have found a duplicate CSE tree
5385 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5386 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5387 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5388 unsigned optCSEstart; // The first local variable number that is a CSE
5389 unsigned optCSEcount; // The total count of CSE's introduced.
5390 unsigned optCSEweight; // The weight of the current block when we are
5391 // scanning for CSE expressions
5393 bool optIsCSEcandidate(GenTreePtr tree);
5395 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5397 bool lclNumIsTrueCSE(unsigned lclNum) const
5399 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5402 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5404 bool lclNumIsCSE(unsigned lclNum) const
5406 return lvaTable[lclNum].lvIsCSE;
5410 bool optConfigDisableCSE();
5411 bool optConfigDisableCSE2();
5413 void optOptimizeCSEs();
5415 #endif // FEATURE_ANYCSE
5423 unsigned ivaVar; // Variable we are interested in, or -1
5424 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5425 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5426 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5427 callInterf ivaMaskCall; // What kind of calls are there?
5430 static callInterf optCallInterf(GenTreePtr call);
5433 // VN based copy propagation.
5434 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5435 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5436 LclNumToGenTreePtrStack;
5438 // Kill set to track variables with intervening definitions.
5439 VARSET_TP optCopyPropKillSet;
5441 // Copy propagation functions.
5442 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5443 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5444 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5445 bool optIsSsaLocal(GenTreePtr tree);
5446 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5447 void optVnCopyProp();
5449 /**************************************************************************
5450 * Early value propagation
5451 *************************************************************************/
5457 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5461 static unsigned GetHashCode(SSAName ssaNm)
5463 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5466 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5468 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5472 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5473 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5474 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5475 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5476 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5478 unsigned optMethodFlags;
5480 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5481 // No throughput diff was found with backward walk bound between 3-8.
5482 static const int optEarlyPropRecurBound = 5;
5484 enum class optPropKind
5492 bool gtIsVtableRef(GenTreePtr tree);
5493 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5494 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5495 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5496 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5497 bool optEarlyPropRewriteTree(GenTreePtr tree);
5498 bool optDoEarlyPropForBlock(BasicBlock* block);
5499 bool optDoEarlyPropForFunc();
5500 void optEarlyProp();
5501 void optFoldNullCheck(GenTreePtr tree);
5502 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5505 /**************************************************************************
5506 * Value/Assertion propagation
5507 *************************************************************************/
5509 // Data structures for assertion prop
5510 BitVecTraits* apTraits;
5514 enum optAssertionKind
5529 O1K_ARRLEN_OPER_BND,
5530 O1K_ARRLEN_LOOP_BND,
5531 O1K_CONSTANT_LOOP_BND,
5552 optAssertionKind assertionKind;
5555 unsigned lclNum; // assigned to or property of this local var number
5563 struct AssertionDscOp1
5565 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5572 struct AssertionDscOp2
5574 optOp2Kind kind; // a const or copy assignment
5578 ssize_t iconVal; // integer
5579 unsigned iconFlags; // gtFlags
5581 struct Range // integer subrange
5595 bool IsArrLenArithBound()
5597 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5599 bool IsArrLenBound()
5601 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5603 bool IsConstantBound()
5605 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5606 op1.kind == O1K_CONSTANT_LOOP_BND);
5608 bool IsBoundsCheckNoThrow()
5610 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5613 bool IsCopyAssertion()
5615 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5618 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5620 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5621 a1->op2.kind == a2->op2.kind;
5624 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5626 if (kind == OAK_EQUAL)
5628 return kind2 == OAK_NOT_EQUAL;
5630 else if (kind == OAK_NOT_EQUAL)
5632 return kind2 == OAK_EQUAL;
5637 static ssize_t GetLowerBoundForIntegralType(var_types type)
5657 static ssize_t GetUpperBoundForIntegralType(var_types type)
5681 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5683 return (op1.kind == that->op1.kind) &&
5684 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5687 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5689 if (op2.kind != that->op2.kind)
5695 case O2K_IND_CNS_INT:
5697 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5699 case O2K_CONST_LONG:
5700 return (op2.lconVal == that->op2.lconVal);
5702 case O2K_CONST_DOUBLE:
5703 // exact match because of positive and negative zero.
5704 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5706 case O2K_LCLVAR_COPY:
5708 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5709 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5712 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5715 // we will return false
5719 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5725 bool Complementary(AssertionDsc* that, bool vnBased)
5727 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5728 HasSameOp2(that, vnBased);
5731 bool Equals(AssertionDsc* that, bool vnBased)
5733 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5737 typedef unsigned short AssertionIndex;
5740 static fgWalkPreFn optAddCopiesCallback;
5741 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5742 unsigned optAddCopyLclNum;
5743 GenTreePtr optAddCopyAsgnNode;
5745 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5746 bool optAssertionPropagated; // set to true if we modified the trees
5747 bool optAssertionPropagatedCurrentStmt;
5749 GenTreePtr optAssertionPropCurrentTree;
5751 AssertionIndex* optComplementaryAssertionMap;
5752 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5753 // using the value of a local var) for each local var
5754 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5755 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5756 AssertionIndex optMaxAssertionCount;
5759 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5760 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5761 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5762 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5763 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5764 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5766 AssertionIndex GetAssertionCount()
5768 return optAssertionCount;
5770 ASSERT_TP* bbJtrueAssertionOut;
5771 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5772 ValueNumToAssertsMap;
5773 ValueNumToAssertsMap* optValueNumToAsserts;
5775 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5777 // Assertion prop helpers.
5778 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5779 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5780 void optAssertionInit(bool isLocalProp);
5781 void optAssertionTraitsInit(AssertionIndex assertionCount);
5782 #if LOCAL_ASSERTION_PROP
5783 void optAssertionReset(AssertionIndex limit);
5784 void optAssertionRemove(AssertionIndex index);
5787 // Assertion prop data flow functions.
5788 void optAssertionPropMain();
5789 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5790 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5791 ASSERT_TP* optInitAssertionDataflowFlags();
5792 ASSERT_TP* optComputeAssertionGen();
5794 // Assertion Gen functions.
5795 void optAssertionGen(GenTreePtr tree);
5796 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5797 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5798 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5799 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5800 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5801 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5803 // Assertion creation functions.
5804 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5805 AssertionIndex optCreateAssertion(GenTreePtr op1,
5807 optAssertionKind assertionKind,
5808 AssertionDsc* assertion);
5809 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5811 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5812 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5813 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5815 void optPrintVnAssertionMapping();
5817 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5819 // Used for respective assertion propagations.
5820 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5821 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5822 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5823 bool optAssertionIsNonNull(GenTreePtr op,
5824 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5826 // Used for Relop propagation.
5827 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5828 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5829 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5831 // Assertion prop for lcl var functions.
5832 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5833 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5835 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5836 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5837 const GenTreePtr tree,
5838 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5839 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5841 // Assertion propagation functions.
5842 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5843 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5844 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5845 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5846 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5847 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5848 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5849 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5850 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5851 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5852 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5853 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5855 // Implied assertion functions.
5856 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5857 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5858 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5859 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5861 ASSERT_VALRET_TP optNewFullAssertSet();
5862 ASSERT_VALRET_TP optNewEmptyAssertSet();
5865 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5866 void optDebugCheckAssertion(AssertionDsc* assertion);
5867 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5869 void optAddCopies();
5870 #endif // ASSERTION_PROP
5872 /**************************************************************************
5874 *************************************************************************/
5877 struct LoopCloneVisitorInfo
5879 LoopCloneContext* context;
5882 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5883 : context(context), loopNum(loopNum), stmt(nullptr)
5888 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5889 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5890 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5891 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5892 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5893 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5894 void optObtainLoopCloningOpts(LoopCloneContext* context);
5895 bool optIsLoopClonable(unsigned loopInd);
5897 bool optCanCloneLoops();
5900 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5902 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5903 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5904 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5905 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5909 void optInsertLoopCloningStress(BasicBlock* head);
5911 #if COUNT_RANGECHECKS
5912 static unsigned optRangeChkRmv;
5913 static unsigned optRangeChkAll;
5922 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5927 RngChkDsc* rcdNextInBucket; // used by the hash table
5929 unsigned short rcdHashValue; // to make matching faster
5930 unsigned short rcdIndex; // 0..optRngChkCount-1
5932 GenTreePtr rcdTree; // the array index tree
5935 unsigned optRngChkCount;
5936 static const size_t optRngChkHashSize;
5938 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5939 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5942 bool optIsNoMore(GenTreePtr op1, GenTreePtr op2, int add1 = 0, int add2 = 0);
5945 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5948 bool optLoopsMarked;
5951 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5952 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5956 XX Does the register allocation and puts the remaining lclVars on the stack XX
5958 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5959 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5963 #ifndef LEGACY_BACKEND
5968 #else // LEGACY_BACKEND
5973 #endif // LEGACY_BACKEND
5975 #ifdef LEGACY_BACKEND
5977 void raAssignVars(); // register allocation
5978 #endif // LEGACY_BACKEND
5980 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
5982 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
5984 void raMarkStkVars();
5987 // Some things are used by both LSRA and regpredict allocators.
