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 #ifndef LEGACY_BACKEND
89 class Lowering; // defined in lower.h
92 // The following are defined in this file, Compiler.h
96 /*****************************************************************************
102 /*****************************************************************************/
105 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
108 // Or the more-general IAllocator interface.
109 void* __cdecl operator new(size_t n, IAllocator* alloc);
110 void* __cdecl operator new[](size_t n, IAllocator* alloc);
112 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
113 // caused these to be ambiguous with the global placement new operators.
114 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
115 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
116 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
118 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
119 #include "loopcloning.h"
121 /*****************************************************************************/
123 /* This is included here and not earlier as it needs the definition of "CSE"
124 * which is defined in the section above */
126 /*****************************************************************************/
128 unsigned genLog2(unsigned value);
129 unsigned genLog2(unsigned __int64 value);
131 var_types genActualType(var_types type);
132 var_types genUnsignedType(var_types type);
133 var_types genSignedType(var_types type);
135 unsigned ReinterpretHexAsDecimal(unsigned);
137 /*****************************************************************************/
140 #ifdef FEATURE_AVX_SUPPORT
141 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
142 #else // !FEATURE_AVX_SUPPORT
143 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
144 #endif // !FEATURE_AVX_SUPPORT
145 #else // !FEATURE_SIMD
146 const unsigned TEMP_MAX_SIZE = sizeof(double);
147 #endif // !FEATURE_SIMD
148 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
150 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
153 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
156 // The following holds the Local var info (scope information)
157 typedef const char* VarName; // Actual ASCII string
160 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
161 IL_OFFSET vsdLifeEnd; // instr offset of end of life
162 unsigned vsdVarNum; // (remapped) LclVarDsc number
165 VarName vsdName; // name of the var
168 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
169 // Also, it is the index of this entry in the info.compVarScopes array,
170 // which is useful since the array is also accessed via the
171 // compEnterScopeList and compExitScopeList sorted arrays.
174 /*****************************************************************************
176 * The following holds the local variable counts and the descriptor table.
179 // This is the location of a definition.
185 DefLoc() : m_blk(nullptr), m_tree(nullptr)
190 // This class encapsulates all info about a local variable that may vary for different SSA names
195 ValueNumPair m_vnPair;
203 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
208 // The constructor. Most things can just be zero'ed.
209 LclVarDsc(Compiler* comp);
211 // note this only packs because var_types is a typedef of unsigned char
212 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
214 unsigned char lvIsParam : 1; // is this a parameter?
215 unsigned char lvIsRegArg : 1; // is this a register argument?
216 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
218 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
219 // is to help determine whether to use block init in the prolog.
220 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
221 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
222 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
223 // variable is in the same register for the entire function.
224 unsigned char lvTracked : 1; // is this a tracked variable?
225 bool lvTrackedNonStruct()
227 return lvTracked && lvType != TYP_STRUCT;
229 unsigned char lvPinned : 1; // is this a pinned variable?
231 unsigned char lvMustInit : 1; // must be initialized
232 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
233 // global location, etc.
234 // We cannot reason reliably about the value of the variable.
235 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
236 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
240 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
242 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
243 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
244 // read/write its value.
245 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
246 // the variable to be
247 // in the stack (at least at those boundaries.)
248 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
249 // particular byte from an int).
250 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
251 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
253 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
254 unsigned char lvRefAssign : 1; // involved in pointer assignment
255 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
256 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
259 unsigned char lvHasILStoreOp : 1; // there is at least one STLOC or STARG on this local
260 unsigned char lvHasMultipleILStoreOp : 1; // there is more than one STLOC on this local
262 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp (if lvIsParam is false), or implicit byref parameter
263 // (if lvIsParam is true)
265 unsigned char lvIsBoolean : 1; // set if variable is boolean
267 unsigned char lvRngOptDone : 1; // considered for range check opt?
268 unsigned char lvLoopInc : 1; // incremented in the loop?
269 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
270 unsigned char lvArrIndx : 1; // used as an array index?
271 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
272 unsigned char lvArrIndxDom : 1; // index dominates loop exit
274 unsigned char lvSingleDef : 1; // variable has a single def
275 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
276 unsigned char lvVolatileHint : 1; // hint for AssertionProp
279 unsigned char lvSpilled : 1; // enregistered variable was spilled
280 #ifndef _TARGET_64BIT_
281 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
282 #endif // !_TARGET_64BIT_
283 #ifdef _TARGET_64BIT_
284 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
287 unsigned char lvKeepType : 1; // Don't change the type of this variable
288 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
290 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
292 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
293 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
294 // 32-bit target. For implicit byref parameters, this gets hijacked between
295 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to indicate whether
296 // references to the arg are being rewritten as references to a promoted shadow local.
297 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
298 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
299 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
300 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
301 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
303 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
304 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
307 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
308 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
309 // with (lvIsRegArg && lvIsHfa())
310 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
311 #endif // FEATURE_HFA
314 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
315 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
316 // Consider cleaning this up so this workaround is not required.
317 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
318 // I.e. there is no longer any reference to the struct directly.
319 // In this case we can simply remove this struct local.
321 #ifndef LEGACY_BACKEND
322 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
323 #endif // !LEGACY_BACKEND
326 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
327 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
328 unsigned char lvSIMDType : 1; // This is a SIMD struct
329 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
330 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
331 #endif // FEATURE_SIMD
332 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
334 unsigned char lvClassIsExact : 1; // lvClassHandle is the exact type
337 unsigned char lvClassInfoUpdated : 1; // true if this var has updated class handle or exactness
341 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
342 // local. For implicit byref parameters, this gets hijacked between
343 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to point to the
344 // struct local created to model the parameter's struct promotion, if any.
345 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
346 // Valid on promoted struct local fields.
349 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
350 unsigned char lvFldOffset;
351 unsigned char lvFldOrdinal;
353 #if FEATURE_MULTIREG_ARGS
354 regNumber lvRegNumForSlot(unsigned slotNum)
360 else if (slotNum == 1)
362 return lvOtherArgReg;
366 assert(false && "Invalid slotNum!");
371 #endif // FEATURE_MULTIREG_ARGS
389 bool lvIsHfaRegArg() const
392 return _lvIsHfaRegArg;
398 void lvSetIsHfaRegArg(bool value = true)
401 _lvIsHfaRegArg = value;
405 bool lvHfaTypeIsFloat() const
408 return _lvHfaTypeIsFloat;
414 void lvSetHfaTypeIsFloat(bool value)
417 _lvHfaTypeIsFloat = value;
421 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
422 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
424 unsigned lvHfaSlots() const
427 assert(lvType == TYP_STRUCT);
429 return lvExactSize / sizeof(float);
430 #else // _TARGET_ARM64_
431 if (lvHfaTypeIsFloat())
433 return lvExactSize / sizeof(float);
437 return lvExactSize / sizeof(double);
439 #endif // _TARGET_ARM64_
442 // lvIsMultiRegArgOrRet()
443 // returns true if this is a multireg LclVar struct used in an argument context
444 // or if this is a multireg LclVar struct assigned from a multireg call
445 bool lvIsMultiRegArgOrRet()
447 return lvIsMultiRegArg || lvIsMultiRegRet;
451 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
452 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
453 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
454 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
455 // to non-zero if the variable gets the same register assignment for its entire
457 #if !defined(_TARGET_64BIT_)
458 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
459 #endif // !defined(_TARGET_64BIT_)
461 regNumberSmall _lvArgReg; // The register in which this argument is passed.
463 #if FEATURE_MULTIREG_ARGS
464 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
465 // Note this is defined but not used by ARM32
466 #endif // FEATURE_MULTIREG_ARGS
468 #ifndef LEGACY_BACKEND
470 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
471 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
473 #endif // !LEGACY_BACKEND
476 // The register number is stored in a small format (8 bits), but the getters return and the setters take
477 // a full-size (unsigned) format, to localize the casts here.
479 /////////////////////
481 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
483 regNumber GetRegNum() const
485 return (regNumber)_lvRegNum;
488 void SetRegNum(regNumber reg)
490 _lvRegNum = (regNumberSmall)reg;
491 assert(_lvRegNum == reg);
494 /////////////////////
496 #if defined(_TARGET_64BIT_)
497 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
499 regNumber GetOtherReg() const
501 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
502 // "unreachable code" warnings
506 void SetOtherReg(regNumber reg)
508 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
509 // "unreachable code" warnings
511 #else // !_TARGET_64BIT_
512 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
514 regNumber GetOtherReg() const
516 return (regNumber)_lvOtherReg;
519 void SetOtherReg(regNumber reg)
521 _lvOtherReg = (regNumberSmall)reg;
522 assert(_lvOtherReg == reg);
524 #endif // !_TARGET_64BIT_
526 /////////////////////
528 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
530 regNumber GetArgReg() const
532 return (regNumber)_lvArgReg;
535 void SetArgReg(regNumber reg)
537 _lvArgReg = (regNumberSmall)reg;
538 assert(_lvArgReg == reg);
541 #if FEATURE_MULTIREG_ARGS
542 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
544 regNumber GetOtherArgReg() const
546 return (regNumber)_lvOtherArgReg;
549 void SetOtherArgReg(regNumber reg)
551 _lvOtherArgReg = (regNumberSmall)reg;
552 assert(_lvOtherArgReg == reg);
554 #endif // FEATURE_MULTIREG_ARGS
557 // Is this is a SIMD struct?
558 bool lvIsSIMDType() const
563 // Is this is a SIMD struct which is used for SIMD intrinsic?
564 bool lvIsUsedInSIMDIntrinsic() const
566 return lvUsedInSIMDIntrinsic;
569 // If feature_simd not enabled, return false
570 bool lvIsSIMDType() const
574 bool lvIsUsedInSIMDIntrinsic() const
580 /////////////////////
582 #ifndef LEGACY_BACKEND
583 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
585 regNumber GetArgInitReg() const
587 return (regNumber)_lvArgInitReg;
590 void SetArgInitReg(regNumber reg)
592 _lvArgInitReg = (regNumberSmall)reg;
593 assert(_lvArgInitReg == reg);
596 /////////////////////
598 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
600 regPairNo GetArgInitRegPair() const
602 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
603 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
607 void SetArgInitRegPair(regPairNo regPair)
609 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
610 _lvArgInitRegPair = (regPairNoSmall)regPair;
611 assert(_lvArgInitRegPair == regPair);
614 /////////////////////
616 bool lvIsRegCandidate() const
618 return lvLRACandidate != 0;
621 bool lvIsInReg() const
623 return lvIsRegCandidate() && (lvRegNum != REG_STK);
626 #else // LEGACY_BACKEND
628 bool lvIsRegCandidate() const
630 return lvTracked != 0;
633 bool lvIsInReg() const
635 return lvRegister != 0;
638 #endif // LEGACY_BACKEND
640 regMaskTP lvRegMask() const
642 regMaskTP regMask = RBM_NONE;
643 if (varTypeIsFloating(TypeGet()))
645 if (lvRegNum != REG_STK)
647 regMask = genRegMaskFloat(lvRegNum, TypeGet());
652 if (lvRegNum != REG_STK)
654 regMask = genRegMask(lvRegNum);
657 // For longs we may have two regs
658 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
660 regMask |= genRegMask(lvOtherReg);
666 regMaskSmall lvPrefReg; // set of regs it prefers to live in
668 unsigned short lvVarIndex; // variable tracking index
669 unsigned short lvRefCnt; // unweighted (real) reference count. For implicit by reference
670 // parameters, this gets hijacked from fgMarkImplicitByRefArgs
671 // through fgMarkDemotedImplicitByRefArgs, to provide a static
672 // appearance count (computed during address-exposed analysis)
673 // that fgMakeOutgoingStructArgCopy consults during global morph
674 // to determine if eliding its copy is legal.
675 unsigned lvRefCntWtd; // weighted reference count
676 int lvStkOffs; // stack offset of home
677 unsigned lvExactSize; // (exact) size of the type in bytes
679 // Is this a promoted struct?
680 // This method returns true only for structs (including SIMD structs), not for
681 // locals that are split on a 32-bit target.
682 // It is only necessary to use this:
683 // 1) if only structs are wanted, and
684 // 2) if Lowering has already been done.
685 // Otherwise lvPromoted is valid.