5989 FrameType rpFrameType;
5990 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
5992 #ifdef LEGACY_BACKEND
5993 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
5995 #endif // LEGACY_BACKEND
5997 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
5999 #if FEATURE_FP_REGALLOC
6000 enum enumConfigRegisterFP
6002 CONFIG_REGISTER_FP_NONE = 0x0,
6003 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6004 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6005 CONFIG_REGISTER_FP_FULL = 0x3,
6007 enumConfigRegisterFP raConfigRegisterFP();
6008 #endif // FEATURE_FP_REGALLOC
6011 regMaskTP raConfigRestrictMaskFP();
6014 #ifndef LEGACY_BACKEND
6015 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6016 #else // LEGACY_BACKEND
6017 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6018 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6019 bool raNewBlocks; // True is we added killing blocks for FPU registers
6020 unsigned rpPasses; // Number of passes made by the register predicter
6021 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6022 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6023 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6024 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6025 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6026 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6027 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6028 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6029 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6030 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6031 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6032 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6034 bool rpRegAllocDone; // Set to true after we have completed register allocation
6036 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6038 void raSetupArgMasks(RegState* r);
6040 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6042 void raDumpVarIntf(); // Dump the variable to variable interference graph
6043 void raDumpRegIntf(); // Dump the variable to register interference graph
6045 void raAdjustVarIntf();
6047 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6049 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6051 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6052 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6054 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6056 static fgWalkPreFn rpMarkRegIntf;
6058 regMaskTP rpPredictAddressMode(
6059 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6061 void rpPredictRefAssign(unsigned lclNum);
6063 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6065 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6067 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6069 void rpPredictRegUse(); // Entry point
6071 unsigned raPredictTreeRegUse(GenTreePtr tree);
6072 unsigned raPredictListRegUse(GenTreePtr list);
6074 void raSetRegVarOrder(var_types regType,
6075 regNumber* customVarOrder,
6076 unsigned* customVarOrderSize,
6078 regMaskTP avoidReg);
6080 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6081 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6082 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6083 void raAddToStkPredict(unsigned val)
6085 unsigned newStkPredict = rpStkPredict + val;
6086 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6087 rpStkPredict = UINT_MAX - 1;
6089 rpStkPredict = newStkPredict;
6093 #if !FEATURE_FP_REGALLOC
6094 void raDispFPlifeInfo();
6098 regMaskTP genReturnRegForTree(GenTreePtr tree);
6099 #endif // LEGACY_BACKEND
6101 /* raIsVarargsStackArg is called by raMaskStkVars and by
6102 lvaSortByRefCount. It identifies the special case
6103 where a varargs function has a parameter passed on the
6104 stack, other than the special varargs handle. Such parameters
6105 require special treatment, because they cannot be tracked
6106 by the GC (their offsets in the stack are not known
6110 bool raIsVarargsStackArg(unsigned lclNum)
6114 LclVarDsc* varDsc = &lvaTable[lclNum];
6116 assert(varDsc->lvIsParam);
6118 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6120 #else // _TARGET_X86_
6124 #endif // _TARGET_X86_
6127 #ifdef LEGACY_BACKEND
6128 // Records the current prediction, if it's better than any previous recorded prediction.
6129 void rpRecordPrediction();
6130 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6131 void rpUseRecordedPredictionIfBetter();
6133 // Data members used in the methods above.
6134 unsigned rpBestRecordedStkPredict;
6135 struct VarRegPrediction
6137 bool m_isEnregistered;
6138 regNumberSmall m_regNum;
6139 regNumberSmall m_otherReg;
6141 VarRegPrediction* rpBestRecordedPrediction;
6142 #endif // LEGACY_BACKEND
6145 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6146 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6150 XX Get to the class and method info from the Execution Engine given XX
6151 XX tokens for the class and method XX
6153 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6154 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6158 /* These are the different addressing modes used to access a local var.
6159 * The JIT has to report the location of the locals back to the EE
6160 * for debugging purposes.
6166 VLT_REG_BYREF, // this type is currently only used for value types on X64
6169 VLT_STK_BYREF, // this type is currently only used for value types on X64
6183 siVarLocType vlType;
6186 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6188 // VLT_REG_BYREF -- the specified register contains the address of the variable
6196 // VLT_STK -- Any 32 bit value which is on the stack
6197 // eg. [ESP+0x20], or [EBP-0x28]
6198 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6199 // eg. mov EAX, [ESP+0x20]; [EAX]
6203 regNumber vlsBaseReg;
6204 NATIVE_OFFSET vlsOffset;
6207 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6216 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6217 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6225 regNumber vlrssBaseReg;
6226 NATIVE_OFFSET vlrssOffset;
6230 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6231 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6237 regNumber vlsrsBaseReg;
6238 NATIVE_OFFSET vlsrsOffset;
6244 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6245 // eg 2 DWords at [ESP+0x10]
6249 regNumber vls2BaseReg;
6250 NATIVE_OFFSET vls2Offset;
6253 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6254 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6261 // VLT_FIXED_VA -- fixed argument of a varargs function.
6262 // The argument location depends on the size of the variable
6263 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6264 // location of the first arg. This argument can then be accessed
6265 // relative to the position of the first arg
6269 unsigned vlfvOffset;
6276 void* rpValue; // pointer to the in-process
6277 // location of the value.
6283 bool vlIsInReg(regNumber reg);
6284 bool vlIsOnStk(regNumber reg, signed offset);
6287 /*************************************************************************/
6292 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6293 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6294 CORINFO_CALLINFO_FLAGS flags,
6295 CORINFO_CALL_INFO* pResult);
6296 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6298 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6299 CORINFO_ACCESS_FLAGS flags,
6300 CORINFO_FIELD_INFO* pResult);
6304 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6306 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6308 bool IsSuperPMIException(unsigned code)
6310 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6312 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6313 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6314 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6315 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6316 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6317 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6318 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6319 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6323 case EXCEPTIONCODE_DebugBreakorAV:
6324 case EXCEPTIONCODE_MC:
6325 case EXCEPTIONCODE_LWM:
6326 case EXCEPTIONCODE_SASM:
6327 case EXCEPTIONCODE_SSYM:
6328 case EXCEPTIONCODE_CALLUTILS:
6329 case EXCEPTIONCODE_TYPEUTILS:
6330 case EXCEPTIONCODE_ASSERT:
6337 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6338 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6340 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6341 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6344 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6345 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6346 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6348 // VOM info, method sigs
6350 void eeGetSig(unsigned sigTok,
6351 CORINFO_MODULE_HANDLE scope,
6352 CORINFO_CONTEXT_HANDLE context,
6353 CORINFO_SIG_INFO* retSig);
6355 void eeGetCallSiteSig(unsigned sigTok,
6356 CORINFO_MODULE_HANDLE scope,
6357 CORINFO_CONTEXT_HANDLE context,
6358 CORINFO_SIG_INFO* retSig);
6360 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6362 // Method entry-points, instrs
6364 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6366 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6368 CORINFO_EE_INFO eeInfo;
6369 bool eeInfoInitialized;
6371 CORINFO_EE_INFO* eeGetEEInfo();
6373 // Gets the offset of a SDArray's first element
6374 unsigned eeGetArrayDataOffset(var_types type);
6375 // Gets the offset of a MDArray's first element
6376 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6378 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6380 // Returns the page size for the target machine as reported by the EE.
6381 inline size_t eeGetPageSize()
6383 #if COR_JIT_EE_VERSION > 460
6384 return eeGetEEInfo()->osPageSize;
6385 #else // COR_JIT_EE_VERSION <= 460
6386 return CORINFO_PAGE_SIZE;
6387 #endif // COR_JIT_EE_VERSION > 460
6390 // Returns the frame size at which we will generate a loop to probe the stack.
6391 inline size_t getVeryLargeFrameSize()
6394 // The looping probe code is 40 bytes, whereas the straight-line probing for
6395 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6396 // or greater, to generate smaller code.
6397 return 2 * eeGetPageSize();
6399 return 3 * eeGetPageSize();
6403 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6405 #if COR_JIT_EE_VERSION > 460
6406 return eeGetEEInfo()->targetAbi == abi;
6408 return CORINFO_DESKTOP_ABI == abi;
6412 inline bool generateCFIUnwindCodes()
6414 #ifdef UNIX_AMD64_ABI
6415 return IsTargetAbi(CORINFO_CORERT_ABI);
6423 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6425 // Debugging support - Line number info
6427 void eeGetStmtOffsets();
6429 unsigned eeBoundariesCount;
6431 struct boundariesDsc
6433 UNATIVE_OFFSET nativeIP;
6435 unsigned sourceReason;
6436 } * eeBoundaries; // Boundaries to report to EE
6437 void eeSetLIcount(unsigned count);
6438 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6442 static void eeDispILOffs(IL_OFFSET offs);
6443 static void eeDispLineInfo(const boundariesDsc* line);
6444 void eeDispLineInfos();
6447 // Debugging support - Local var info
6451 unsigned eeVarsCount;
6453 struct VarResultInfo
6455 UNATIVE_OFFSET startOffset;
6456 UNATIVE_OFFSET endOffset;
6460 void eeSetLVcount(unsigned count);
6461 void eeSetLVinfo(unsigned which,
6462 UNATIVE_OFFSET startOffs,
6463 UNATIVE_OFFSET length,
6468 const siVarLoc& loc);
6472 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6473 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6476 // ICorJitInfo wrappers
6478 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6480 void eeAllocUnwindInfo(BYTE* pHotCode,
6486 CorJitFuncKind funcKind);
6488 void eeSetEHcount(unsigned cEH);
6490 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6492 WORD eeGetRelocTypeHint(void* target);
6494 // ICorStaticInfo wrapper functions
6496 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6498 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6500 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6503 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6504 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6505 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6506 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6508 template <typename ParamType>
6509 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6511 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6514 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6516 // Utility functions
6518 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6521 const wchar_t* eeGetCPString(size_t stringHandle);
6524 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6526 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6527 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6529 static fgWalkPreFn CountSharedStaticHelper;
6530 static bool IsSharedStaticHelper(GenTreePtr tree);
6531 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6533 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6534 // returns true/false if 'field' is a Jit Data offset
6535 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6536 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6537 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6539 /*****************************************************************************/
6544 enum TEMP_USAGE_TYPE
6550 static var_types tmpNormalizeType(var_types type);
6551 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6552 void tmpRlsTemp(TempDsc* temp);
6553 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6556 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6557 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6561 bool tmpAllFree() const;
6564 #ifndef LEGACY_BACKEND
6565 void tmpPreAllocateTemps(var_types type, unsigned count);
6566 #endif // !LEGACY_BACKEND
6569 #ifdef LEGACY_BACKEND
6570 unsigned tmpIntSpillMax; // number of int-sized spill temps
6571 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6572 #endif // LEGACY_BACKEND
6574 unsigned tmpCount; // Number of temps
6575 unsigned tmpSize; // Size of all the temps
6578 // Used by RegSet::rsSpillChk()
6579 unsigned tmpGetCount; // Temps which haven't been released yet
6582 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6584 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6585 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6588 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6589 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6593 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6594 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6598 CodeGenInterface* codeGen;
6600 // The following holds information about instr offsets in terms of generated code.
6604 IPmappingDsc* ipmdNext; // next line# record
6605 IL_OFFSETX ipmdILoffsx; // the instr offset
6606 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6607 bool ipmdIsLabel; // Can this code be a branch label?
6610 // Record the instr offset mapping to the generated code
6612 IPmappingDsc* genIPmappingList;
6613 IPmappingDsc* genIPmappingLast;
6615 // Managed RetVal - A side hash table meant to record the mapping from a
6616 // GT_CALL node to its IL offset. This info is used to emit sequence points
6617 // that can be used by debugger to determine the native offset at which the
6618 // managed RetVal will be available.