686 bool lvPromotedStruct()
688 #if !defined(_TARGET_64BIT_)
689 return (lvPromoted && !varTypeIsLong(lvType));
690 #else // defined(_TARGET_64BIT_)
692 #endif // defined(_TARGET_64BIT_)
695 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
697 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
698 // where the struct itself is no longer used because all access is via its member fields.
699 // When that happens, the struct is marked as unused and its type has been changed to
700 // TYP_INT (to keep the GC tracking code from looking at it).
701 // See Compiler::raAssignVars() for details. For example:
702 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
703 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
704 // float V03.f1 (offs=0x00) -> V12 tmp7
705 // f8 (last use) (last use) $345
706 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
707 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
709 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
711 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
712 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
713 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
714 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
715 if ((lvType == TYP_SIMD12) && !lvIsParam)
717 assert(lvExactSize == 12);
720 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
722 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
725 const size_t lvArgStackSize() const;
727 unsigned lvSlotNum; // original slot # (if remapped)
729 typeInfo lvVerTypeInfo; // type info needed for verification
731 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
733 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
735 BYTE* lvGcLayout; // GC layout info for structs
738 BlockSet lvRefBlks; // Set of blocks that contain refs
739 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
740 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
742 var_types TypeGet() const
744 return (var_types)lvType;
746 bool lvStackAligned() const
748 assert(lvIsStructField);
749 return ((lvFldOffset % sizeof(void*)) == 0);
751 bool lvNormalizeOnLoad() const
753 return varTypeIsSmall(TypeGet()) &&
754 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
755 (lvIsParam || lvAddrExposed || lvIsStructField);
758 bool lvNormalizeOnStore()
760 return varTypeIsSmall(TypeGet()) &&
761 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
762 !(lvIsParam || lvAddrExposed || lvIsStructField);
765 void lvaResetSortAgainFlag(Compiler* pComp);
766 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
767 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
768 void setPrefReg(regNumber regNum, Compiler* pComp);
769 void addPrefReg(regMaskTP regMask, Compiler* pComp);
770 bool IsFloatRegType() const
772 return isFloatRegType(lvType) || lvIsHfaRegArg();
774 var_types GetHfaType() const
776 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
778 void SetHfaType(var_types type)
780 assert(varTypeIsFloating(type));
781 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
784 #ifndef LEGACY_BACKEND
785 var_types lvaArgType();
788 PerSsaArray lvPerSsaData;
791 // Keep track of the # of SsaNames, for a bounds check.
792 unsigned lvNumSsaNames;
795 // Returns the address of the per-Ssa data for the given ssaNum (which is required
796 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
797 // not an SSA variable).
798 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
800 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
801 assert(SsaConfig::RESERVED_SSA_NUM == 0);
802 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
803 assert(zeroBased < lvNumSsaNames);
804 return &lvPerSsaData.GetRef(zeroBased);
809 void PrintVarReg() const
811 if (isRegPairType(TypeGet()))
813 printf("%s:%s", getRegName(lvOtherReg), // hi32
814 getRegName(lvRegNum)); // lo32
818 printf("%s", getRegName(lvRegNum));
823 }; // class LclVarDsc
826 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
827 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
831 XX The temporary lclVars allocated by the compiler for code generation XX
833 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
834 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
837 /*****************************************************************************
839 * The following keeps track of temporaries allocated in the stack frame
840 * during code-generation (after register allocation). These spill-temps are
841 * only used if we run out of registers while evaluating a tree.
843 * These are different from the more common temps allocated by lvaGrabTemp().
854 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
862 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
866 0); // temps must have a negative number (so they have a different number from all local variables)
867 tdOffs = BAD_TEMP_OFFSET;
871 IMPL_LIMITATION("too many spill temps");
876 bool tdLegalOffset() const
878 return tdOffs != BAD_TEMP_OFFSET;
882 int tdTempOffs() const
884 assert(tdLegalOffset());
887 void tdSetTempOffs(int offs)
890 assert(tdLegalOffset());
892 void tdAdjustTempOffs(int offs)
895 assert(tdLegalOffset());
898 int tdTempNum() const
903 unsigned tdTempSize() const
907 var_types tdTempType() const
913 // interface to hide linearscan implementation from rest of compiler
914 class LinearScanInterface
917 virtual void doLinearScan() = 0;
918 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
919 virtual bool willEnregisterLocalVars() const = 0;
922 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
924 // Information about arrays: their element type and size, and the offset of the first element.
925 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
926 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
927 // for example, in value numbering of array index expressions.
930 var_types m_elemType;
931 CORINFO_CLASS_HANDLE m_elemStructType;
933 unsigned m_elemOffset;
935 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
939 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
940 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
945 // This enumeration names the phases into which we divide compilation. The phases should completely
946 // partition a compilation.
949 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
950 #include "compphases.h"
954 extern const char* PhaseNames[];
955 extern const char* PhaseEnums[];
956 extern const LPCWSTR PhaseShortNames[];
958 // The following enum provides a simple 1:1 mapping to CLR API's
959 enum API_ICorJitInfo_Names
961 #define DEF_CLR_API(name) API_##name,
962 #include "ICorJitInfo_API_names.h"
966 //---------------------------------------------------------------
970 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
971 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
972 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
973 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
974 // by "m_timerFailure" being true.
975 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
978 #ifdef FEATURE_JIT_METHOD_PERF
979 // The string names of the phases.
980 static const char* PhaseNames[];
982 static bool PhaseHasChildren[];
983 static int PhaseParent[];
984 static bool PhaseReportsIRSize[];
986 unsigned m_byteCodeBytes;
987 unsigned __int64 m_totalCycles;
988 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
989 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
990 #if MEASURE_CLRAPI_CALLS
991 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
992 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
995 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
997 // For better documentation, we call EndPhase on
998 // non-leaf phases. We should also call EndPhase on the
999 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
1000 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
1001 // We add all such "redundant end phase" intervals to this variable below; we print
1002 // it out in a report, so we can verify that it is, indeed, very small. If it ever
1003 // isn't, this means that we're doing something significant between the end of the last
1004 // declared subphase and the end of its parent.
1005 unsigned __int64 m_parentPhaseEndSlop;
1006 bool m_timerFailure;
1008 #if MEASURE_CLRAPI_CALLS
1009 // The following measures the time spent inside each individual CLR API call.
1010 unsigned m_allClrAPIcalls;
1011 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1012 unsigned __int64 m_allClrAPIcycles;
1013 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1014 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1015 #endif // MEASURE_CLRAPI_CALLS
1017 CompTimeInfo(unsigned byteCodeBytes);
1021 #ifdef FEATURE_JIT_METHOD_PERF
1023 #if MEASURE_CLRAPI_CALLS
1024 struct WrapICorJitInfo;
1027 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1028 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1029 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1030 // threads, so it is protected by a lock.
1031 // This class is intended to be used as a singleton type, with only a single instance.
1032 class CompTimeSummaryInfo
1034 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1035 static CritSecObject s_compTimeSummaryLock;
1039 CompTimeInfo m_total;
1040 CompTimeInfo m_maximum;
1042 int m_numFilteredMethods;
1043 CompTimeInfo m_filtered;
1045 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1046 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1047 // If any OS interaction fails, returns 0.0.
1048 double CyclesPerSecond();
1050 // This can use what ever data you want to determine if the value to be added
1051 // belongs in the filtered section (it's always included in the unfiltered section)
1052 bool IncludedInFilteredData(CompTimeInfo& info);
1055 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1056 static CompTimeSummaryInfo s_compTimeSummary;
1058 CompTimeSummaryInfo()
1059 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1063 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1064 // This is thread safe.
1065 void AddInfo(CompTimeInfo& info, bool includePhases);
1067 // Print the summary information to "f".
1068 // This is not thread-safe; assumed to be called by only one thread.
1069 void Print(FILE* f);
1072 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1073 // and when the current phase started. This is intended to be part of a Compilation object. This is
1074 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1078 unsigned __int64 m_start; // Start of the compilation.
1079 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1080 #if MEASURE_CLRAPI_CALLS
1081 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1082 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1083 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1084 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1085 static double s_cyclesPerSec; // Cached for speedier measurements
1088 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1090 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1092 static CritSecObject s_csvLock; // Lock to protect the time log file.
1093 void PrintCsvMethodStats(Compiler* comp);
1096 void* operator new(size_t);
1097 void* operator new[](size_t);
1098 void operator delete(void*);
1099 void operator delete[](void*);
1102 // Initialized the timer instance
1103 JitTimer(unsigned byteCodeSize);
1105 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1107 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1110 static void PrintCsvHeader();
1112 // Ends the current phase (argument is for a redundant check).
1113 void EndPhase(Compiler* compiler, Phases phase);
1115 #if MEASURE_CLRAPI_CALLS
1116 // Start and end a timed CLR API call.
1117 void CLRApiCallEnter(unsigned apix);
1118 void CLRApiCallLeave(unsigned apix);
1119 #endif // MEASURE_CLRAPI_CALLS
1121 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1122 // and adds it to "sum".
1123 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1125 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1126 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1127 // "m_info" to true.
1128 bool GetThreadCycles(unsigned __int64* cycles)
1130 bool res = CycleTimer::GetThreadCyclesS(cycles);
1133 m_info.m_timerFailure = true;
1138 #endif // FEATURE_JIT_METHOD_PERF
1140 //------------------- Function/Funclet info -------------------------------
1141 enum FuncKind : BYTE
1143 FUNC_ROOT, // The main/root function (always id==0)
1144 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1145 FUNC_FILTER, // a funclet associated with an EH filter
1154 BYTE funFlags; // Currently unused, just here for padding
1155 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1156 // funclet. It is only valid if funKind field indicates this is a
1157 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1159 #if defined(_TARGET_AMD64_)
1161 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1162 emitLocation* startLoc;
1163 emitLocation* endLoc;
1164 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1165 emitLocation* coldEndLoc;
1166 UNWIND_INFO unwindHeader;
1167 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1168 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1169 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1170 unsigned unwindCodeSlot;
1172 #ifdef UNIX_AMD64_ABI
1173 jitstd::vector<CFI_CODE>* cfiCodes;
1174 #endif // UNIX_AMD64_ABI
1176 #elif defined(_TARGET_ARMARCH_)
1178 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1179 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1180 // Note: we only have a pointer here instead of the actual object,
1181 // to save memory in the JIT case (compared to the NGEN case),
1182 // where we don't have any cold section.
1183 // Note 2: we currently don't support hot/cold splitting in functions
1184 // with EH, so uwiCold will be NULL for all funclets.
1186 #endif // _TARGET_ARMARCH_
1188 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1189 // that isn't shared between the main function body and funclets.
1192 struct fgArgTabEntry
1195 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1198 otherRegNum = REG_NA;
1199 isStruct = false; // is this a struct arg
1201 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1203 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1205 // it will point at the actual argument in the gtCallLateArgs list.
1206 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1208 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1210 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1212 unsigned numRegs; // Count of number of registers that this argument uses
1214 // A slot is a pointer sized region in the OutArg area.
1215 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1216 unsigned numSlots; // Count of number of slots that this argument uses
1218 unsigned alignment; // 1 or 2 (slots/registers)
1219 unsigned lateArgInx; // index into gtCallLateArgs list
1220 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1222 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1223 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1224 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1225 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1226 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1227 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1228 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1229 // previous arguments.
1230 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1231 // to be on the stack despite its arg list position.
1233 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1234 bool isStruct : 1; // True if this is a struct arg
1236 regNumber otherRegNum; // The (second) register to use when passing this argument.
1238 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1239 #elif !defined(_TARGET_64BIT_)
1240 __declspec(property(get = getIsStruct)) bool isStruct;
1243 return varTypeIsStruct(node);
1245 #endif // !_TARGET_64BIT_
1248 void SetIsHfaRegArg(bool hfaRegArg)
1250 isHfaRegArg = hfaRegArg;
1253 void SetIsBackFilled(bool backFilled)
1255 isBackFilled = backFilled;
1258 bool IsBackFilled() const
1260 return isBackFilled;
1262 #else // !_TARGET_ARM_
1263 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1265 void SetIsHfaRegArg(bool hfaRegArg)
1269 void SetIsBackFilled(bool backFilled)
1273 bool IsBackFilled() const
1277 #endif // !_TARGET_ARM_
1283 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1285 //-------------------------------------------------------------------------
1287 // The class fgArgInfo is used to handle the arguments
1288 // when morphing a GT_CALL node.
1293 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1294 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1295 unsigned argCount; // Updatable arg count value
1296 unsigned nextSlotNum; // Updatable slot count value
1297 unsigned stkLevel; // Stack depth when we make this call (for x86)
1299 #if defined(UNIX_X86_ABI)
1300 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1301 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1302 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1303 // Computed dynamically during codegen, based on stkSizeBytes and the current
1304 // stack level (genStackLevel) when the first stack adjustment is made for
1308 #if FEATURE_FIXED_OUT_ARGS
1309 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1312 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1313 bool hasRegArgs; // true if we have one or more register arguments
1314 bool hasStackArgs; // true if we have one or more stack arguments
1315 bool argsComplete; // marker for state
1316 bool argsSorted; // marker for state
1317 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1320 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1323 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1324 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1326 fgArgTabEntryPtr AddRegArg(
1327 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1329 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1330 fgArgTabEntryPtr AddRegArg(
1337 const bool isStruct,
1338 const regNumber otherRegNum = REG_NA,
1339 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1340 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1342 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1346 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1348 void RemorphReset();
1349 fgArgTabEntryPtr RemorphRegArg(
1350 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1352 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1354 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1356 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1358 void ArgsComplete();
1362 void EvalArgsToTemps();
1364 void RecordStkLevel(unsigned stkLvl);
1365 unsigned RetrieveStkLevel();
1371 fgArgTabEntryPtr* ArgTable()
1375 unsigned GetNextSlotNum()
1385 return hasStackArgs;
1387 bool AreArgsComplete() const
1389 return argsComplete;
1391 #if FEATURE_FIXED_OUT_ARGS
1392 unsigned GetOutArgSize() const
1396 void SetOutArgSize(unsigned newVal)
1398 outArgSize = newVal;
1400 #endif // FEATURE_FIXED_OUT_ARGS
1402 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1404 #if defined(UNIX_X86_ABI)
1405 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1406 #endif // defined(UNIX_X86_ABI)
1409 void SetStkSizeBytes(unsigned newStkSizeBytes)
1411 #if defined(UNIX_X86_ABI)
1412 stkSizeBytes = newStkSizeBytes;
1413 #endif // defined(UNIX_X86_ABI)
1416 #if defined(UNIX_X86_ABI)
1417 unsigned GetStkAlign()
1421 unsigned GetStkSizeBytes() const
1423 return stkSizeBytes;
1425 bool IsStkAlignmentDone() const
1427 return alignmentDone;
1429 void SetStkAlignmentDone()
1431 alignmentDone = true;
1433 #endif // defined(UNIX_X86_ABI)
1435 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1436 GenTreePtr GetLateArg(unsigned argIndex);
1440 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1441 // We have the ability to mark source expressions with "Test Labels."
1442 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1443 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1445 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1448 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1449 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1450 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1451 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1452 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1455 struct TestLabelAndNum
1460 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1465 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1467 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1470 // This class implements the "IAllocator" interface, so that we can use
1471 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1473 class CompAllocator : public IAllocator
1476 #if MEASURE_MEM_ALLOC
1480 CompAllocator(Compiler* comp, CompMemKind cmk)
1482 #if MEASURE_MEM_ALLOC
1488 inline void* Alloc(size_t sz);
1490 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1492 // For the compiler's no-release allocator, free operations are no-ops.
1499 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1500 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1502 XX The big guy. The sections are currently organized as : XX
1504 XX o GenTree and BasicBlock XX
1516 XX o PrologScopeInfo XX
1517 XX o CodeGenerator XX
1522 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1523 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1528 friend class emitter;
1529 friend class UnwindInfo;
1530 friend class UnwindFragmentInfo;
1531 friend class UnwindEpilogInfo;
1532 friend class JitTimer;
1533 friend class LinearScan;
1534 friend class fgArgInfo;
1535 friend class Rationalizer;
1537 friend class Lowering;
1538 friend class CSE_DataFlow;
1539 friend class CSE_Heuristic;
1540 friend class CodeGenInterface;
1541 friend class CodeGen;
1542 friend class LclVarDsc;
1543 friend class TempDsc;
1545 friend class ObjectAllocator;
1547 #ifndef _TARGET_64BIT_
1548 friend class DecomposeLongs;
1549 #endif // !_TARGET_64BIT_
1552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1553 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1555 XX Misc structs definitions XX
1557 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1558 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1562 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1581 bool dumpIRDataflow;
1582 bool dumpIRBlockHeaders;
1584 LPCWSTR dumpIRPhase;
1585 LPCWSTR dumpIRFormat;
1587 bool shouldUseVerboseTrees();
1588 bool asciiTrees; // If true, dump trees using only ASCII characters
1589 bool shouldDumpASCIITrees();
1590 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1591 bool shouldUseVerboseSsa();
1592 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1593 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1595 const char* VarNameToStr(VarName name)
1600 DWORD expensiveDebugCheckLevel;
1603 #if FEATURE_MULTIREG_RET
1604 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1605 #endif // FEATURE_MULTIREG_RET
1608 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1609 #endif // ARM_SOFTFP
1611 //-------------------------------------------------------------------------
1612 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1613 // HFAs are one to four element structs where each element is the same
1614 // type, either all float or all double. They are treated specially
1615 // in the ARM Procedure Call Standard, specifically, they are passed in
1616 // floating-point registers instead of the general purpose registers.
1619 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1620 bool IsHfa(GenTreePtr tree);
1622 var_types GetHfaType(GenTreePtr tree);
1623 unsigned GetHfaCount(GenTreePtr tree);
1625 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1626 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1628 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1629 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1631 //-------------------------------------------------------------------------
1632 // The following is used for validating format of EH table
1636 typedef struct EHNodeDsc* pEHNodeDsc;
1638 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1639 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1652 EHBlockType ehnBlockType; // kind of EH block
1653 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1654 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1655 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1657 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1658 pEHNodeDsc ehnChild; // leftmost nested block
1660 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1661 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1663 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1664 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1666 inline void ehnSetTryNodeType()
1668 ehnBlockType = TryNode;
1670 inline void ehnSetFilterNodeType()
1672 ehnBlockType = FilterNode;
1674 inline void ehnSetHandlerNodeType()
1676 ehnBlockType = HandlerNode;
1678 inline void ehnSetFinallyNodeType()
1680 ehnBlockType = FinallyNode;
1682 inline void ehnSetFaultNodeType()
1684 ehnBlockType = FaultNode;
1687 inline BOOL ehnIsTryBlock()
1689 return ehnBlockType == TryNode;
1691 inline BOOL ehnIsFilterBlock()
1693 return ehnBlockType == FilterNode;
1695 inline BOOL ehnIsHandlerBlock()
1697 return ehnBlockType == HandlerNode;
1699 inline BOOL ehnIsFinallyBlock()
1701 return ehnBlockType == FinallyNode;
1703 inline BOOL ehnIsFaultBlock()
1705 return ehnBlockType == FaultNode;
1708 // returns true if there is any overlap between the two nodes
1709 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1711 if (node1->ehnStartOffset < node2->ehnStartOffset)
1713 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1717 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1721 // fails with BADCODE if inner is not completely nested inside outer
1722 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1724 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1728 //-------------------------------------------------------------------------
1729 // Exception handling functions
1732 #if !FEATURE_EH_FUNCLETS
1734 bool ehNeedsShadowSPslots()
1736 return (info.compXcptnsCount || opts.compDbgEnC);
1739 // 0 for methods with no EH
1740 // 1 for methods with non-nested EH, or where only the try blocks are nested
1741 // 2 for a method with a catch within a catch
1743 unsigned ehMaxHndNestingCount;
1745 #endif // !FEATURE_EH_FUNCLETS
1747 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1748 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1750 bool bbInCatchHandlerILRange(BasicBlock* blk);
1751 bool bbInFilterILRange(BasicBlock* blk);
1752 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1753 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1754 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1755 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1756 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1758 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1759 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1761 // Returns true if "block" is the start of a try region.
1762 bool bbIsTryBeg(BasicBlock* block);
1764 // Returns true if "block" is the start of a handler or filter region.
1765 bool bbIsHandlerBeg(BasicBlock* block);
1767 // Returns true iff "block" is where control flows if an exception is raised in the
1768 // try region, and sets "*regionIndex" to the index of the try for the handler.
1769 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1770 // block of the filter, but not for the filter's handler.
1771 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1773 bool ehHasCallableHandlers();
1775 // Return the EH descriptor for the given region index.
1776 EHblkDsc* ehGetDsc(unsigned regionIndex);
1778 // Return the EH index given a region descriptor.
1779 unsigned ehGetIndex(EHblkDsc* ehDsc);
1781 // Return the EH descriptor index of the enclosing try, for the given region index.
1782 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1784 // Return the EH descriptor index of the enclosing handler, for the given region index.
1785 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1787 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1788 // block is not in a 'try' region).
1789 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1791 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1792 // if this block is not in a filter or handler region).
1793 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1795 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1796 // nullptr if this block's exceptions propagate to caller).
1797 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1799 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1800 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1801 bool ehIsBlockEHLast(BasicBlock* block);
1803 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1805 // Return the region index of the most nested EH region this block is in.
1806 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1808 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1809 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1811 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1812 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1813 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1814 // (It can never be a filter.)
1815 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1817 // A block has been deleted. Update the EH table appropriately.
1818 void ehUpdateForDeletedBlock(BasicBlock* block);
1820 // Determine whether a block can be deleted while preserving the EH normalization rules.
1821 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1823 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1824 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1826 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1827 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1828 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1829 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1830 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1831 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1832 // lives in a filter.)
1833 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1835 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1836 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1837 // (nullptr if the last block is the last block in the program).
1838 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1839 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1842 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1843 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1844 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1847 #if FEATURE_EH_FUNCLETS
1848 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1849 // if there is a filter that protects a region with a nested EH clause (such as a
1850 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1851 // genFuncletProlog() for more details. However, the VM seems to use it for more
1852 // purposes, maybe including debugging. Until we are sure otherwise, always create
1853 // a PSPSym for functions with any EH.
1854 bool ehNeedsPSPSym() const
1858 #else // _TARGET_X86_
1859 return compHndBBtabCount > 0;
1860 #endif // _TARGET_X86_
1863 bool ehAnyFunclets(); // Are there any funclets in this function?
1864 unsigned ehFuncletCount(); // Return the count of funclets in the function
1866 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1867 #else // !FEATURE_EH_FUNCLETS
1868 bool ehAnyFunclets()
1872 unsigned ehFuncletCount()
1877 unsigned bbThrowIndex(BasicBlock* blk)
1879 return blk->bbTryIndex;
1880 } // Get the index to use as the cache key for sharing throw blocks
1881 #endif // !FEATURE_EH_FUNCLETS
1883 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1884 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1885 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1886 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1887 // convenient to also consider it a predecessor.)
1888 flowList* BlockPredsWithEH(BasicBlock* blk);
1890 // This table is useful for memoization of the method above.
1891 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1893 BlockToFlowListMap* m_blockToEHPreds;
1894 BlockToFlowListMap* GetBlockToEHPreds()
1896 if (m_blockToEHPreds == nullptr)
1898 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1900 return m_blockToEHPreds;
1903 void* ehEmitCookie(BasicBlock* block);
1904 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1906 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1908 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1910 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1912 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1914 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1916 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1918 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1920 void fgAllocEHTable();
1922 void fgRemoveEHTableEntry(unsigned XTnum);
1924 #if FEATURE_EH_FUNCLETS
1926 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1928 #endif // FEATURE_EH_FUNCLETS
1932 #endif // !FEATURE_EH
1934 void fgSortEHTable();
1936 // Causes the EH table to obey some well-formedness conditions, by inserting
1937 // empty BB's when necessary:
1938 // * No block is both the first block of a handler and the first block of a try.
1939 // * No block is the first block of multiple 'try' regions.
1940 // * No block is the last block of multiple EH regions.