6620 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6621 // favor of a side table for two reasons: 1) We need IL offset for only those
6622 // GT_CALL nodes (created during importation) that correspond to an IL call and
6623 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6624 // structure and IL offset is needed only when generating debuggable code. Therefore
6625 // it is desirable to avoid memory size penalty in retail scenarios.
6626 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6627 CallSiteILOffsetTable;
6628 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6630 unsigned genReturnLocal; // Local number for the return value when applicable.
6631 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6633 // The following properties are part of CodeGenContext. Getters are provided here for
6634 // convenience and backward compatibility, but the properties can only be set by invoking
6635 // the setter on CodeGenContext directly.
6637 __declspec(property(get = getEmitter)) emitter* genEmitter;
6638 emitter* getEmitter()
6640 return codeGen->getEmitter();
6643 const bool isFramePointerUsed()
6645 return codeGen->isFramePointerUsed();
6648 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6649 bool getInterruptible()
6651 return codeGen->genInterruptible;
6653 void setInterruptible(bool value)
6655 codeGen->setInterruptible(value);
6659 const bool genDoubleAlign()
6661 return codeGen->doDoubleAlign();
6663 DWORD getCanDoubleAlign();
6664 bool shouldDoubleAlign(unsigned refCntStk,
6666 unsigned refCntWtdReg,
6667 unsigned refCntStkParam,
6668 unsigned refCntWtdStkDbl);
6669 #endif // DOUBLE_ALIGN
6671 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6672 bool getFullPtrRegMap()
6674 return codeGen->genFullPtrRegMap;
6676 void setFullPtrRegMap(bool value)
6678 codeGen->setFullPtrRegMap(value);
6681 // Things that MAY belong either in CodeGen or CodeGenContext
6683 #if FEATURE_EH_FUNCLETS
6684 FuncInfoDsc* compFuncInfos;
6685 unsigned short compCurrFuncIdx;
6686 unsigned short compFuncInfoCount;
6688 unsigned short compFuncCount()
6690 assert(fgFuncletsCreated);
6691 return compFuncInfoCount;
6694 #else // !FEATURE_EH_FUNCLETS
6696 // This is a no-op when there are no funclets!
6697 void genUpdateCurrentFunclet(BasicBlock* block)
6702 FuncInfoDsc compFuncInfoRoot;
6704 static const unsigned compCurrFuncIdx = 0;
6706 unsigned short compFuncCount()
6711 #endif // !FEATURE_EH_FUNCLETS
6713 FuncInfoDsc* funCurrentFunc();
6714 void funSetCurrentFunc(unsigned funcIdx);
6715 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6716 unsigned int funGetFuncIdx(BasicBlock* block);
6720 VARSET_TP compCurLife; // current live variables
6721 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6723 template <bool ForCodeGen>
6724 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6726 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6728 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6731 template <bool ForCodeGen>
6732 void compUpdateLife(GenTreePtr tree);
6734 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6735 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6736 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6737 template <bool ForCodeGen>
6738 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6740 template <bool ForCodeGen>
6741 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6743 // Gets a register mask that represent the kill set for a helper call since
6744 // not all JIT Helper calls follow the standard ABI on the target architecture.
6745 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6747 // Gets a register mask that represent the kill set for a NoGC helper call.
6748 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6751 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6752 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6753 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6754 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6755 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6756 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6757 #endif // _TARGET_ARM_
6759 // 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
6761 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6763 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6764 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6765 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6766 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6767 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6768 // for the tracked var indices of the field vars, as in a live var set).
6769 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6771 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6773 if (m_promotedStructDeathVars == nullptr)
6775 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6777 return m_promotedStructDeathVars;
6781 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6782 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6786 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6787 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6790 #if !defined(__GNUC__)
6791 #pragma region Unwind information
6796 // Infrastructure functions: start/stop/reserve/emit.
6799 void unwindBegProlog();
6800 void unwindEndProlog();
6801 void unwindBegEpilog();
6802 void unwindEndEpilog();
6803 void unwindReserve();
6804 void unwindEmit(void* pHotCode, void* pColdCode);
6807 // Specific unwind information functions: called by code generation to indicate a particular
6808 // prolog or epilog unwindable instruction has been generated.
6811 void unwindPush(regNumber reg);
6812 void unwindAllocStack(unsigned size);
6813 void unwindSetFrameReg(regNumber reg, unsigned offset);
6814 void unwindSaveReg(regNumber reg, unsigned offset);
6816 #if defined(_TARGET_ARM_)
6817 void unwindPushMaskInt(regMaskTP mask);
6818 void unwindPushMaskFloat(regMaskTP mask);
6819 void unwindPopMaskInt(regMaskTP mask);
6820 void unwindPopMaskFloat(regMaskTP mask);
6821 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6822 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6823 // called via unwindPadding().
6824 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6825 // instruction and the current location.
6826 #endif // _TARGET_ARM_
6828 #if defined(_TARGET_ARM64_)
6830 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6831 // instruction and the current location.
6832 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6833 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6834 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6835 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6836 void unwindSaveNext(); // unwind code: save_next
6837 void unwindReturn(regNumber reg); // ret lr
6838 #endif // defined(_TARGET_ARM64_)
6841 // Private "helper" functions for the unwind implementation.
6845 #if FEATURE_EH_FUNCLETS
6846 void unwindGetFuncLocations(FuncInfoDsc* func,
6847 bool getHotSectionData,
6848 /* OUT */ emitLocation** ppStartLoc,
6849 /* OUT */ emitLocation** ppEndLoc);
6850 #endif // FEATURE_EH_FUNCLETS
6852 void unwindReserveFunc(FuncInfoDsc* func);
6853 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6855 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6857 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6858 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6860 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6862 #if defined(_TARGET_AMD64_)
6864 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6866 void unwindBegPrologWindows();
6867 void unwindPushWindows(regNumber reg);
6868 void unwindAllocStackWindows(unsigned size);
6869 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6870 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6872 #ifdef UNIX_AMD64_ABI
6873 void unwindBegPrologCFI();
6874 void unwindPushCFI(regNumber reg);
6875 void unwindAllocStackCFI(unsigned size);
6876 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6877 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6878 int mapRegNumToDwarfReg(regNumber reg);
6879 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6880 #endif // UNIX_AMD64_ABI
6881 #elif defined(_TARGET_ARM_)
6883 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6884 void unwindPushPopMaskFloat(regMaskTP mask);
6885 void unwindSplit(FuncInfoDsc* func);
6887 #endif // _TARGET_ARM_
6889 #if !defined(__GNUC__)
6890 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6894 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6895 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6899 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6900 XX that contains the distinguished, well-known SIMD type definitions). XX
6902 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6903 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6906 // Get highest available instruction set for floating point codegen
6907 InstructionSet getFloatingPointInstructionSet()
6909 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6912 return InstructionSet_AVX;
6917 return InstructionSet_SSE3_4;
6921 assert(canUseSSE2());
6922 return InstructionSet_SSE2;
6924 assert(!"getFPInstructionSet() is not implemented for target arch");
6926 return InstructionSet_NONE;
6930 // Get highest available instruction set for SIMD codegen
6931 InstructionSet getSIMDInstructionSet()
6933 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6934 return getFloatingPointInstructionSet();
6936 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6938 return InstructionSet_NONE;
6944 // Should we support SIMD intrinsics?
6947 // Have we identified any SIMD types?
6948 // This is currently used by struct promotion to avoid getting type information for a struct
6949 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6951 bool _usesSIMDTypes;
6952 bool usesSIMDTypes()
6954 return _usesSIMDTypes;
6956 void setUsesSIMDTypes(bool value)
6958 _usesSIMDTypes = value;
6961 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6962 // that require indexed access to the individual fields of the vector, which is not well supported
6963 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6964 unsigned lvaSIMDInitTempVarNum;
6967 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6968 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6969 CORINFO_CLASS_HANDLE SIMDIntHandle;
6970 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6971 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6972 CORINFO_CLASS_HANDLE SIMDShortHandle;
6973 CORINFO_CLASS_HANDLE SIMDByteHandle;
6974 CORINFO_CLASS_HANDLE SIMDLongHandle;
6975 CORINFO_CLASS_HANDLE SIMDUIntHandle;
6976 CORINFO_CLASS_HANDLE SIMDULongHandle;
6977 CORINFO_CLASS_HANDLE SIMDVector2Handle;
6978 CORINFO_CLASS_HANDLE SIMDVector3Handle;
6979 CORINFO_CLASS_HANDLE SIMDVector4Handle;
6980 CORINFO_CLASS_HANDLE SIMDVectorHandle;
6982 // Get the handle for a SIMD type.
6983 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
6985 if (simdBaseType == TYP_FLOAT)
6990 return SIMDVector2Handle;
6992 return SIMDVector3Handle;
6994 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
6996 return SIMDVector4Handle;
7005 assert(simdType == getSIMDVectorType());
7006 switch (simdBaseType)
7009 return SIMDFloatHandle;
7011 return SIMDDoubleHandle;
7013 return SIMDIntHandle;
7015 return SIMDUShortHandle;
7017 return SIMDUShortHandle;
7019 return SIMDUByteHandle;
7021 return SIMDShortHandle;
7023 return SIMDByteHandle;
7025 return SIMDLongHandle;
7027 return SIMDUIntHandle;
7029 return SIMDULongHandle;
7031 assert(!"Didn't find a class handle for simdType");
7033 return NO_CLASS_HANDLE;
7037 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7038 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7039 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7041 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7042 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7043 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7044 bool isSIMDTypeLocal(GenTree* tree)
7046 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7049 // Returns true if the type of the tree is a byref of TYP_SIMD
7050 bool isAddrOfSIMDType(GenTree* tree)
7052 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7054 switch (tree->OperGet())
7057 return varTypeIsSIMD(tree->gtGetOp1());
7059 case GT_LCL_VAR_ADDR:
7060 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7063 return isSIMDTypeLocal(tree);
7070 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7072 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7073 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7074 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7077 // Returns base type of a TYP_SIMD local.
7078 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7079 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7081 if (isSIMDTypeLocal(tree))
7083 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7089 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7091 return info.compCompHnd->isInSIMDModule(clsHnd);
7094 bool isSIMDClass(typeInfo* pTypeInfo)
7096 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7099 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7100 // if it is not a SIMD type or is an unsupported base type.