1941 void fgNormalizeEH();
1942 bool fgNormalizeEHCase1();
1943 bool fgNormalizeEHCase2();
1944 bool fgNormalizeEHCase3();
1947 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1948 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1949 void fgVerifyHandlerTab();
1950 void fgDispHandlerTab();
1953 bool fgNeedToSortEHTable;
1955 void verInitEHTree(unsigned numEHClauses);
1956 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1957 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1958 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1959 void verCheckNestingLevel(EHNodeDsc* initRoot);
1962 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1963 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1965 XX GenTree and BasicBlock XX
1967 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1969 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1970 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1973 // Functions to create nodes
1974 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1977 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1979 // For binary opers.
1980 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1982 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1984 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1985 var_types type = TYP_I_IMPL,
1986 GenTreePtr op1 = nullptr,
1987 GenTreePtr op2 = nullptr);
1989 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1991 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1993 GenTreePtr gtNewJmpTableNode();
1994 GenTreePtr gtNewIconHandleNode(
1995 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1997 unsigned gtTokenToIconFlags(unsigned token);
1999 GenTreePtr gtNewIconEmbHndNode(void* value,
2002 unsigned handle1 = 0,
2003 void* handle2 = nullptr,
2004 void* compileTimeHandle = nullptr);
2006 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2007 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2008 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2009 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2011 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2013 GenTreePtr gtNewLconNode(__int64 value);
2015 GenTreePtr gtNewDconNode(double value);
2017 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2019 GenTreePtr gtNewZeroConNode(var_types type);
2021 GenTreePtr gtNewOneConNode(var_types type);
2024 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2025 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2028 GenTreeBlk* gtNewBlkOpNode(
2029 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
2031 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2033 GenTree* gtNewPutArgReg(var_types type, GenTreePtr arg, regNumber argReg);
2036 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
2039 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2040 void gtSetObjGcInfo(GenTreeObj* objNode);
2041 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2042 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
2044 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2046 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
2048 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2049 CORINFO_METHOD_HANDLE handle,
2051 GenTreeArgList* args,
2052 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2054 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
2056 GenTreeArgList* args,
2057 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2059 GenTreeCall* gtNewHelperCallNode(unsigned helper,
2062 GenTreeArgList* args = nullptr);
2064 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2067 GenTreeSIMD* gtNewSIMDNode(
2068 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2069 GenTreeSIMD* gtNewSIMDNode(var_types type,
2072 SIMDIntrinsicID simdIntrinsicID,
2075 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2078 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2079 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2080 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2082 GenTreePtr gtNewCodeRef(BasicBlock* block);
2084 GenTreePtr gtNewFieldRef(
2085 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2087 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2089 GenTreeArgList* gtNewArgList(GenTreePtr op);
2090 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2091 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2093 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2094 static fgArgTabEntryPtr gtArgEntryByNode(GenTreeCall* call, GenTreePtr node);
2095 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2096 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2098 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2100 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2102 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2103 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2104 CORINFO_ACCESS_FLAGS access,
2105 CORINFO_FIELD_INFO* pFieldInfo,
2107 CORINFO_CLASS_HANDLE structType,
2110 GenTreePtr gtNewNothingNode();
2112 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2114 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2116 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2118 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2120 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2122 //------------------------------------------------------------------------
2123 // Other GenTree functions
2125 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2127 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2128 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2129 // IntCnses with value `deepVarVal`.
2130 GenTreePtr gtCloneExpr(
2131 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2133 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2134 // `varNum` to int constants with value `varVal`.
2135 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2137 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2140 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2142 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2144 // Returns "true" iff the complexity (not formally defined, but first interpretation
2145 // is #of nodes in subtree) of "tree" is greater than "limit".
2146 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2147 // before they have been set.)
2148 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2150 bool gtCompareTree(GenTree* op1, GenTree* op2);
2152 GenTreePtr gtReverseCond(GenTree* tree);
2154 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2156 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2158 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2160 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2163 unsigned gtHashValue(GenTree* tree);
2165 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2168 void gtPrepareCost(GenTree* tree);
2169 bool gtIsLikelyRegVar(GenTree* tree);
2171 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2173 // Returns true iff the secondNode can be swapped with firstNode.
2174 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2176 unsigned gtSetEvalOrder(GenTree* tree);
2178 #if FEATURE_STACK_FP_X87
2180 void gtComputeFPlvls(GenTreePtr tree);
2181 #endif // FEATURE_STACK_FP_X87
2183 void gtSetStmtInfo(GenTree* stmt);
2185 // Returns "true" iff "node" has any of the side effects in "flags".
2186 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2188 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2189 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2191 // Appends 'expr' in front of 'list'
2192 // 'list' will typically start off as 'nullptr'
2193 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2194 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2196 void gtExtractSideEffList(GenTreePtr expr,
2198 unsigned flags = GTF_SIDE_EFFECT,
2199 bool ignoreRoot = false);
2201 GenTreePtr gtGetThisArg(GenTreeCall* call);
2203 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2204 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2205 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2206 // the given "fldHnd", is such an object pointer.
2207 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2209 // Return true if call is a recursive call; return false otherwise.
2210 // Note when inlining, this looks for calls back to the root method.
2211 bool gtIsRecursiveCall(GenTreeCall* call)
2213 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2216 //-------------------------------------------------------------------------
2218 GenTreePtr gtFoldExpr(GenTreePtr tree);
2221 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2222 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2223 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2224 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2225 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2226 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2227 // optimizations for now.
2228 __attribute__((optnone))
2230 gtFoldExprConst(GenTreePtr tree);
2231 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2232 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2233 bool gtTryRemoveBoxUpstreamEffects(GenTreePtr tree);
2235 //-------------------------------------------------------------------------
2236 // Get the handle, if any.
2237 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2238 // Get the handle, and assert if not found.
2239 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2240 // Get the handle for a ref type.
2241 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2243 //-------------------------------------------------------------------------
2244 // Functions to display the trees
2247 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2249 void gtDispVN(GenTreePtr tree);
2250 void gtDispConst(GenTreePtr tree);
2251 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2252 void gtDispNodeName(GenTreePtr tree);
2253 void gtDispRegVal(GenTreePtr tree);
2265 void gtDispChild(GenTreePtr child,
2266 IndentStack* indentStack,
2268 __in_opt const char* msg = nullptr,
2269 bool topOnly = false);
2270 void gtDispTree(GenTreePtr tree,
2271 IndentStack* indentStack = nullptr,
2272 __in_opt const char* msg = nullptr,
2273 bool topOnly = false,
2274 bool isLIR = false);
2275 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2276 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2277 char* gtGetLclVarName(unsigned lclNum);
2278 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2279 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2280 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2281 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2282 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2283 void gtDispFieldSeq(FieldSeqNode* pfsn);
2285 void gtDispRange(LIR::ReadOnlyRange const& range);
2287 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2289 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2301 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2302 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2305 static fgWalkPreFn gtAssertColonCond;
2307 static fgWalkPreFn gtMarkColonCond;
2308 static fgWalkPreFn gtClearColonCond;
2310 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2311 bool gtHasCatchArg(GenTreePtr tree);
2312 bool gtHasUnmanagedCall(GenTreePtr tree);
2314 typedef ArrayStack<GenTree*> GenTreeStack;
2316 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2317 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2319 //=========================================================================
2320 // BasicBlock functions
2322 // This is a debug flag we will use to assert when creating block during codegen
2323 // as this interferes with procedure splitting. If you know what you're doing, set
2324 // it to true before creating the block. (DEBUG only)
2325 bool fgSafeBasicBlockCreation;
2328 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2331 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2332 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2336 XX The variables to be used by the code generator. XX
2338 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2339 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2343 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2344 // be placed in the stack frame and it's fields must be laid out sequentially.
2346 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2347 // a local variable that can be enregistered or placed in the stack frame.
2348 // The fields do not need to be laid out sequentially
2350 enum lvaPromotionType
2352 PROMOTION_TYPE_NONE, // The struct local is not promoted
2353 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2354 // and its field locals are independent of its parent struct local.
2355 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2356 // but its field locals depend on its parent struct local.
2359 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2360 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2362 /*****************************************************************************/
2364 enum FrameLayoutState
2367 INITIAL_FRAME_LAYOUT,
2368 PRE_REGALLOC_FRAME_LAYOUT,
2369 REGALLOC_FRAME_LAYOUT,
2370 TENTATIVE_FRAME_LAYOUT,
2375 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2376 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2377 bool lvaSortAgain; // true: We need to sort the lvaTable
2378 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2379 unsigned lvaCount; // total number of locals
2381 unsigned lvaRefCount; // total number of references to locals
2382 LclVarDsc* lvaTable; // variable descriptor table
2383 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2385 LclVarDsc** lvaRefSorted; // table sorted by refcount
2387 unsigned short lvaTrackedCount; // actual # of locals being tracked
2388 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2390 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2391 // Only for AMD64 System V cache the first caller stack homed argument.
2392 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2393 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2396 VARSET_TP lvaTrackedVars; // set of tracked variables
2398 #ifndef _TARGET_64BIT_
2399 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2401 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2403 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2404 // It that changes, this changes. VarSets from different epochs
2405 // cannot be meaningfully combined.
2407 unsigned GetCurLVEpoch()
2412 // reverse map of tracked number to var number
2413 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2415 #ifdef LEGACY_BACKEND
2416 // variable interference graph
2417 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2420 // variable preference graph
2421 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2425 // # of procs compiled a with double-aligned stack
2426 static unsigned s_lvaDoubleAlignedProcsCount;
2430 // Getters and setters for address-exposed and do-not-enregister local var properties.
2431 bool lvaVarAddrExposed(unsigned varNum);
2432 void lvaSetVarAddrExposed(unsigned varNum);
2433 bool lvaVarDoNotEnregister(unsigned varNum);
2435 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2436 enum DoNotEnregisterReason
2441 DNER_VMNeedsStackAddr,
2442 DNER_LiveInOutOfHandler,
2443 DNER_LiveAcrossUnmanagedCall,
2444 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2445 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2446 DNER_DepField, // It is a field of a dependently promoted struct
2447 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2448 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2449 #if !defined(LEGACY_BACKEND) && !defined(_TARGET_64BIT_)
2450 DNER_LongParamField, // It is a decomposed field of a long parameter.
2452 #ifdef JIT32_GCENCODER
2457 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2459 unsigned lvaVarargsHandleArg;
2461 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2463 #endif // _TARGET_X86_
2465 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2466 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2467 #if FEATURE_FIXED_OUT_ARGS
2468 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2470 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2471 // that tracks whether the lock has been taken
2473 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2474 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2475 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2477 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2478 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2480 #if FEATURE_FIXED_OUT_ARGS
2481 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2482 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2483 #endif // FEATURE_FIXED_OUT_ARGS
2486 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2487 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2488 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2489 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2490 // this variable to be this scratch word whenever struct promotion occurs.
2491 unsigned lvaPromotedStructAssemblyScratchVar;
2492 #endif // _TARGET_ARM_
2495 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2496 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2499 unsigned lvaGenericsContextUseCount;
2501 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2502 // CORINFO_GENERICS_CTXT_FROM_THIS?
2503 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2505 //-------------------------------------------------------------------------
2506 // All these frame offsets are inter-related and must be kept in sync
2508 #if !FEATURE_EH_FUNCLETS
2509 // This is used for the callable handlers
2510 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2511 #endif // FEATURE_EH_FUNCLETS
2513 unsigned lvaCachedGenericContextArgOffs;
2514 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2517 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2519 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2521 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2522 // after the reg predict we will use a computed maxTmpSize
2523 // which is based upon the number of spill temps predicted by reg predict
2524 // All this is necessary because if we under-estimate the size of the spill
2525 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2527 // Pre codegen max spill temp size.
2528 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2530 //-------------------------------------------------------------------------
2532 unsigned lvaGetMaxSpillTempSize();
2534 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2535 #endif // _TARGET_ARM_
2536 void lvaAssignFrameOffsets(FrameLayoutState curState);
2537 void lvaFixVirtualFrameOffsets();
2539 #ifndef LEGACY_BACKEND
2540 void lvaUpdateArgsWithInitialReg();
2541 #endif // !LEGACY_BACKEND
2543 void lvaAssignVirtualFrameOffsetsToArgs();
2544 #ifdef UNIX_AMD64_ABI
2545 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2546 #else // !UNIX_AMD64_ABI
2547 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2548 #endif // !UNIX_AMD64_ABI
2549 void lvaAssignVirtualFrameOffsetsToLocals();
2550 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2551 #ifdef _TARGET_AMD64_
2552 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2553 bool lvaIsCalleeSavedIntRegCountEven();
2555 void lvaAlignFrame();
2556 void lvaAssignFrameOffsetsToPromotedStructs();
2557 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2560 void lvaDumpRegLocation(unsigned lclNum);
2561 void lvaDumpFrameLocation(unsigned lclNum);
2562 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2563 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2564 // layout state defined by lvaDoneFrameLayout
2567 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2568 // to avoid bugs from borderline cases.
2569 #define MAX_FrameSize 0x3FFFFFFF
2570 void lvaIncrementFrameSize(unsigned size);
2572 unsigned lvaFrameSize(FrameLayoutState curState);
2574 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2575 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2577 // Returns the caller-SP-relative offset for the local variable "varNum."
2578 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2580 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2581 int lvaGetSPRelativeOffset(unsigned varNum);
2583 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2584 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2586 //------------------------ For splitting types ----------------------------
2588 void lvaInitTypeRef();
2590 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2591 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2592 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2593 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2594 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2595 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2597 void lvaInitVarDsc(LclVarDsc* varDsc,
2599 CorInfoType corInfoType,
2600 CORINFO_CLASS_HANDLE typeHnd,
2601 CORINFO_ARG_LIST_HANDLE varList,
2602 CORINFO_SIG_INFO* varSig);
2604 static unsigned lvaTypeRefMask(var_types type);
2606 var_types lvaGetActualType(unsigned lclNum);
2607 var_types lvaGetRealType(unsigned lclNum);
2609 //-------------------------------------------------------------------------
2613 unsigned lvaLclSize(unsigned varNum);
2614 unsigned lvaLclExactSize(unsigned varNum);
2616 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2618 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2619 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2620 // the return result.
2621 bool lvaLclVarRefsAccum(
2622 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2624 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2625 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2626 // and (destructively) unions "trkedVars" into "*result".
2627 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2629 ALLVARSET_VALARG_TP allVars,
2630 VARSET_VALARG_TP trkdVars);
2632 bool lvaHaveManyLocals() const;
2634 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2635 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2636 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2639 void lvaSortByRefCount();
2640 void lvaDumpRefCounts();
2642 void lvaMarkLocalVars(BasicBlock* block);
2644 void lvaMarkLocalVars(); // Local variable ref-counting
2646 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2648 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2650 void lvaIncRefCnts(GenTreePtr tree);
2651 void lvaDecRefCnts(GenTreePtr tree);
2653 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2654 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2655 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2658 struct lvaStressLclFldArgs
2660 Compiler* m_pCompiler;
2664 static fgWalkPreFn lvaStressLclFldCB;
2665 void lvaStressLclFld();
2667 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2668 void lvaDispVarSet(VARSET_VALARG_TP set);
2673 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2675 int lvaFrameAddress(int varNum, bool* pFPbased);
2678 bool lvaIsParameter(unsigned varNum);
2679 bool lvaIsRegArgument(unsigned varNum);
2680 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2681 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2682 // that writes to arg0
2684 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2685 // (this is an overload of lvIsTemp because there are no temp parameters).
2686 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2687 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2688 bool lvaIsImplicitByRefLocal(unsigned varNum)
2690 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2691 LclVarDsc* varDsc = &(lvaTable[varNum]);
2692 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2694 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2697 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2701 // Returns true if this local var is a multireg struct
2702 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2704 // If the local is a TYP_STRUCT, get/set a class handle describing it
2705 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2706 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2708 // If the local is TYP_REF, set or update the associated class information.
2709 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2710 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2711 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2712 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2714 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2716 // Info about struct fields
2717 struct lvaStructFieldInfo
2719 CORINFO_FIELD_HANDLE fldHnd;
2720 unsigned char fldOffset;
2721 unsigned char fldOrdinal;
2724 CORINFO_CLASS_HANDLE fldTypeHnd;
2727 // Info about struct to be promoted.
2728 struct lvaStructPromotionInfo
2730 CORINFO_CLASS_HANDLE typeHnd;
2732 bool requiresScratchVar;
2735 unsigned char fieldCnt;
2736 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2738 lvaStructPromotionInfo()
2739 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2744 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2745 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2746 lvaStructPromotionInfo* StructPromotionInfo,
2748 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2749 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2750 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2751 #if !defined(_TARGET_64BIT_)
2752 void lvaPromoteLongVars();
2753 #endif // !defined(_TARGET_64BIT_)
2754 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2755 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2756 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2757 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2758 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2759 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2760 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2762 #if defined(FEATURE_SIMD)
2763 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2765 assert(varDsc->lvType == TYP_SIMD12);
2766 assert(varDsc->lvExactSize == 12);
2768 #if defined(_TARGET_64BIT_)
2769 assert(varDsc->lvSize() == 16);
2770 #endif // defined(_TARGET_64BIT_)
2772 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2773 // already does this calculation. However, we also need to prevent mapping types if the var is a
2774 // dependently promoted struct field, which must remain its exact size within its parent struct.
2775 // However, we don't know this until late, so we may have already pretended the field is bigger
2777 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2786 #endif // defined(FEATURE_SIMD)
2788 BYTE* lvaGetGcLayout(unsigned varNum);
2789 bool lvaTypeIsGC(unsigned varNum);
2790 unsigned lvaGSSecurityCookie; // LclVar number
2791 bool lvaTempsHaveLargerOffsetThanVars();
2793 unsigned lvaSecurityObject; // variable representing the security object on the stack
2794 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2796 #if FEATURE_EH_FUNCLETS
2797 unsigned lvaPSPSym; // variable representing the PSPSym
2800 InlineInfo* impInlineInfo;
2801 InlineStrategy* m_inlineStrategy;
2803 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2804 Compiler* impInlineRoot();
2806 #if defined(DEBUG) || defined(INLINE_DATA)
2807 unsigned __int64 getInlineCycleCount()
2809 return m_compCycles;
2811 #endif // defined(DEBUG) || defined(INLINE_DATA)
2813 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2814 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2816 //=========================================================================
2818 //=========================================================================
2821 //---------------- Local variable ref-counting ----------------------------
2824 BasicBlock* lvaMarkRefsCurBlock;
2825 GenTreePtr lvaMarkRefsCurStmt;
2827 BasicBlock::weight_t lvaMarkRefsWeight;
2829 void lvaMarkLclRefs(GenTreePtr tree);
2831 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2832 void SetVolatileHint(LclVarDsc* varDsc);
2834 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2835 PerSsaArray lvMemoryPerSsaData;
2836 unsigned lvMemoryNumSsaNames;
2839 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2840 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2841 // not an SSA variable).
2842 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2844 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2845 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2847 assert(ssaNum < lvMemoryNumSsaNames);
2848 return &lvMemoryPerSsaData.GetRef(ssaNum);
2852 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2853 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2857 XX Imports the given method and converts it to semantic trees XX
2859 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2860 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2866 void impImport(BasicBlock* method);
2868 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2869 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2870 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2871 CORINFO_CLASS_HANDLE impGetStringClass();
2872 CORINFO_CLASS_HANDLE impGetObjectClass();
2874 //=========================================================================
2876 //=========================================================================
2879 //-------------------- Stack manipulation ---------------------------------
2881 unsigned impStkSize; // Size of the full stack
2883 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2885 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2887 struct SavedStack // used to save/restore stack contents.
2889 unsigned ssDepth; // number of values on stack
2890 StackEntry* ssTrees; // saved tree values
2893 bool impIsPrimitive(CorInfoType type);
2894 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2896 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2898 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2899 void impPushNullObjRefOnStack();
2900 StackEntry impPopStack();
2901 StackEntry& impStackTop(unsigned n = 0);
2902 unsigned impStackHeight();
2904 void impSaveStackState(SavedStack* savePtr, bool copy);
2905 void impRestoreStackState(SavedStack* savePtr);
2907 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2908 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2909 CORINFO_CALL_INFO* pCallInfo);
2911 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2913 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2915 bool impCanPInvokeInline();
2916 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2917 void impCheckForPInvokeCall(
2918 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2919 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2920 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2922 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2923 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2924 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2926 var_types impImportCall(OPCODE opcode,
2927 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2928 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2930 GenTreePtr newobjThis,
2932 CORINFO_CALL_INFO* callInfo,
2933 IL_OFFSET rawILOffset);
2935 void impDevirtualizeCall(GenTreeCall* call,
2937 CORINFO_METHOD_HANDLE* method,
2938 unsigned* methodFlags,
2939 CORINFO_CONTEXT_HANDLE* contextHandle,
2940 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2942 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2944 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2946 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2949 var_types impImportJitTestLabelMark(int numArgs);
2952 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2954 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2956 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2957 CORINFO_ACCESS_FLAGS access,
2958 CORINFO_FIELD_INFO* pFieldInfo,
2961 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2963 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2965 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2967 void impImportLeave(BasicBlock* block);
2968 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2969 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2970 CORINFO_CLASS_HANDLE clsHnd,
2971 CORINFO_METHOD_HANDLE method,
2972 CORINFO_SIG_INFO* sig,
2976 CorInfoIntrinsics* pIntrinsicID);
2977 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2978 CORINFO_SIG_INFO* sig,
2981 CorInfoIntrinsics intrinsicID);
2982 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2984 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2986 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2987 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2988 CORINFO_THIS_TRANSFORM transform);
2990 //----------------- Manipulating the trees and stmts ----------------------
2992 GenTreePtr impTreeList; // Trees for the BB being imported
2993 GenTreePtr impTreeLast; // The last tree for the current BB
2997 CHECK_SPILL_ALL = -1,
2998 CHECK_SPILL_NONE = -2
3002 void impBeginTreeList();
3003 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
3004 void impEndTreeList(BasicBlock* block);
3005 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3006 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3007 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3008 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3009 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3010 void impAssignTempGen(unsigned tmp,
3013 GenTreePtr* pAfterStmt = nullptr,
3014 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3015 BasicBlock* block = nullptr);
3016 void impAssignTempGen(unsigned tmpNum,
3018 CORINFO_CLASS_HANDLE structHnd,
3020 GenTreePtr* pAfterStmt = nullptr,
3021 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3022 BasicBlock* block = nullptr);
3023 GenTreePtr impCloneExpr(GenTreePtr tree,
3025 CORINFO_CLASS_HANDLE structHnd,
3027 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3028 GenTreePtr impAssignStruct(GenTreePtr dest,
3030 CORINFO_CLASS_HANDLE structHnd,
3032 GenTreePtr* pAfterStmt = nullptr,
3033 BasicBlock* block = nullptr);
3034 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3036 CORINFO_CLASS_HANDLE structHnd,
3038 GenTreePtr* pAfterStmt = nullptr,
3039 BasicBlock* block = nullptr);
3041 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3042 CORINFO_CLASS_HANDLE structHnd,
3046 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3047 BYTE* gcLayout = nullptr,
3048 unsigned* numGCVars = nullptr,
3049 var_types* simdBaseType = nullptr);
3051 GenTreePtr impNormStructVal(GenTreePtr structVal,
3052 CORINFO_CLASS_HANDLE structHnd,
3054 bool forceNormalization = false);
3056 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3057 BOOL* pRuntimeLookup = nullptr,
3058 BOOL mustRestoreHandle = FALSE,
3059 BOOL importParent = FALSE);
3061 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3062 BOOL* pRuntimeLookup = nullptr,
3063 BOOL mustRestoreHandle = FALSE)
3065 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3068 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3069 CORINFO_LOOKUP* pLookup,
3071 void* compileTimeHandle);
3073 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3075 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3076 CORINFO_LOOKUP* pLookup,
3077 void* compileTimeHandle);
3079 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3081 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3082 CorInfoHelpFunc helper,
3084 GenTreeArgList* arg = nullptr,
3085 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3087 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3089 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3092 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3093 CORINFO_CLASS_HANDLE typeClass,
3097 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3098 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3099 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3100 static bool IsMathIntrinsic(GenTreePtr tree);
3103 //----------------- Importing the method ----------------------------------
3105 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3108 unsigned impCurOpcOffs;
3109 const char* impCurOpcName;
3110 bool impNestedStackSpill;
3112 // For displaying instrs with generated native code (-n:B)
3113 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3114 void impNoteLastILoffs();
3117 /* IL offset of the stmt currently being imported. It gets set to
3118 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3119 updated at IL offsets for which we have to report mapping info.