7101 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7103 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7105 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7108 // Get SIMD Intrinsic info given the method handle.
7109 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7110 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7111 CORINFO_METHOD_HANDLE methodHnd,
7112 CORINFO_SIG_INFO* sig,
7115 var_types* baseType,
7116 unsigned* sizeBytes);
7118 // Pops and returns GenTree node from importers type stack.
7119 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7120 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7122 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7123 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7125 // Creates a GT_SIMD tree for Select operation
7126 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7128 unsigned simdVectorSize,
7133 // Creates a GT_SIMD tree for Min/Max operation
7134 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7135 CORINFO_CLASS_HANDLE typeHnd,
7137 unsigned simdVectorSize,
7141 // Transforms operands and returns the SIMD intrinsic to be applied on
7142 // transformed operands to obtain given relop result.
7143 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7144 CORINFO_CLASS_HANDLE typeHnd,
7145 unsigned simdVectorSize,
7146 var_types* baseType,
7150 // Creates a GT_SIMD tree for Abs intrinsic.
7151 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7153 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7154 // Transforms operands and returns the SIMD intrinsic to be applied on
7155 // transformed operands to obtain == comparison result.
7156 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7157 unsigned simdVectorSize,
7161 // Transforms operands and returns the SIMD intrinsic to be applied on
7162 // transformed operands to obtain > comparison result.
7163 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7164 unsigned simdVectorSize,
7168 // Transforms operands and returns the SIMD intrinsic to be applied on
7169 // transformed operands to obtain >= comparison result.
7170 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7171 unsigned simdVectorSize,
7175 // Transforms operands and returns the SIMD intrinsic to be applied on
7176 // transformed operands to obtain >= comparison result in case of int32
7177 // and small int base type vectors.
7178 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7179 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7180 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7182 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7183 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7184 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7185 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7186 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7188 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7189 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7190 GenTreePtr newobjThis,
7191 CORINFO_CLASS_HANDLE clsHnd,
7192 CORINFO_METHOD_HANDLE method,
7193 CORINFO_SIG_INFO* sig,
7196 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7198 // Whether SIMD vector occupies part of SIMD register.
7199 // SSE2: vector2f/3f are considered sub register SIMD types.
7200 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7201 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7203 unsigned sizeBytes = 0;
7204 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7205 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7208 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7210 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7213 // Get the type for the hardware SIMD vector.
7214 // This is the maximum SIMD type supported for this target.
7215 var_types getSIMDVectorType()
7217 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7224 assert(canUseSSE2());
7228 assert(!"getSIMDVectorType() unimplemented on target arch");
7233 // Get the size of the SIMD type in bytes
7234 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7236 unsigned sizeBytes = 0;
7237 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7241 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7242 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7244 // Get the the number of elements of basetype of SIMD vector given by its type handle
7245 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7247 // Get preferred alignment of SIMD type.
7248 int getSIMDTypeAlignment(var_types simdType);
7250 // Get the number of bytes in a SIMD Vector for the current compilation.
7251 unsigned getSIMDVectorRegisterByteLength()
7253 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7256 return YMM_REGSIZE_BYTES;
7260 assert(canUseSSE2());
7261 return XMM_REGSIZE_BYTES;
7264 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7269 // The minimum and maximum possible number of bytes in a SIMD vector.
7270 unsigned int maxSIMDStructBytes()
7272 return getSIMDVectorRegisterByteLength();
7274 unsigned int minSIMDStructBytes()
7276 return emitTypeSize(TYP_SIMD8);
7279 #ifdef FEATURE_AVX_SUPPORT
7280 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7281 static const unsigned maxPossibleSIMDStructBytes = 32;
7282 #else // !FEATURE_AVX_SUPPORT
7283 static const unsigned maxPossibleSIMDStructBytes = 16;
7284 #endif // !FEATURE_AVX_SUPPORT
7286 // Returns the codegen type for a given SIMD size.
7287 var_types getSIMDTypeForSize(unsigned size)
7289 var_types simdType = TYP_UNDEF;
7292 simdType = TYP_SIMD8;
7294 else if (size == 12)
7296 simdType = TYP_SIMD12;
7298 else if (size == 16)
7300 simdType = TYP_SIMD16;
7302 #ifdef FEATURE_AVX_SUPPORT
7303 else if (size == 32)
7305 simdType = TYP_SIMD32;
7307 #endif // FEATURE_AVX_SUPPORT
7310 noway_assert(!"Unexpected size for SIMD type");
7315 unsigned getSIMDInitTempVarNum()
7317 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7319 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7320 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7322 return lvaSIMDInitTempVarNum;
7325 #endif // FEATURE_SIMD
7328 //------------------------------------------------------------------------
7329 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7331 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7332 // candidate for enregistration.
7334 unsigned largestEnregisterableStructSize()
7337 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7338 if (vectorRegSize > TARGET_POINTER_SIZE)
7340 return vectorRegSize;
7343 #endif // FEATURE_SIMD
7345 return TARGET_POINTER_SIZE;
7350 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7351 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7352 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7354 // Is this var is of type simd struct?
7355 bool lclVarIsSIMDType(unsigned varNum)
7357 LclVarDsc* varDsc = lvaTable + varNum;
7358 return varDsc->lvIsSIMDType();
7361 // Is this Local node a SIMD local?
7362 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7364 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7367 // Returns true if the TYP_SIMD locals on stack are aligned at their
7368 // preferred byte boundary specified by getSIMDTypeAlignment().
7370 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7371 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7372 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7373 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7374 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7375 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7376 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7379 bool isSIMDTypeLocalAligned(unsigned varNum)
7381 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7382 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7385 int off = lvaFrameAddress(varNum, &ebpBased);
7386 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7387 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7388 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7391 #endif // FEATURE_SIMD
7396 // Whether SSE2 is available
7397 bool canUseSSE2() const
7399 #ifdef _TARGET_XARCH_
7400 return opts.compCanUseSSE2;
7406 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7407 bool CanUseSSE3_4() const
7409 #ifdef _TARGET_XARCH_
7410 return opts.compCanUseSSE3_4;
7416 bool canUseAVX() const
7418 #ifdef FEATURE_AVX_SUPPORT
7419 return opts.compCanUseAVX;
7426 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7427 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7431 XX Generic info about the compilation and the method being compiled. XX
7432 XX It is responsible for driving the other phases. XX
7433 XX It is also responsible for all the memory management. XX
7435 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7436 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7440 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7442 InlineResult* compInlineResult; // The result of importing the inlinee method.
7444 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7445 bool compJmpOpUsed; // Does the method do a JMP
7446 bool compLongUsed; // Does the method use TYP_LONG
7447 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7448 bool compTailCallUsed; // Does the method do a tailcall
7449 bool compLocallocUsed; // Does the method use localloc.
7450 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7451 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7452 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7454 // NOTE: These values are only reliable after
7455 // the importing is completely finished.
7457 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7458 // we can iterate over these efficiently.
7460 #if CPU_USES_BLOCK_MOVE
7461 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7465 // State information - which phases have completed?
7466 // These are kept together for easy discoverability
7468 bool bRangeAllowStress;
7469 bool compCodeGenDone;
7470 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7471 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7472 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7473 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7476 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7477 bool fgLocalVarLivenessChanged;
7479 bool compStackProbePrologDone;
7481 #ifndef LEGACY_BACKEND
7483 #endif // !LEGACY_BACKEND
7484 bool compRationalIRForm;
7486 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7488 bool compGeneratingProlog;
7489 bool compGeneratingEpilog;
7490 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7491 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7492 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7493 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7494 bool getNeedsGSSecurityCookie() const
7496 return compNeedsGSSecurityCookie;
7498 void setNeedsGSSecurityCookie()
7500 compNeedsGSSecurityCookie = true;
7503 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7504 // frame layout calculations, this is the level we are currently
7507 //---------------------------- JITing options -----------------------------
7520 JitFlags* jitFlags; // all flags passed from the EE
7521 unsigned compFlags; // method attributes
7523 codeOptimize compCodeOpt; // what type of code optimizations
7527 #ifdef _TARGET_XARCH_
7528 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7529 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7531 #ifdef FEATURE_AVX_SUPPORT
7532 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7533 #endif // FEATURE_AVX_SUPPORT
7534 #endif // _TARGET_XARCH_
7536 // optimize maximally and/or favor speed over size?
7538 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7539 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7540 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7541 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7542 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7544 // Maximun number of locals before turning off the inlining
7545 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7548 unsigned instrCount;
7549 unsigned lvRefCount;
7550 bool compMinOptsIsSet;
7552 bool compMinOptsIsUsed;
7554 inline bool MinOpts()
7556 assert(compMinOptsIsSet);
7557 compMinOptsIsUsed = true;
7560 inline bool IsMinOptsSet()
7562 return compMinOptsIsSet;
7565 inline bool MinOpts()
7569 inline bool IsMinOptsSet()
7571 return compMinOptsIsSet;
7574 inline void SetMinOpts(bool val)
7576 assert(!compMinOptsIsUsed);
7577 assert(!compMinOptsIsSet || (compMinOpts == val));
7579 compMinOptsIsSet = true;
7582 // true if the CLFLG_* for an optimization is set.
7583 inline bool OptEnabled(unsigned optFlag)
7585 return !!(compFlags & optFlag);
7588 #ifdef FEATURE_READYTORUN_COMPILER
7589 inline bool IsReadyToRun()
7591 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7594 inline bool IsReadyToRun()
7600 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7601 // PInvoke transitions inline (e.g. when targeting CoreRT).
7602 inline bool ShouldUsePInvokeHelpers()
7604 #if COR_JIT_EE_VERSION > 460
7605 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7611 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7613 inline bool IsReversePInvoke()
7615 #if COR_JIT_EE_VERSION > 460
7616 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7622 // true if we must generate code compatible with JIT32 quirks
7623 inline bool IsJit32Compat()
7625 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7626 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7632 // true if we must generate code compatible with Jit64 quirks
7633 inline bool IsJit64Compat()
7635 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7636 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7637 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7644 bool compScopeInfo; // Generate the LocalVar info ?
7645 bool compDbgCode; // Generate debugger-friendly code?
7646 bool compDbgInfo; // Gather debugging info?
7649 #ifdef PROFILING_SUPPORTED
7650 bool compNoPInvokeInlineCB;
7652 static const bool compNoPInvokeInlineCB;
7656 bool compGcChecks; // Check arguments and return values to ensure they are sane
7657 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7658 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7662 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7663 // to be allocated on the stack.