3120 It also includes flag bits, so use jitGetILoffs()
3121 to get the actual IL offset value.
3124 IL_OFFSETX impCurStmtOffs;
3125 void impCurStmtOffsSet(IL_OFFSET offs);
3127 void impNoteBranchOffs();
3129 unsigned impInitBlockLineInfo();
3131 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3132 bool impIsThis(GenTreePtr obj);
3133 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3134 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3135 bool impIsAnySTLOC(OPCODE opcode)
3137 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3138 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3141 GenTreeArgList* impPopList(unsigned count, CORINFO_SIG_INFO* sig, GenTreeArgList* prefixTree = nullptr);
3143 GenTreeArgList* impPopRevList(unsigned count, CORINFO_SIG_INFO* sig, unsigned skipReverseCount = 0);
3146 * Get current IL offset with stack-empty info incoporated
3148 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3150 //---------------- Spilling the importer stack ----------------------------
3152 // The maximum number of bytes of IL processed without clean stack state.
3153 // It allows to limit the maximum tree size and depth.
3154 static const unsigned MAX_TREE_SIZE = 200;
3155 bool impCanSpillNow(OPCODE prevOpcode);
3161 SavedStack pdSavedStack;
3162 ThisInitState pdThisPtrInit;
3165 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3166 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3168 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3169 ExpandArray<BYTE> impPendingBlockMembers;
3171 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3172 // Operates on the map in the top-level ancestor.
3173 BYTE impGetPendingBlockMember(BasicBlock* blk)
3175 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3178 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3179 // Operates on the map in the top-level ancestor.
3180 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3182 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3185 bool impCanReimport;
3187 bool impSpillStackEntry(unsigned level,
3191 bool bAssertOnRecursion,
3196 void impSpillStackEnsure(bool spillLeaves = false);
3197 void impEvalSideEffects();
3198 void impSpillSpecialSideEff();
3199 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3200 void impSpillValueClasses();
3201 void impSpillEvalStack();
3202 static fgWalkPreFn impFindValueClasses;
3203 void impSpillLclRefs(ssize_t lclNum);
3205 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3207 void impImportBlockCode(BasicBlock* block);
3209 void impReimportMarkBlock(BasicBlock* block);
3210 void impReimportMarkSuccessors(BasicBlock* block);
3212 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3214 void impImportBlockPending(BasicBlock* block);
3216 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3217 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3218 // for the block, but instead, just re-uses the block's existing EntryState.
3219 void impReimportBlockPending(BasicBlock* block);
3221 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3223 void impImportBlock(BasicBlock* block);
3225 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3226 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3227 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3228 // the variables that will be used -- and for all the predecessors of those successors, and the
3229 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3230 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3231 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3232 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3233 // of local variable numbers, so we represent them with the base local variable number), returns that.
3234 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3235 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3236 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3237 // on which kind of member of the clique the block is).
3238 unsigned impGetSpillTmpBase(BasicBlock* block);
3240 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3241 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3242 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3243 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3244 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3245 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3246 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3247 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3248 // successors receive a native int. Similarly float and double are unified to double.
3249 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3250 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3251 // predecessors, so they insert an upcast if needed).
3252 void impReimportSpillClique(BasicBlock* block);
3254 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3255 // block, and represent the predecessor and successor members of the clique currently being computed.
3256 // *** Access to these will need to be locked in a parallel compiler.
3257 ExpandArray<BYTE> impSpillCliquePredMembers;
3258 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3266 // Abstract class for receiving a callback while walking a spill clique
3267 class SpillCliqueWalker
3270 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3273 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3274 class SetSpillTempsBase : public SpillCliqueWalker
3279 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3282 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3285 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3286 class ReimportSpillClique : public SpillCliqueWalker
3291 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3294 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3297 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3298 // predecessor or successor within the spill clique
3299 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3301 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3302 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3303 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3304 void impRetypeEntryStateTemps(BasicBlock* blk);
3306 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3307 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3309 void impPushVar(GenTree* op, typeInfo tiRetVal);
3310 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3311 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3313 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3315 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3316 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3317 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3320 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3323 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3324 struct BlockListNode
3327 BlockListNode* m_next;
3328 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3331 void* operator new(size_t sz, Compiler* comp);
3333 BlockListNode* impBlockListNodeFreeList;
3335 BlockListNode* AllocBlockListNode();
3336 void FreeBlockListNode(BlockListNode* node);
3338 bool impIsValueType(typeInfo* pTypeInfo);
3339 var_types mangleVarArgsType(var_types type);
3342 regNumber getCallArgIntRegister(regNumber floatReg);
3343 regNumber getCallArgFloatRegister(regNumber intReg);
3344 #endif // FEATURE_VARARG
3347 static unsigned jitTotalMethodCompiled;
3351 static LONG jitNestingLevel;
3354 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3356 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3358 // STATIC inlining decision based on the IL code.
3359 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3360 CORINFO_METHOD_INFO* methInfo,
3362 InlineResult* inlineResult);
3364 void impCheckCanInline(GenTreePtr call,
3365 CORINFO_METHOD_HANDLE fncHandle,
3367 CORINFO_CONTEXT_HANDLE exactContextHnd,
3368 InlineCandidateInfo** ppInlineCandidateInfo,
3369 InlineResult* inlineResult);
3371 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3372 GenTreePtr curArgVal,
3374 InlineResult* inlineResult);
3376 void impInlineInitVars(InlineInfo* pInlineInfo);
3378 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3380 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3382 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3384 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3385 GenTreePtr variableBeingDereferenced,
3386 InlArgInfo* inlArgInfo);
3388 void impMarkInlineCandidate(GenTreePtr call,
3389 CORINFO_CONTEXT_HANDLE exactContextHnd,
3390 bool exactContextNeedsRuntimeLookup,
3391 CORINFO_CALL_INFO* callInfo);
3393 bool impTailCallRetTypeCompatible(var_types callerRetType,
3394 CORINFO_CLASS_HANDLE callerRetTypeClass,
3395 var_types calleeRetType,
3396 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3398 bool impIsTailCallILPattern(bool tailPrefixed,
3400 const BYTE* codeAddrOfNextOpcode,
3401 const BYTE* codeEnd,
3403 bool* IsCallPopRet = nullptr);
3405 bool impIsImplicitTailCallCandidate(
3406 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3408 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3411 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3412 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3416 XX Info about the basic-blocks, their contents and the flow analysis XX
3418 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3419 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3423 BasicBlock* fgFirstBB; // Beginning of the basic block list
3424 BasicBlock* fgLastBB; // End of the basic block list
3425 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3426 #if FEATURE_EH_FUNCLETS
3427 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3429 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3431 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3432 unsigned fgEdgeCount; // # of control flow edges between the BBs
3433 unsigned fgBBcount; // # of BBs in the method
3435 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3437 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3438 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3439 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3440 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3442 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3443 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3444 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3445 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3446 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3447 // index). The arrays are of size fgBBNumMax + 1.
3448 unsigned* fgDomTreePreOrder;
3449 unsigned* fgDomTreePostOrder;
3451 bool fgBBVarSetsInited;
3453 // Allocate array like T* a = new T[fgBBNumMax + 1];
3454 // Using helper so we don't keep forgetting +1.
3455 template <typename T>
3456 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3458 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3461 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3462 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3463 // cannot be meaningfully combined. Note that new blocks can be created with higher
3464 // block numbers without changing the basic block epoch. These blocks *cannot*
3465 // participate in a block set until the blocks are all renumbered, causing the epoch
3466 // to change. This is useful if continuing to use previous block sets is valuable.
3467 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3468 unsigned fgCurBBEpoch;
3470 unsigned GetCurBasicBlockEpoch()
3472 return fgCurBBEpoch;
3475 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3476 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3477 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3478 unsigned fgCurBBEpochSize;
3480 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3481 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3482 unsigned fgBBSetCountInSizeTUnits;
3484 void NewBasicBlockEpoch()
3486 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3488 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3490 fgCurBBEpochSize = fgBBNumMax + 1;
3491 fgBBSetCountInSizeTUnits =
3492 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3495 // All BlockSet objects are now invalid!
3496 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3497 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3501 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3502 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3503 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3504 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3506 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3507 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3508 // array of size_t bitsets), then print that out.
3509 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3516 void EnsureBasicBlockEpoch()
3518 if (fgCurBBEpochSize != fgBBNumMax + 1)
3520 NewBasicBlockEpoch();
3524 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3525 void fgEnsureFirstBBisScratch();
3526 bool fgFirstBBisScratch();
3527 bool fgBBisScratch(BasicBlock* block);
3529 void fgExtendEHRegionBefore(BasicBlock* block);
3530 void fgExtendEHRegionAfter(BasicBlock* block);
3532 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3534 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3536 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3539 BasicBlock* nearBlk,
3540 bool putInFilter = false,
3541 bool runRarely = false,
3542 bool insertAtEnd = false);
3544 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3546 bool runRarely = false,
3547 bool insertAtEnd = false);
3549 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3551 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3552 BasicBlock* afterBlk,
3553 unsigned xcptnIndex,
3554 bool putInTryRegion);
3556 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3557 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3558 void fgUnlinkBlock(BasicBlock* block);
3560 unsigned fgMeasureIR();
3562 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3563 bool fgMultipleNots;
3566 bool fgModified; // True if the flow graph has been modified recently
3567 bool fgComputePredsDone; // Have we computed the bbPreds list
3568 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3569 bool fgDomsComputed; // Have we computed the dominator sets?
3570 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3572 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3573 bool fgHasPostfix; // any postfix ++/-- found?
3574 unsigned fgIncrCount; // number of increment nodes found
3576 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3580 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3581 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3584 bool fgRemoveRestOfBlock; // true if we know that we will throw
3585 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3587 // There are two modes for ordering of the trees.
3588 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3589 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3590 // by traversing the tree according to the order of the operands.
3591 // - In FGOrderLinear, the dominant ordering is the linear order.
3598 FlowGraphOrder fgOrder;
3600 // The following are boolean flags that keep track of the state of internal data structures
3602 bool fgStmtListThreaded;
3603 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3604 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3605 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3606 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3607 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3608 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3609 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3610 // This is derived from the profile data
3611 // or is BB_UNITY_WEIGHT when we don't have profile data
3613 #if FEATURE_EH_FUNCLETS
3614 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3615 #endif // FEATURE_EH_FUNCLETS
3617 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3618 // since fgMorphTree can be called from several places
3620 bool impBoxTempInUse; // the temp below is valid and available
3621 unsigned impBoxTemp; // a temporary that is used for boxing
3624 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3625 // and we are trying to compile again in a "safer", minopts mode?
3629 unsigned impInlinedCodeSize;
3632 //-------------------------------------------------------------------------
3638 void fgTransformFatCalli();
3642 void fgRemoveEmptyTry();
3644 void fgRemoveEmptyFinally();
3646 void fgMergeFinallyChains();
3648 void fgCloneFinally();
3650 void fgCleanupContinuation(BasicBlock* continuation);
3652 void fgUpdateFinallyTargetFlags();
3654 void fgClearAllFinallyTargetBits();
3656 void fgAddFinallyTargetFlags();
3658 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3659 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3660 // when this is necessary.
3661 bool fgNeedToAddFinallyTargetBits;
3662 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3664 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3665 BasicBlock* handler,
3666 BlockToBlockMap& continuationMap);
3668 GenTreePtr fgGetCritSectOfStaticMethod();
3670 #if FEATURE_EH_FUNCLETS
3672 void fgAddSyncMethodEnterExit();
3674 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3676 void fgConvertSyncReturnToLeave(BasicBlock* block);
3678 #endif // FEATURE_EH_FUNCLETS
3680 void fgAddReversePInvokeEnterExit();
3682 bool fgMoreThanOneReturnBlock();
3684 // The number of separate return points in the method.
3685 unsigned fgReturnCount;
3687 void fgAddInternal();
3689 bool fgFoldConditional(BasicBlock* block);
3691 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3692 void fgMorphBlocks();
3694 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3696 void fgCheckArgCnt();
3697 void fgSetOptions();
3700 static fgWalkPreFn fgAssertNoQmark;
3701 void fgPreExpandQmarkChecks(GenTreePtr expr);
3702 void fgPostExpandQmarkChecks();
3703 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3706 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3708 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3709 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3710 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3711 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3712 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3714 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3715 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3716 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3717 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3719 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3720 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3721 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3722 void fgExpandQmarkNodes();
3726 // Do "simple lowering." This functionality is (conceptually) part of "general"
3727 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3728 void fgSimpleLowering();
3730 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3732 GenTreePtr fgInitThisClass();
3734 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3736 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3738 inline bool backendRequiresLocalVarLifetimes()
3740 #if defined(LEGACY_BACKEND)
3743 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3747 void fgLocalVarLiveness();
3749 void fgLocalVarLivenessInit();
3751 #ifdef LEGACY_BACKEND
3752 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3754 void fgPerNodeLocalVarLiveness(GenTree* node);
3756 void fgPerBlockLocalVarLiveness();
3758 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3760 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3762 // This is used in the liveness computation, as a temporary. When we use the
3763 // arbitrary-length VarSet representation, it is better not to allocate a new one
3765 VARSET_TP fgMarkIntfUnionVS;
3767 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3769 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3771 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3773 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3775 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3777 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3779 void fgComputeLifeTrackedLocalUse(VARSET_TP& life, LclVarDsc& varDsc, GenTreeLclVarCommon* node);
3780 bool fgComputeLifeTrackedLocalDef(VARSET_TP& life,
3781 VARSET_VALARG_TP keepAliveVars,
3783 GenTreeLclVarCommon* node);
3784 void fgComputeLifeUntrackedLocal(VARSET_TP& life,
3785 VARSET_VALARG_TP keepAliveVars,
3787 GenTreeLclVarCommon* lclVarNode,
3789 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3791 void fgComputeLife(VARSET_TP& life,
3792 GenTreePtr startNode,
3794 VARSET_VALARG_TP volatileVars,
3795 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3797 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3799 bool fgRemoveDeadStore(GenTree** pTree,
3801 VARSET_VALARG_TP life,
3803 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3805 // For updating liveset during traversal AFTER fgComputeLife has completed
3806 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3807 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3809 // Returns the set of live variables after endTree,
3810 // assuming that liveSet is the set of live variables BEFORE tree.
3811 // Requires that fgComputeLife has completed, and that tree is in the same
3812 // statement as endTree, and that it comes before endTree in execution order
3814 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3816 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3817 while (tree != nullptr && tree != endTree->gtNext)
3819 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3820 tree = tree->gtNext;
3822 assert(tree == endTree->gtNext);
3826 void fgInterBlockLocalVarLiveness();
3828 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3829 // "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
3830 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3831 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3832 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3833 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3834 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3836 if (m_opAsgnVarDefSsaNums == nullptr)
3838 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3840 return m_opAsgnVarDefSsaNums;
3843 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3844 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3845 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3847 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3849 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3850 // Except: assumes that lcl is a def, and if it is
3851 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3852 // rather than the "use" SSA number recorded in the tree "lcl".
3853 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3855 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3856 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3857 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3858 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3859 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3861 // (byref addrS1 = &s1,
3862 // *(addrS1 * offsetof(f0)) = s2f0,
3864 // *(addrS1 * offsetof(fn)) = s2fn)
3866 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3867 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3868 // give it SSA names and value numbers?
3870 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3871 // end with an instance of the structure below, whose fields are described in the declaration.
3872 struct IndirectAssignmentAnnotation
3874 unsigned m_lclNum; // The local num that is being indirectly assigned.
3875 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3876 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3877 // be the singleton field sequence "g". The individual assignments would
3878 // further append the fields of "s.g" to that.
3879 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3880 // structure has a single field).
3881 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3882 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3885 IndirectAssignmentAnnotation(unsigned lclNum,
3886 FieldSeqNode* fldSeq,
3888 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3889 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3890 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3894 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3895 NodeToIndirAssignMap;
3896 NodeToIndirAssignMap* m_indirAssignMap;
3897 NodeToIndirAssignMap* GetIndirAssignMap()
3899 if (m_indirAssignMap == nullptr)
3901 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3902 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3903 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3905 return m_indirAssignMap;
3908 // Performs SSA conversion.
3911 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3912 void fgResetForSsa();
3914 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3916 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3917 inline bool fgExcludeFromSsa(unsigned lclNum);
3919 // The value numbers for this compilation.
3920 ValueNumStore* vnStore;
3923 ValueNumStore* GetValueNumStore()
3928 // Do value numbering (assign a value number to each
3930 void fgValueNumber();
3932 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3933 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3934 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3935 // match the element type of the array or fldSeq. When this type doesn't match
3936 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3938 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3941 FieldSeqNode* fldSeq,
3945 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3946 // has been parsed to yield the other input arguments. If evaluation of the address
3947 // can raise exceptions, those should be captured in the exception set "excVN."
3948 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3949 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3950 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3951 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3952 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3954 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3955 CORINFO_CLASS_HANDLE elemTypeEq,
3959 FieldSeqNode* fldSeq);
3961 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3962 // by evaluating the array index expression "tree". Returns the value number resulting from
3963 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3964 // "GT_IND" that does the dereference, and it is given the returned value number.
3965 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3967 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3968 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3970 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3972 // Utility functions for fgValueNumber.
3974 // Perform value-numbering for the trees in "blk".
3975 void fgValueNumberBlock(BasicBlock* blk);
3977 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3978 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3979 // assumed for the memoryKind at the start "entryBlk".
3980 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3982 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3983 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3984 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3986 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3988 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3990 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3991 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3992 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3994 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3995 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3997 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3998 // value in that SSA #.
3999 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
4001 // The input 'tree' is a leaf node that is a constant
4002 // Assign the proper value number to the tree
4003 void fgValueNumberTreeConst(GenTreePtr tree);
4005 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
4006 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4008 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4010 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
4012 // Does value-numbering for a block assignment.
4013 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4015 // Does value-numbering for a cast tree.
4016 void fgValueNumberCastTree(GenTreePtr tree);
4018 // Does value-numbering for an intrinsic tree.
4019 void fgValueNumberIntrinsic(GenTreePtr tree);
4021 // Does value-numbering for a call. We interpret some helper calls.
4022 void fgValueNumberCall(GenTreeCall* call);
4024 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4025 void fgUpdateArgListVNs(GenTreeArgList* args);
4027 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4028 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4030 // Requires "helpCall" to be a helper call. Assigns it a value number;
4031 // we understand the semantics of some of the calls. Returns "true" if
4032 // the call may modify the heap (we assume arbitrary memory side effects if so).
4033 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4035 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4036 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4038 // These are the current value number for the memory implicit variables while
4039 // doing value numbering. These are the value numbers under the "liberal" interpretation
4040 // of memory values; the "conservative" interpretation needs no VN, since every access of
4041 // memory yields an unknown value.
4042 ValueNum fgCurMemoryVN[MemoryKindCount];
4044 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4045 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4046 // is 1, and the rest is an encoding of "elemTyp".
4047 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4049 if (elemStructType != nullptr)
4051 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4052 varTypeIsIntegral(elemTyp));
4053 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4054 return elemStructType;
4058 elemTyp = varTypeUnsignedToSigned(elemTyp);
4059 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4062 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4063 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4064 // the struct type of the element).
4065 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4067 size_t clsHndVal = size_t(clsHnd);
4068 if (clsHndVal & 0x1)
4070 return var_types(clsHndVal >> 1);
4078 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4079 var_types getJitGCType(BYTE gcType);
4081 enum structPassingKind
4083 SPK_Unknown, // Invalid value, never returned
4084 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4085 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4086 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4087 // parameters registers are used, then the stack will be used)
4088 // for X86 passed on the stack, for ARM32 passed in registers
4089 // or the stack or split between registers and the stack.
4090 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4092 }; // The struct is passed/returned by reference to a copy/buffer.
4094 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4095 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4096 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4097 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4099 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4101 // Get the type that is used to pass values of the given struct type.
4102 // If you have already retrieved the struct size then pass it as the optional third argument
4104 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4105 structPassingKind* wbPassStruct,
4106 unsigned structSize = 0);
4108 // Get the type that is used to return values of the given struct type.
4109 // If you have already retrieved the struct size then pass it as the optional third argument
4111 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4112 structPassingKind* wbPassStruct = nullptr,
4113 unsigned structSize = 0);
4116 // Print a representation of "vnp" or "vn" on standard output.
4117 // If "level" is non-zero, we also print out a partial expansion of the value.
4118 void vnpPrint(ValueNumPair vnp, unsigned level);
4119 void vnPrint(ValueNum vn, unsigned level);
4122 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4124 // Dominator computation member functions
4125 // Not exposed outside Compiler
4127 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4129 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4130 // flow graph. We first assume the fields bbIDom on each
4131 // basic block are invalid. This computation is needed later
4132 // by fgBuildDomTree to build the dominance tree structure.
4133 // Based on: A Simple, Fast Dominance Algorithm
4134 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4136 void fgCompDominatedByExceptionalEntryBlocks();
4138 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4139 // Note: this is relatively slow compared to calling fgDominate(),
4140 // especially if dealing with a single block versus block check.
4142 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4144 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4146 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4148 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4150 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4152 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4153 // processed in topological sort, this function takes care of that.
4155 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4157 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4158 // Returns this as a set.
4160 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4161 // root nodes. Returns this as a set.
4164 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4167 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4168 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4171 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4172 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4173 // && postOrder(A) >= postOrder(B) making the computation O(1).
4174 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4176 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4178 void fgUpdateChangedFlowGraph();
4181 // Compute the predecessors of the blocks in the control flow graph.
4182 void fgComputePreds();
4184 // Remove all predecessor information.
4185 void fgRemovePreds();
4187 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4188 // before the full predecessors lists are computed.
4189 void fgComputeCheapPreds();
4192 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4194 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4204 // Initialize the per-block variable sets (used for liveness analysis).
4205 void fgInitBlockVarSets();
4207 // true if we've gone through and created GC Poll calls.
4208 bool fgGCPollsCreated;
4209 void fgMarkGCPollBlocks();
4210 void fgCreateGCPolls();
4211 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4213 // Requires that "block" is a block that returns from
4214 // a finally. Returns the number of successors (jump targets of
4215 // of blocks in the covered "try" that did a "LEAVE".)
4216 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4218 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4219 // a finally. Returns its "i"th successor (jump targets of
4220 // of blocks in the covered "try" that did a "LEAVE".)
4221 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4222 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4225 // Factor out common portions of the impls of the methods above.
4226 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4229 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4230 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4231 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4232 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4233 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4234 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4235 // we leave the entry associated with the block, but it will no longer be accessed.)
4236 struct SwitchUniqueSuccSet
4238 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4239 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4242 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4243 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4244 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4245 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4248 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4249 BlockToSwitchDescMap;
4252 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4253 // iteration over only the distinct successors.
4254 BlockToSwitchDescMap* m_switchDescMap;
4257 BlockToSwitchDescMap* GetSwitchDescMap()
4259 if (m_switchDescMap == nullptr)
4261 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4263 return m_switchDescMap;
4266 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4267 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4268 // we don't accidentally look up and return the wrong switch data.
4269 void InvalidateUniqueSwitchSuccMap()
4271 m_switchDescMap = nullptr;
4274 // Requires "switchBlock" to be a block that ends in a switch. Returns
4275 // the corresponding SwitchUniqueSuccSet.
4276 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4278 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4279 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4280 // remove it from "this", and ensure that "to" is a member.