7664 // It will be set to true in the following cases:
7665 // 1. When the method being compiled has a declarative security
7666 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7667 // This is also the case when we inject a prolog and epilog in the method.
7669 // 2. When the method being compiled has imperative security (i.e. the method
7670 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7672 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7674 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7675 // which gets reported as a GC root to stackwalker.
7676 // (See also ICodeManager::GetAddrOfSecurityObject.)
7683 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7684 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7688 #ifdef UNIX_AMD64_ABI
7689 // This flag is indicating if there is a need to align the frame.
7690 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7691 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7692 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7693 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7694 // there are calls and making sure the frame alignment logic is executed.
7695 bool compNeedToAlignFrame;
7696 #endif // UNIX_AMD64_ABI
7698 bool compProcedureSplitting; // Separate cold code from hot code
7700 bool genFPorder; // Preserve FP order (operations are non-commutative)
7701 bool genFPopt; // Can we do frame-pointer-omission optimization?
7702 bool altJit; // True if we are an altjit and are compiling this method
7705 bool optRepeat; // Repeat optimizer phases k times
7706 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7707 bool dspCode; // Display native code generated
7708 bool dspEHTable; // Display the EH table reported to the VM
7709 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7710 bool dspEmit; // Display emitter output
7711 bool dspLines; // Display source-code lines intermixed with native code output
7712 bool dmpHex; // Display raw bytes in hex of native code output
7713 bool varNames; // Display variables names in native code output
7714 bool disAsm; // Display native code as it is generated
7715 bool disAsmSpilled; // Display native code when any register spilling occurs
7716 bool disDiffable; // Makes the Disassembly code 'diff-able'
7717 bool disAsm2; // Display native code after it is generated using external disassembler
7718 bool dspOrder; // Display names of each of the methods that we ngen/jit
7719 bool dspUnwind; // Display the unwind info output
7720 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7721 bool compLongAddress; // Force using large pseudo instructions for long address
7722 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7723 bool dspGCtbls; // Display the GC tables
7727 bool doLateDisasm; // Run the late disassembler
7728 #endif // LATE_DISASM
7730 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7731 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7732 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7733 static const bool dspGCtbls = true;
7736 // We need stack probes to guarantee that we won't trigger a stack overflow
7737 // when calling unmanaged code until they get a chance to set up a frame, because
7738 // the EE will have no idea where it is.
7740 // We will only be doing this currently for hosted environments. Unfortunately
7741 // we need to take care of stubs, so potentially, we will have to do the probes
7742 // for any call. We have a plan for not needing for stubs though
7743 bool compNeedStackProbes;
7745 #ifdef PROFILING_SUPPORTED
7746 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7747 // This option helps make the JIT behave as if it is running under a profiler.
7748 bool compJitELTHookEnabled;
7749 #endif // PROFILING_SUPPORTED
7751 #if FEATURE_TAILCALL_OPT
7752 // Whether opportunistic or implicit tail call optimization is enabled.
7753 bool compTailCallOpt;
7754 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7755 bool compTailCallLoopOpt;
7759 static const bool compUseSoftFP = true;
7760 #else // !ARM_SOFTFP
7761 static const bool compUseSoftFP = false;
7764 GCPollType compGCPollType;
7768 static bool s_pAltJitExcludeAssembliesListInitialized;
7769 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7774 template <typename T>
7777 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7780 template <typename T>
7783 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7786 static int dspTreeID(GenTree* tree)
7788 return tree->gtTreeID;
7790 static void printTreeID(GenTree* tree)
7792 if (tree == nullptr)
7798 printf("[%06d]", dspTreeID(tree));
7805 #define STRESS_MODES \
7809 /* "Variations" stress areas which we try to mix up with each other. */ \
7810 /* These should not be exhaustively used as they might */ \
7811 /* hide/trivialize other areas */ \
7813 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7814 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7815 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7816 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7817 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7818 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7819 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7820 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7821 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7822 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7823 STRESS_MODE(NULL_OBJECT_CHECK) \
7824 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7825 STRESS_MODE(RANDOM_INLINE) \
7827 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7829 /* "Check" stress areas that can be exhaustively used if we */ \
7830 /* dont care about performance at all */ \
7832 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7833 STRESS_MODE(CHK_FLOW_UPDATE) \
7834 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7836 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7840 #define STRESS_MODE(mode) STRESS_##mode,
7847 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7848 BYTE compActiveStressModes[STRESS_COUNT];
7851 #define MAX_STRESS_WEIGHT 100
7853 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7857 bool compInlineStress()
7859 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7862 bool compRandomInlineStress()
7864 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7869 bool compTailCallStress()
7872 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7878 codeOptimize compCodeOpt()
7881 // Switching between size & speed has measurable throughput impact
7882 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7883 // DEBUG, but should generate identical code between CHK & RET builds,
7884 // so that's not acceptable.
7885 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7886 // Investigate the cause of the throughput regression.
7888 return opts.compCodeOpt;
7890 return BLENDED_CODE;
7894 //--------------------- Info about the procedure --------------------------
7898 COMP_HANDLE compCompHnd;
7899 CORINFO_MODULE_HANDLE compScopeHnd;
7900 CORINFO_CLASS_HANDLE compClassHnd;
7901 CORINFO_METHOD_HANDLE compMethodHnd;
7902 CORINFO_METHOD_INFO* compMethodInfo;
7904 BOOL hasCircularClassConstraints;
7905 BOOL hasCircularMethodConstraints;
7907 #if defined(DEBUG) || defined(LATE_DISASM)
7908 const char* compMethodName;
7909 const char* compClassName;
7910 const char* compFullName;
7911 #endif // defined(DEBUG) || defined(LATE_DISASM)
7913 #if defined(DEBUG) || defined(INLINE_DATA)
7914 // Method hash is logcally const, but computed
7916 mutable unsigned compMethodHashPrivate;
7917 unsigned compMethodHash() const;
7918 #endif // defined(DEBUG) || defined(INLINE_DATA)
7920 #ifdef PSEUDORANDOM_NOP_INSERTION
7921 // things for pseudorandom nop insertion
7922 unsigned compChecksum;
7926 // The following holds the FLG_xxxx flags for the method we're compiling.
7929 // The following holds the class attributes for the method we're compiling.
7930 unsigned compClassAttr;
7932 const BYTE* compCode;
7933 IL_OFFSET compILCodeSize; // The IL code size
7934 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7935 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7936 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7937 // (2) the code is hot/cold split, and we issued less code than we expected
7938 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7940 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7941 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7942 bool compIsContextful : 1; // contextful method
7943 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7944 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7945 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7946 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7947 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7949 var_types compRetType; // Return type of the method as declared in IL
7950 var_types compRetNativeType; // Normalized return type as per target arch ABI
7951 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7952 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7953 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7954 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7955 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7956 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7957 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7958 unsigned compMaxStack;
7959 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7960 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7962 unsigned compCallUnmanaged; // count of unmanaged calls
7963 unsigned compLvFrameListRoot; // lclNum for the Frame root
7964 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7965 // You should generally use compHndBBtabCount instead: it is the
7966 // current number of EH clauses (after additions like synchronized
7967 // methods and funclets, and removals like unreachable code deletion).
7969 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7970 // and the VM expects that, or the JIT is a "self-host" compiler
7971 // (e.g., x86 hosted targeting x86) and the VM expects that.
7973 /* The following holds IL scope information about local variables.
7976 unsigned compVarScopesCount;
7977 VarScopeDsc* compVarScopes;
7979 /* The following holds information about instr offsets for
7980 * which we need to report IP-mappings
7983 IL_OFFSET* compStmtOffsets; // sorted
7984 unsigned compStmtOffsetsCount;
7985 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
7987 #define CPU_X86 0x0100 // The generic X86 CPU
7988 #define CPU_X86_PENTIUM_4 0x0110
7990 #define CPU_X64 0x0200 // The generic x64 CPU
7991 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
7992 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
7994 #define CPU_ARM 0x0300 // The generic ARM CPU
7996 unsigned genCPU; // What CPU are we running on
7999 // Returns true if the method being compiled returns a non-void and non-struct value.
8000 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8001 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8002 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8003 // Methods returning such structs are considered to return non-struct return value and
8004 // this method returns true in that case.
8005 bool compMethodReturnsNativeScalarType()
8007 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8010 // Returns true if the method being compiled returns RetBuf addr as its return value
8011 bool compMethodReturnsRetBufAddr()
8013 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8014 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8016 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8017 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8018 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8019 // methods with hidden RetBufArg.
8021 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8022 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8023 // returning the address of RetBuf.
8025 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8026 // to be returned in RAX.
8027 CLANG_FORMAT_COMMENT_ANCHOR;
8029 #ifdef _TARGET_AMD64_
8030 return (info.compRetBuffArg != BAD_VAR_NUM);
8031 #else // !_TARGET_AMD64_
8032 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8033 #endif // !_TARGET_AMD64_
8036 // Returns true if the method returns a value in more than one return register
8037 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8038 // TODO-ARM64: Does this apply for ARM64 too?
8039 bool compMethodReturnsMultiRegRetType()
8041 #if FEATURE_MULTIREG_RET
8042 #if defined(_TARGET_X86_)
8043 // On x86 only 64-bit longs are returned in multiple registers
8044 return varTypeIsLong(info.compRetNativeType);
8045 #else // targets: X64-UNIX, ARM64 or ARM32
8046 // On all other targets that support multireg return values:
8047 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8048 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8049 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8050 #endif // TARGET_XXX
8052 #else // not FEATURE_MULTIREG_RET
8054 // For this architecture there are no multireg returns
8057 #endif // FEATURE_MULTIREG_RET
8060 #if FEATURE_MULTIREG_ARGS
8061 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8062 // return the gcPtr layout for the pointers sized fields
8063 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8064 #endif // FEATURE_MULTIREG_ARGS
8066 // Returns true if the method being compiled returns a value
8067 bool compMethodHasRetVal()
8069 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8070 compMethodReturnsMultiRegRetType();
8075 void compDispLocalVars();
8079 //-------------------------- Global Compiler Data ------------------------------------
8082 static unsigned s_compMethodsCount; // to produce unique label names
8083 unsigned compGenTreeID;
8086 BasicBlock* compCurBB; // the current basic block in process
8087 GenTreePtr compCurStmt; // the current statement in process
8089 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8092 // The following is used to create the 'method JIT info' block.
8093 size_t compInfoBlkSize;
8094 BYTE* compInfoBlkAddr;
8096 EHblkDsc* compHndBBtab; // array of EH data
8097 unsigned compHndBBtabCount; // element count of used elements in EH data array
8098 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8100 #if defined(_TARGET_X86_)
8102 //-------------------------------------------------------------------------
8103 // Tracking of region covered by the monitor in synchronized methods
8104 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8105 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8107 #endif // !_TARGET_X86_
8109 Phases previousCompletedPhase; // the most recently completed phase
8111 //-------------------------------------------------------------------------
8112 // The following keeps track of how many bytes of local frame space we've
8113 // grabbed so far in the current function, and how many argument bytes we
8114 // need to pop when we return.
8117 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8119 // Count of callee-saved regs we pushed in the prolog.
8120 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8121 // In case of Amd64 this doesn't include float regs saved on stack.
8122 unsigned compCalleeRegsPushed;
8124 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8125 // Mask of callee saved float regs on stack.
8126 regMaskTP compCalleeFPRegsSavedMask;
8128 #ifdef _TARGET_AMD64_
8129 // Quirk for VS debug-launch scenario to work:
8130 // Bytes of padding between save-reg area and locals.
8131 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8132 unsigned compVSQuirkStackPaddingNeeded;
8133 bool compQuirkForPPPflag;
8136 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8138 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8139 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8140 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8142 //-------------------------------------------------------------------------
8144 static void compStartup(); // One-time initialization
8145 static void compShutdown(); // One-time finalization
8147 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8150 static void compDisplayStaticSizes(FILE* fout);
8152 //------------ Some utility functions --------------
8154 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8155 void** ppIndirection); /* OUT */
8157 // Several JIT/EE interface functions return a CorInfoType, and also return a
8158 // class handle as an out parameter if the type is a value class. Returns the
8159 // size of the type these describe.
8160 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8163 // Components used by the compiler may write unit test suites, and
8164 // have them run within this method. They will be run only once per process, and only
8165 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8166 // These should fail by asserting.
8167 void compDoComponentUnitTestsOnce();
8170 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8171 CORINFO_MODULE_HANDLE classPtr,
8172 COMP_HANDLE compHnd,
8173 CORINFO_METHOD_INFO* methodInfo,
8174 void** methodCodePtr,
8175 ULONG* methodCodeSize,
8176 JitFlags* compileFlags);
8177 void compCompileFinish();
8178 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8179 COMP_HANDLE compHnd,
8180 CORINFO_METHOD_INFO* methodInfo,
8181 void** methodCodePtr,
8182 ULONG* methodCodeSize,
8183 JitFlags* compileFlags,
8184 CorInfoInstantiationVerification instVerInfo);
8186 ArenaAllocator* compGetAllocator();
8188 #if MEASURE_MEM_ALLOC
8190 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8194 unsigned allocCnt; // # of allocs
8195 UINT64 allocSz; // total size of those alloc.
8196 UINT64 allocSzMax; // Maximum single allocation.
8197 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8198 UINT64 nraTotalSizeAlloc;
8199 UINT64 nraTotalSizeUsed;
8201 static const char* s_CompMemKindNames[]; // Names of the kinds.
8203 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8205 for (int i = 0; i < CMK_Count; i++)
8207 allocSzByKind[i] = 0;
8210 MemStats(const MemStats& ms)
8211 : allocCnt(ms.allocCnt)
8212 , allocSz(ms.allocSz)
8213 , allocSzMax(ms.allocSzMax)
8214 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8215 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8217 for (int i = 0; i < CMK_Count; i++)
8219 allocSzByKind[i] = ms.allocSzByKind[i];
8223 // Until we have ubiquitous constructors.
8226 this->MemStats::MemStats();
8229 void AddAlloc(size_t sz, CompMemKind cmk)
8233 if (sz > allocSzMax)
8237 allocSzByKind[cmk] += sz;
8240 void Print(FILE* f); // Print these stats to f.
8241 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8243 MemStats genMemStats;
8245 struct AggregateMemStats : public MemStats
8249 AggregateMemStats() : MemStats(), nMethods(0)
8253 void Add(const MemStats& ms)
8256 allocCnt += ms.allocCnt;
8257 allocSz += ms.allocSz;
8258 allocSzMax = max(allocSzMax, ms.allocSzMax);
8259 for (int i = 0; i < CMK_Count; i++)
8261 allocSzByKind[i] += ms.allocSzByKind[i];
8263 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8264 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8267 void Print(FILE* f); // Print these stats to jitstdout.
8270 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8271 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8272 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8274 #endif // MEASURE_MEM_ALLOC
8276 #if LOOP_HOIST_STATS
8277 unsigned m_loopsConsidered;
8278 bool m_curLoopHasHoistedExpression;
8279 unsigned m_loopsWithHoistedExpressions;
8280 unsigned m_totalHoistedExpressions;
8282 void AddLoopHoistStats();
8283 void PrintPerMethodLoopHoistStats();
8285 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8286 static unsigned s_loopsConsidered;
8287 static unsigned s_loopsWithHoistedExpressions;
8288 static unsigned s_totalHoistedExpressions;
8290 static void PrintAggregateLoopHoistStats(FILE* f);
8291 #endif // LOOP_HOIST_STATS
8293 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8294 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8295 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8296 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8297 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8298 void compFreeMem(void*);
8300 bool compIsForImportOnly();
8301 bool compIsForInlining();
8302 bool compDonotInline();
8305 const char* compLocalVarName(unsigned varNum, unsigned offs);
8306 VarName compVarName(regNumber reg, bool isFloatReg = false);
8307 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8308 const char* compRegPairName(regPairNo regPair);
8309 const char* compRegNameForSize(regNumber reg, size_t size);
8310 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8311 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8312 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8315 //-------------------------------------------------------------------------
8317 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8319 struct VarScopeMapInfo
8321 VarScopeListNode* head;
8322 VarScopeListNode* tail;
8323 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8325 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8332 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8333 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8335 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8336 VarNumToScopeDscMap;
8338 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8339 VarNumToScopeDscMap* compVarScopeMap;
8341 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8343 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8345 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8347 void compInitVarScopeMap();
8349 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8350 // enter scope, sorted by instr offset
8351 unsigned compNextEnterScope;
8353 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8354 // go out of scope, sorted by instr offset
8355 unsigned compNextExitScope;
8357 void compInitScopeLists();
8359 void compResetScopeLists();
8361 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8363 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8365 void compProcessScopesUntil(unsigned offset,
8367 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8368 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8371 void compDispScopeLists();
8374 bool compIsProfilerHookNeeded();
8376 //-------------------------------------------------------------------------
8377 /* Statistical Data Gathering */
8379 void compJitStats(); // call this function and enable
8380 // various ifdef's below for statistical data
8383 void compCallArgStats();
8384 static void compDispCallArgStats(FILE* fout);
8387 //-------------------------------------------------------------------------
8394 ArenaAllocator* compAllocator;
8397 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8398 // suitable for use by utilcode collection types.
8399 IAllocator* compAsIAllocator;
8401 #if MEASURE_MEM_ALLOC
8402 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8403 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8404 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8406 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8408 #endif // MEASURE_MEM_ALLOC
8410 void compFunctionTraceStart();
8411 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8414 size_t compMaxUncheckedOffsetForNullObject;
8416 void compInitOptions(JitFlags* compileFlags);
8418 void compSetProcessor();
8419 void compInitDebuggingInfo();
8420 void compSetOptimizationLevel();
8421 #ifdef _TARGET_ARMARCH_
8422 bool compRsvdRegCheck(FrameLayoutState curState);
8424 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8426 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8427 void ResetOptAnnotations();
8429 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8430 void RecomputeLoopInfo();
8432 #ifdef PROFILING_SUPPORTED
8433 // Data required for generating profiler Enter/Leave/TailCall hooks
8435 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8436 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8437 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8440 #ifdef _TARGET_AMD64_
8441 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8444 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8445 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8447 IAllocator* getAllocator()
8449 return compAsIAllocator;
8452 #if MEASURE_MEM_ALLOC
8453 IAllocator* getAllocatorBitset()
8455 return compAsIAllocatorBitset;
8457 IAllocator* getAllocatorGC()
8459 return compAsIAllocatorGC;
8461 IAllocator* getAllocatorLoopHoist()
8463 return compAsIAllocatorLoopHoist;
8465 #else // !MEASURE_MEM_ALLOC
8466 IAllocator* getAllocatorBitset()
8468 return compAsIAllocator;
8470 IAllocator* getAllocatorGC()
8472 return compAsIAllocator;
8474 IAllocator* getAllocatorLoopHoist()
8476 return compAsIAllocator;
8478 #endif // !MEASURE_MEM_ALLOC
8481 IAllocator* getAllocatorDebugOnly()
8483 #if MEASURE_MEM_ALLOC
8484 return compAsIAllocatorDebugOnly;
8485 #else // !MEASURE_MEM_ALLOC
8486 return compAsIAllocator;
8487 #endif // !MEASURE_MEM_ALLOC
8492 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8493 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8497 XX Checks for type compatibility and merges types XX
8499 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8500 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8504 // Set to TRUE if verification cannot be skipped for this method
8505 // If we detect unverifiable code, we will lazily check
8506 // canSkipMethodVerification() to see if verification is REALLY needed.
8507 BOOL tiVerificationNeeded;
8509 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8510 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8511 BOOL tiIsVerifiableCode;
8513 // Set to TRUE if runtime callout is needed for this method
8514 BOOL tiRuntimeCalloutNeeded;
8516 // Set to TRUE if security prolog/epilog callout is needed for this method
8517 // Note: This flag is different than compNeedSecurityCheck.
8518 // compNeedSecurityCheck means whether or not a security object needs
8519 // to be allocated on the stack, which is currently true for EnC as well.
8520 // tiSecurityCalloutNeeded means whether or not security callouts need
8521 // to be inserted in the jitted code.
8522 BOOL tiSecurityCalloutNeeded;
8524 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8525 // This support is necessary to suport attributes that are not described in
8526 // for example, signatures. For example, the permanent home byref (byref that
8527 // points to the gc heap), isn't a property of method signatures, therefore,
8528 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8529 // but when deciding if we need to reimport a block, we need to take these
8531 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8533 // Returns TRUE if child is equal to or a subtype of parent.
8534 // normalisedForStack indicates that both types are normalised for the stack
8535 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8537 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8538 // *pDest is modified to represent the merged type. Sets "*changed" to true
8539 // if this changes "*pDest".
8540 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8542 // Set pDest from the primitive value type.
8543 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8545 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8548 // <BUGNUM> VSW 471305
8549 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8550 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8551 // We use a "short" as we need to push/pop this scope.
8553 short compRegSetCheckLevel;
8557 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8558 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8560 XX IL verification stuff XX
8563 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8564 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8568 // The following is used to track liveness of local variables, initialization
8569 // of valueclass constructors, and type safe use of IL instructions.
8571 // dynamic state info needed for verification
8572 EntryState verCurrentState;
8574 // this ptr of object type .ctors are considered intited only after
8575 // the base class ctor is called, or an alternate ctor is called.
8576 // An uninited this ptr can be used to access fields, but cannot
8577 // be used to call a member function.
8578 BOOL verTrackObjCtorInitState;
8580 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8582 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8583 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8584 void verInitCurrentState();
8585 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8587 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8588 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8589 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8591 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8592 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8593 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8594 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8595 typeInfo verMakeTypeInfo(CorInfoType ciType,
8596 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8597 BOOL verIsSDArray(typeInfo ti);
8598 typeInfo verGetArrayElemType(typeInfo ti);
8600 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8601 BOOL verNeedsVerification();
8602 BOOL verIsByRefLike(const typeInfo& ti);
8603 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8605 // generic type variables range over types that satisfy IsBoxable
8606 BOOL verIsBoxable(const typeInfo& ti);
8608 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8609 DEBUGARG(unsigned line));
8610 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8611 DEBUGARG(unsigned line));
8612 bool verCheckTailCallConstraint(OPCODE opcode,
8613 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8614 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8615 // on a type parameter?
8616 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8617 // return false to the caller.
8618 // If false, it will throw.
8620 bool verIsBoxedValueType(typeInfo ti);
8622 void verVerifyCall(OPCODE opcode,
8623 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8624 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8626 bool readonlyCall, // is this a "readonly." call?
8627 const BYTE* delegateCreateStart,
8628 const BYTE* codeAddr,
8629 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8631 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8633 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8634 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8635 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8636 const CORINFO_FIELD_INFO& fieldInfo,
8637 const typeInfo* tiThis,
8639 BOOL allowPlainStructAsThis = FALSE);
8640 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8641 void verVerifyThisPtrInitialised();
8642 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8644 // Register allocator
8645 void raInitStackFP();
8646 void raEnregisterVarsPrePassStackFP();
8647 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8648 void raEnregisterVarsPostPassStackFP();
8649 void raGenerateFPRefCounts();
8650 void raEnregisterVarsStackFP();
8651 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8653 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8654 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8656 // returns true if enregistering v1 would save more mem accesses than v2
8657 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8660 void raDumpHeightsStackFP();
8661 void raDumpVariableRegIntfFloat();
8664 #if FEATURE_STACK_FP_X87
8666 // Currently, we use FP transition blocks in only 2 situations:
8668 // -conditional jump on longs where FP stack differs with target: it's not strictly
8669 // necessary, but its low frequency and the code would get complicated if we try to
8670 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8671 // minimize the way we generate the jump code.
8672 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8673 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8675 // However, transition blocks have 2 problems
8677 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8678 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8679 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8680 // in the right place without preordering them), this causes us to have to generate the transition
8681 // blocks in the cold area if we want procedure splitting.
8684 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8685 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8686 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8687 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8688 // a big change in the exception.
8690 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8691 // optimizations. For these 2 cases:
8693 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8694 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8695 // a switch statement.
8697 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8698 // current procedure splitting and exception code have.
8699 bool compMayHaveTransitionBlocks;
8701 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8703 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8705 unsigned raCntStkStackFP;
8706 unsigned raCntWtdStkDblStackFP;
8707 unsigned raCntStkParamDblStackFP;
8709 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8710 // TODO: Do we want to put this in LclVarDsc?
8711 unsigned raPayloadStackFP[lclMAX_TRACKED];
8712 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8714 // Useful for debugging
8715 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8717 #endif // FEATURE_STACK_FP_X87
8720 // One line log function. Default level is 0. Increasing it gives you
8721 // more log information
8723 // levels are currently unused: #define JITDUMP(level,...) ();
8724 void JitLogEE(unsigned level, const char* fmt, ...);
8726 bool compDebugBreak;
8728 bool compJitHaltMethod();
8733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8734 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8736 XX GS Security checks for unsafe buffers XX
8738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8739 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8742 struct ShadowParamVarInfo
8744 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8745 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8747 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8749 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8750 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8751 // slots and update all trees to refer to shadow slots is done immediately after
8752 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8753 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8754 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8755 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8756 // creating a shadow slot even though this routine returns true.
8758 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8759 // required. There are two cases under which a reg arg could potentially be used from its
8761 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8762 // b) LSRA spills it
8764 // Possible solution to address case (a)
8765 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8766 // in this routine. Note that live out of exception handler is something we may not be
8767 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8768 // Therefore, for methods with exception handling and need GS cookie check we might have
8769 // to take conservative approach.
8771 // Possible solution to address case (b)
8772 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8773 // create a new spill temp if the method needs GS cookie check.
8774 return varDsc->lvIsParam;
8775 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8776 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8783 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8788 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8789 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8790 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8792 void gsGSChecksInitCookie(); // Grabs cookie variable
8793 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8794 bool gsFindVulnerableParams(); // Shadow param analysis code
8795 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8797 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8798 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8800 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8801 // This can be overwritten by setting complus_JITInlineSize env variable.
8803 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8806 #ifdef FEATURE_JIT_METHOD_PERF
8807 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8808 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8810 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8811 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8813 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8815 #if MEASURE_CLRAPI_CALLS
8816 // Thin wrappers that call into JitTimer (if present).
8817 inline void CLRApiCallEnter(unsigned apix);
8818 inline void CLRApiCallLeave(unsigned apix);
8821 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8822 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8827 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8828 // These variables are associated with maintaining SQM data about compile time.
8829 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8830 // in the current compilation.
8831 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8832 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8833 // the inlining phase in the current compilation.
8834 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8836 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8837 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8838 // type-loading and class initialization).
8839 void RecordStateAtEndOfInlining();
8840 // Assumes being called at the end of compilation. Update the SQM state.
8841 void RecordStateAtEndOfCompilation();
8843 #ifdef FEATURE_CLRSQM
8844 // Does anything SQM related necessary at process shutdown time.
8845 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8846 #endif // FEATURE_CLRSQM
8849 #if FUNC_INFO_LOGGING
8850 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8851 // filename to write it to.
8852 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8853 #endif // FUNC_INFO_LOGGING
8855 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8857 // Is the compilation in a full trust context?
8858 bool compIsFullTrust();
8860 #ifndef FEATURE_TRACELOGGING
8861 // Should we actually fire the noway assert body and the exception handler?
8862 bool compShouldThrowOnNoway();
8863 #else // FEATURE_TRACELOGGING
8864 // Should we actually fire the noway assert body and the exception handler?
8865 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8867 // Telemetry instance to use per method compilation.
8868 JitTelemetry compJitTelemetry;
8870 // Get common parameters that have to be logged with most telemetry data.
8871 void compGetTelemetryDefaults(const char** assemblyName,
8872 const char** scopeName,
8873 const char** methodName,
8874 unsigned* methodHash);
8875 #endif // !FEATURE_TRACELOGGING
8879 NodeToTestDataMap* m_nodeTestData;
8881 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8882 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8883 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8884 // Current kept in this.
8886 NodeToTestDataMap* GetNodeTestData()
8888 Compiler* compRoot = impInlineRoot();
8889 if (compRoot->m_nodeTestData == nullptr)
8891 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8893 return compRoot->m_nodeTestData;
8896 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8898 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8899 // currently occur in the AST graph.
8900 NodeToIntMap* FindReachableNodesInNodeTestData();
8902 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8903 // test data, associate that data with "to".
8904 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8906 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8907 // have annotations, attach similar annotations to the corresponding nodes in "to".
8908 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8910 // These are the methods that test that the various conditions implied by the
8911 // test attributes are satisfied.
8912 void JitTestCheckSSA(); // SSA builder tests.
8913 void JitTestCheckVN(); // Value numbering tests.
8916 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8918 FieldSeqStore* m_fieldSeqStore;
8920 FieldSeqStore* GetFieldSeqStore()
8922 Compiler* compRoot = impInlineRoot();
8923 if (compRoot->m_fieldSeqStore == nullptr)
8925 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8926 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8927 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8929 return compRoot->m_fieldSeqStore;
8932 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8934 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8935 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8936 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8937 // attach the field sequence directly to the address node.
8938 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8940 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8942 // Don't need to worry about inlining here
8943 if (m_zeroOffsetFieldMap == nullptr)
8945 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8947 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8948 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8950 return m_zeroOffsetFieldMap;
8953 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8954 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8955 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8956 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8957 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8958 // record the the field sequence using the ZeroOffsetFieldMap described above.
8960 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8961 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8962 // CoreRT. Such case is handled same as the default case.
8963 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8965 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8967 NodeToArrayInfoMap* m_arrayInfoMap;
8969 NodeToArrayInfoMap* GetArrayInfoMap()
8971 Compiler* compRoot = impInlineRoot();
8972 if (compRoot->m_arrayInfoMap == nullptr)
8974 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8975 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8976 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
8978 return compRoot->m_arrayInfoMap;
8981 NodeToUnsignedMap* m_heapSsaMap;
8983 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
8984 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
8985 // all the possible heap states are possible initial states of the corresponding catch block(s).)
8986 NodeToUnsignedMap* GetHeapSsaMap()
8988 Compiler* compRoot = impInlineRoot();
8989 if (compRoot->m_heapSsaMap == nullptr)
8991 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8992 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8993 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
8995 return compRoot->m_heapSsaMap;
8998 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
8999 CORINFO_CLASS_HANDLE m_refAnyClass;
9000 CORINFO_FIELD_HANDLE GetRefanyDataField()
9002 if (m_refAnyClass == nullptr)
9004 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9006 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9008 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9010 if (m_refAnyClass == nullptr)
9012 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9014 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9018 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9020 #if ALLVARSET_COUNTOPS
9021 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9024 static HelperCallProperties s_helperCallProperties;
9026 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9027 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9028 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9030 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9033 unsigned __int8* offset0,
9034 unsigned __int8* offset1);
9035 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9036 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9038 void fgMorphMultiregStructArgs(GenTreeCall* call);
9039 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9041 }; // end of class Compiler
9043 // Inline methods of CompAllocator.
9044 void* CompAllocator::Alloc(size_t sz)
9046 #if MEASURE_MEM_ALLOC
9047 return m_comp->compGetMem(sz, m_cmk);
9049 return m_comp->compGetMem(sz);
9053 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9055 #if MEASURE_MEM_ALLOC
9056 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9058 return m_comp->compGetMemArray(elems, elemSize);
9062 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9063 inline LclVarDsc::LclVarDsc(Compiler* comp)
9064 : // Initialize the ArgRegs to REG_STK.
9065 // The morph will do the right thing to change
9066 // to the right register if passed in register.
9069 #if FEATURE_MULTIREG_ARGS
9070 _lvOtherArgReg(REG_STK)
9072 #endif // FEATURE_MULTIREG_ARGS
9074 lvRefBlks(BlockSetOps::UninitVal())
9076 #endif // ASSERTION_PROP
9077 lvPerSsaData(comp->getAllocator())
9082 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9083 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9085 XX Miscellaneous Compiler stuff XX
9087 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9088 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9091 // Values used to mark the types a stack slot is used for
9093 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9094 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9095 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9096 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9097 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9098 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9099 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9100 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9102 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9104 /*****************************************************************************
9106 * Variables to keep track of total code amounts.
9111 extern size_t grossVMsize;
9112 extern size_t grossNCsize;
9113 extern size_t totalNCsize;
9115 extern unsigned genMethodICnt;
9116 extern unsigned genMethodNCnt;
9117 extern size_t gcHeaderISize;
9118 extern size_t gcPtrMapISize;
9119 extern size_t gcHeaderNSize;
9120 extern size_t gcPtrMapNSize;
9122 #endif // DISPLAY_SIZES
9124 /*****************************************************************************
9126 * Variables to keep track of basic block counts (more data on 1 BB methods)
9129 #if COUNT_BASIC_BLOCKS
9130 extern Histogram bbCntTable;
9131 extern Histogram bbOneBBSizeTable;
9134 /*****************************************************************************
9136 * Used by optFindNaturalLoops to gather statistical information such as
9137 * - total number of natural loops
9138 * - number of loops with 1, 2, ... exit conditions
9139 * - number of loops that have an iterator (for like)
9140 * - number of loops that have a constant iterator
9145 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9146 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9147 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9148 extern unsigned totalLoopCount; // counts the total number of natural loops
9149 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9150 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9151 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9152 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9154 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9155 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9156 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9157 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9158 extern Histogram loopCountTable; // Histogram of loop counts
9159 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9161 #endif // COUNT_LOOPS
9163 /*****************************************************************************
9164 * variables to keep track of how many iterations we go in a dataflow pass
9169 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9170 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9172 #endif // DATAFLOW_ITER
9174 #if MEASURE_BLOCK_SIZE
9175 extern size_t genFlowNodeSize;
9176 extern size_t genFlowNodeCnt;
9177 #endif // MEASURE_BLOCK_SIZE
9179 #if MEASURE_NODE_SIZE
9180 struct NodeSizeStats
9185 genTreeNodeSize = 0;
9186 genTreeNodeActualSize = 0;
9189 size_t genTreeNodeCnt;
9190 size_t genTreeNodeSize; // The size we allocate
9191 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9192 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9193 // a smaller node to a larger one. TODO-Cleanup: add stats on
9194 // SetOper()/ChangeOper() usage to quanitfy this.
9196 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9197 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9198 extern Histogram genTreeNcntHist;
9199 extern Histogram genTreeNsizHist;
9200 #endif // MEASURE_NODE_SIZE
9202 /*****************************************************************************
9203 * Count fatal errors (including noway_asserts).
9207 extern unsigned fatal_badCode;
9208 extern unsigned fatal_noWay;
9209 extern unsigned fatal_NOMEM;
9210 extern unsigned fatal_noWayAssertBody;
9212 extern unsigned fatal_noWayAssertBodyArgs;
9214 extern unsigned fatal_NYI;
9215 #endif // MEASURE_FATAL
9217 /*****************************************************************************
9221 #ifdef _TARGET_XARCH_
9223 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9224 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9225 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9227 const instruction INS_AND = INS_and;
9228 const instruction INS_OR = INS_or;
9229 const instruction INS_XOR = INS_xor;
9230 const instruction INS_NEG = INS_neg;
9231 const instruction INS_TEST = INS_test;
9232 const instruction INS_MUL = INS_imul;
9233 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9234 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9235 const instruction INS_BREAKPOINT = INS_int3;
9236 const instruction INS_ADDC = INS_adc;
9237 const instruction INS_SUBC = INS_sbb;
9238 const instruction INS_NOT = INS_not;
9244 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9245 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9246 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9248 const instruction INS_AND = INS_and;
9249 const instruction INS_OR = INS_orr;
9250 const instruction INS_XOR = INS_eor;
9251 const instruction INS_NEG = INS_rsb;
9252 const instruction INS_TEST = INS_tst;
9253 const instruction INS_MUL = INS_mul;
9254 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9255 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9256 const instruction INS_BREAKPOINT = INS_bkpt;
9257 const instruction INS_ADDC = INS_adc;
9258 const instruction INS_SUBC = INS_sbc;
9259 const instruction INS_NOT = INS_mvn;
9263 #ifdef _TARGET_ARM64_
9265 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9266 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9267 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9269 const instruction INS_AND = INS_and;
9270 const instruction INS_OR = INS_orr;
9271 const instruction INS_XOR = INS_eor;
9272 const instruction INS_NEG = INS_neg;
9273 const instruction INS_TEST = INS_tst;
9274 const instruction INS_MUL = INS_mul;
9275 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9276 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9277 const instruction INS_BREAKPOINT = INS_bkpt;
9278 const instruction INS_ADDC = INS_adc;
9279 const instruction INS_SUBC = INS_sbc;
9280 const instruction INS_NOT = INS_mvn;
9284 /*****************************************************************************/
9286 extern const BYTE genTypeSizes[];
9287 extern const BYTE genTypeAlignments[];
9288 extern const BYTE genTypeStSzs[];
9289 extern const BYTE genActualTypes[];
9291 /*****************************************************************************/
9293 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9294 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9297 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9298 #elif defined(_TARGET_ARM64_)
9299 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9302 /*****************************************************************************/
9304 #define REG_CORRUPT regNumber(REG_NA + 1)
9305 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9306 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9308 /*****************************************************************************/
9310 extern BasicBlock dummyBB;
9312 /*****************************************************************************/
9313 /*****************************************************************************/
9315 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9316 // nodes of a statement in execution order.
9317 // __stmt: a GT_STMT type GenTree*
9318 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9320 #define foreach_treenode_execution_order(__node, __stmt) \
9321 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9323 // foreach_block: An iterator over all blocks in the function.
9324 // __compiler: the Compiler* object
9325 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9327 #define foreach_block(__compiler, __block) \
9328 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9330 /*****************************************************************************/
9331 /*****************************************************************************/
9335 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9337 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9338 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9340 XX Debugging helpers XX
9342 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9343 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9346 /*****************************************************************************/
9347 /* The following functions are intended to be called from the debugger, to dump
9348 * various data structures. The can be used in the debugger Watch or Quick Watch
9349 * windows. They are designed to be short to type and take as few arguments as
9350 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9351 * See the function definition comment for more details.
9354 void cBlock(Compiler* comp, BasicBlock* block);
9355 void cBlocks(Compiler* comp);
9356 void cBlocksV(Compiler* comp);
9357 void cTree(Compiler* comp, GenTree* tree);
9358 void cTrees(Compiler* comp);
9359 void cEH(Compiler* comp);
9360 void cVar(Compiler* comp, unsigned lclNum);
9361 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9362 void cVars(Compiler* comp);
9363 void cVarsFinal(Compiler* comp);
9364 void cBlockPreds(Compiler* comp, BasicBlock* block);
9365 void cReach(Compiler* comp);
9366 void cDoms(Compiler* comp);
9367 void cLiveness(Compiler* comp);
9368 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9370 void cFuncIR(Compiler* comp);
9371 void cBlockIR(Compiler* comp, BasicBlock* block);
9372 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9373 void cTreeIR(Compiler* comp, GenTree* tree);
9374 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9375 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9376 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9377 int cOperandIR(Compiler* comp, GenTree* operand);
9378 int cLeafIR(Compiler* comp, GenTree* tree);
9379 int cIndirIR(Compiler* comp, GenTree* tree);
9380 int cListIR(Compiler* comp, GenTree* list);
9381 int cSsaNumIR(Compiler* comp, GenTree* tree);
9382 int cValNumIR(Compiler* comp, GenTree* tree);
9383 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9385 void dBlock(BasicBlock* block);
9388 void dTree(GenTree* tree);
9391 void dVar(unsigned lclNum);
9392 void dVarDsc(LclVarDsc* varDsc);
9395 void dBlockPreds(BasicBlock* block);
9399 void dCVarSet(VARSET_VALARG_TP vars);
9401 void dVarSet(VARSET_VALARG_TP vars);
9402 void dRegMask(regMaskTP mask);
9405 void dBlockIR(BasicBlock* block);
9406 void dTreeIR(GenTree* tree);
9407 void dLoopIR(Compiler::LoopDsc* loop);
9408 void dLoopNumIR(unsigned loopNum);
9409 int dTabStopIR(int curr, int tabstop);
9410 int dTreeTypeIR(GenTree* tree);
9411 int dTreeKindsIR(GenTree* tree);
9412 int dTreeFlagsIR(GenTree* tree);
9413 int dOperandIR(GenTree* operand);
9414 int dLeafIR(GenTree* tree);
9415 int dIndirIR(GenTree* tree);
9416 int dListIR(GenTree* list);
9417 int dSsaNumIR(GenTree* tree);
9418 int dValNumIR(GenTree* tree);
9419 int dDependsIR(GenTree* comma);
9422 GenTree* dFindTree(GenTree* tree, unsigned id);
9423 GenTree* dFindTree(unsigned id);
9424 GenTreeStmt* dFindStmt(unsigned id);
9425 BasicBlock* dFindBlock(unsigned bbNum);
9429 #include "compiler.hpp" // All the shared inline functions
9431 /*****************************************************************************/
9432 #endif //_COMPILER_H_
9433 /*****************************************************************************/