4281 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4283 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4284 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4286 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4288 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4290 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4292 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4294 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4296 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4298 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4300 void fgRemoveBlockAsPred(BasicBlock* block);
4302 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4304 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4306 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4308 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4310 flowList* fgAddRefPred(BasicBlock* block,
4311 BasicBlock* blockPred,
4312 flowList* oldEdge = nullptr,
4313 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4316 void fgFindBasicBlocks();
4318 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4320 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4322 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4323 bool putInTryRegion,
4324 BasicBlock* startBlk,
4326 BasicBlock* nearBlk,
4327 BasicBlock* jumpBlk,
4330 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4332 void fgRemoveEmptyBlocks();
4334 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4336 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4338 void fgCreateLoopPreHeader(unsigned lnum);
4340 void fgUnreachableBlock(BasicBlock* block);
4342 void fgRemoveConditionalJump(BasicBlock* block);
4344 BasicBlock* fgLastBBInMainFunction();
4346 BasicBlock* fgEndBBAfterMainFunction();
4348 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4350 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4352 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4354 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4356 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4358 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4360 bool fgRenumberBlocks();
4362 bool fgExpandRarelyRunBlocks();
4364 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4366 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4368 enum FG_RELOCATE_TYPE
4370 FG_RELOCATE_TRY, // relocate the 'try' region
4371 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4373 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4375 #if FEATURE_EH_FUNCLETS
4376 #if defined(_TARGET_ARM_)
4377 void fgClearFinallyTargetBit(BasicBlock* block);
4378 #endif // defined(_TARGET_ARM_)
4379 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4380 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4381 void fgInsertFuncletPrologBlock(BasicBlock* block);
4382 void fgCreateFuncletPrologBlocks();
4383 void fgCreateFunclets();
4384 #else // !FEATURE_EH_FUNCLETS
4385 bool fgRelocateEHRegions();
4386 #endif // !FEATURE_EH_FUNCLETS
4388 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4390 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4392 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4394 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4396 bool fgOptimizeEmptyBlock(BasicBlock* block);
4398 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4400 bool fgOptimizeBranch(BasicBlock* bJump);
4402 bool fgOptimizeSwitchBranches(BasicBlock* block);
4404 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4406 bool fgOptimizeSwitchJumps();
4408 void fgPrintEdgeWeights();
4410 void fgComputeEdgeWeights();
4412 void fgReorderBlocks();
4414 void fgDetermineFirstColdBlock();
4416 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4418 bool fgUpdateFlowGraph(bool doTailDup = false);
4420 void fgFindOperOrder();
4422 // method that returns if you should split here
4423 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4425 void fgSetBlockOrder();
4427 void fgRemoveReturnBlock(BasicBlock* block);
4429 /* Helper code that has been factored out */
4430 inline void fgConvertBBToThrowBB(BasicBlock* block);
4432 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4433 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4434 GenTreePtr fgMakeTmpArgNode(
4435 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4437 // The following check for loops that don't execute calls
4438 bool fgLoopCallMarked;
4440 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4441 void fgLoopCallMark();
4443 void fgMarkLoopHead(BasicBlock* block);
4445 unsigned fgGetCodeEstimate(BasicBlock* block);
4448 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4449 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4450 bool fgDumpFlowGraph(Phases phase);
4452 #endif // DUMP_FLOWGRAPHS
4457 void fgDispBBLiveness(BasicBlock* block);
4458 void fgDispBBLiveness();
4459 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4460 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4461 void fgDispBasicBlocks(bool dumpTrees = false);
4462 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4463 void fgDumpBlock(BasicBlock* block);
4464 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4466 static fgWalkPreFn fgStress64RsltMulCB;
4467 void fgStress64RsltMul();
4468 void fgDebugCheckUpdate();
4469 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4470 void fgDebugCheckBlockLinks();
4471 void fgDebugCheckLinks(bool morphTrees = false);
4472 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4473 void fgDebugCheckFlags(GenTreePtr tree);
4474 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4475 void fgDebugCheckTryFinallyExits();
4478 #ifdef LEGACY_BACKEND
4479 static void fgOrderBlockOps(GenTreePtr tree,
4483 GenTreePtr* opsPtr, // OUT
4484 regMaskTP* regsPtr); // OUT
4485 #endif // LEGACY_BACKEND
4487 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4488 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4489 void fgTraverseRPO();
4491 //--------------------- Walking the trees in the IR -----------------------
4496 fgWalkPreFn* wtprVisitorFn;
4497 fgWalkPostFn* wtpoVisitorFn;
4498 void* pCallbackData; // user-provided data
4499 bool wtprLclsOnly; // whether to only visit lclvar nodes
4500 GenTreePtr parent; // parent of current node, provided to callback
4501 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4503 bool printModified; // callback can use this
4507 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4508 fgWalkPreFn* visitor,
4509 void* pCallBackData = nullptr,
4510 bool lclVarsOnly = false,
4511 bool computeStack = false);
4513 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4514 fgWalkPreFn* preVisitor,
4515 fgWalkPostFn* postVisitor,
4516 void* pCallBackData = nullptr);
4518 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4522 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4523 fgWalkPostFn* visitor,
4524 void* pCallBackData = nullptr,
4525 bool computeStack = false);
4527 // An fgWalkPreFn that looks for expressions that have inline throws in
4528 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4529 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4530 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4531 // properly propagated to parent trees). It returns WALK_CONTINUE
4533 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4534 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4535 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4537 /**************************************************************************
4539 *************************************************************************/
4542 friend class SsaBuilder;
4543 friend struct ValueNumberState;
4545 //--------------------- Detect the basic blocks ---------------------------
4547 BasicBlock** fgBBs; // Table of pointers to the BBs
4549 void fgInitBBLookup();
4550 BasicBlock* fgLookupBB(unsigned addr);
4552 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4554 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4556 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4558 void fgLinkBasicBlocks();
4560 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4562 void fgCheckBasicBlockControlFlow();
4564 void fgControlFlowPermitted(BasicBlock* blkSrc,
4565 BasicBlock* blkDest,
4566 BOOL IsLeave = false /* is the src a leave block */);
4568 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4570 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4572 void fgAdjustForAddressExposedOrWrittenThis();
4574 bool fgProfileData_ILSizeMismatch;
4575 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4576 ULONG fgProfileBufferCount;
4577 ULONG fgNumProfileRuns;
4579 unsigned fgStressBBProf()
4582 unsigned result = JitConfig.JitStressBBProf();
4585 if (compStressCompile(STRESS_BB_PROFILE, 15))
4596 bool fgHaveProfileData();
4597 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4598 void fgInstrumentMethod();
4601 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4602 // or if we have some fake profile data for the stress mode
4603 bool fgIsUsingProfileWeights()
4605 return (fgHaveProfileData() || fgStressBBProf());
4608 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4609 // or BB_UNITY_WEIGHT when we aren't using profile data.
4610 unsigned fgProfileRunsCount()
4612 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4615 //-------- Insert a statement at the start or end of a basic block --------
4619 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4623 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4625 public: // Used by linear scan register allocation
4626 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4629 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4630 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4632 public: // Used by linear scan register allocation
4633 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4636 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4638 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4640 // Create a new temporary variable to hold the result of *ppTree,
4641 // and transform the graph accordingly.
4642 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4643 GenTree* fgMakeMultiUse(GenTree** ppTree);
4646 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4647 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4648 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4650 //-------- Determine the order in which the trees will be evaluated -------
4652 unsigned fgTreeSeqNum;
4653 GenTree* fgTreeSeqLst;
4654 GenTree* fgTreeSeqBeg;
4656 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4657 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4658 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4659 void fgSetStmtSeq(GenTree* tree);
4660 void fgSetBlockOrder(BasicBlock* block);
4662 //------------------------- Morphing --------------------------------------
4664 unsigned fgPtrArgCntCur;
4665 unsigned fgPtrArgCntMax;
4666 hashBv* fgOutgoingArgTemps;
4667 hashBv* fgCurrentlyInUseArgTemps;
4669 bool compCanEncodePtrArgCntMax();
4671 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4674 void fgMoveOpsLeft(GenTreePtr tree);
4677 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4679 bool fgIsThrow(GenTreePtr tree);
4681 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4682 bool fgIsBlockCold(BasicBlock* block);
4684 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4686 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4688 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4690 bool fgMorphRelopToQmark(GenTreePtr tree);
4692 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4693 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4694 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4695 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4696 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4697 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4698 // small; hence the other fields of MorphAddrContext.
4699 enum MorphAddrContextKind
4704 struct MorphAddrContext
4706 MorphAddrContextKind m_kind;
4707 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4708 // top-level indirection and here have been constants.
4709 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4710 // In that case, is the sum of those constant offsets.
4712 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4717 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4718 static MorphAddrContext s_CopyBlockMAC;
4721 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4722 var_types* baseTypeOut,
4724 unsigned* simdSizeOut,
4725 bool ignoreUsedInSIMDIntrinsic = false);
4726 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4727 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4728 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4729 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4731 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4732 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4733 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4735 #endif // FEATURE_SIMD
4736 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4737 GenTreePtr fgMorphCast(GenTreePtr tree);
4738 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4739 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4741 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4744 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4745 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4747 void fgFixupStructReturn(GenTreePtr call);
4748 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4749 bool fgAddrCouldBeNull(GenTreePtr addr);
4750 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4751 bool fgCanFastTailCall(GenTreeCall* call);
4752 void fgMorphTailCall(GenTreeCall* call);
4753 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4754 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4755 fgArgTabEntryPtr argTabEntry,
4757 IL_OFFSETX callILOffset,
4758 GenTreePtr tmpAssignmentInsertionPoint,
4759 GenTreePtr paramAssignmentInsertionPoint);
4760 static int fgEstimateCallStackSize(GenTreeCall* call);
4761 GenTreePtr fgMorphCall(GenTreeCall* call);
4762 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4763 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4765 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4766 static fgWalkPreFn fgFindNonInlineCandidate;
4768 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4769 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4770 CORINFO_RESOLVED_TOKEN* ldftnToken);
4771 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4772 void fgAssignSetVarDef(GenTreePtr tree);
4773 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4774 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4775 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4776 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4777 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4778 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4779 void fgMorphUnsafeBlk(GenTreeObj* obj);
4780 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4781 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4782 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4783 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4784 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4785 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4786 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4788 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4789 GenTreePtr fgMorphConst(GenTreePtr tree);
4792 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4795 #if LOCAL_ASSERTION_PROP
4796 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4797 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4799 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4801 GenTreeStmt* fgMorphStmt;
4803 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4804 // used when morphing big offset.
4806 //----------------------- Liveness analysis -------------------------------
4808 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4809 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4811 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4812 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4813 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4815 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4817 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4819 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4820 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4822 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4823 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4825 void fgExtendDbgScopes();
4826 void fgExtendDbgLifetimes();
4829 void fgDispDebugScopes();
4832 //-------------------------------------------------------------------------
4834 // The following keeps track of any code we've added for things like array
4835 // range checking or explicit calls to enable GC, and so on.
4840 AddCodeDsc* acdNext;
4841 BasicBlock* acdDstBlk; // block to which we jump
4843 SpecialCodeKind acdKind; // what kind of a special block is this?
4844 unsigned short acdStkLvl;
4848 static unsigned acdHelper(SpecialCodeKind codeKind);
4850 AddCodeDsc* fgAddCodeList;
4852 bool fgRngChkThrowAdded;
4853 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4855 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4857 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4860 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4863 bool fgIsCodeAdded();
4865 bool fgIsThrowHlpBlk(BasicBlock* block);
4866 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4868 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4870 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4871 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4872 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4873 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4874 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4876 #if FEATURE_MULTIREG_RET
4877 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4878 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4879 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4880 #endif // FEATURE_MULTIREG_RET
4882 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4885 static fgWalkPreFn fgDebugCheckInlineCandidates;
4887 void CheckNoFatPointerCandidatesLeft();
4888 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4891 void fgPromoteStructs();
4892 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4893 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4895 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4896 void fgMarkImplicitByRefArgs();
4898 // Change implicit byrefs' types from struct to pointer, and for any that were
4899 // promoted, create new promoted struct temps.
4900 void fgRetypeImplicitByRefArgs();
4902 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4903 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4904 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4906 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4907 void fgMarkDemotedImplicitByRefArgs();
4909 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4910 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4911 void fgMarkAddressExposedLocals();
4912 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4914 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4916 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4918 // The given local variable, required to be a struct variable, is being assigned via
4919 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4920 // the variable is not enregistered, and is therefore not promoted independently.
4921 void fgLclFldAssign(unsigned lclNum);
4923 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4924 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4925 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4926 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4929 bool fgPrintInlinedMethods;
4932 bool fgIsBigOffset(size_t offset);
4934 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4935 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4936 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4937 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4938 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4941 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4942 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4946 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4947 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4954 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4957 void optRemoveRangeCheck(
4958 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4959 bool optIsRangeCheckRemovable(GenTreePtr tree);
4962 static fgWalkPreFn optValidRangeCheckIndex;
4963 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4966 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4968 /**************************************************************************
4970 *************************************************************************/
4973 // Do hoisting for all loops.
4974 void optHoistLoopCode();
4976 // To represent sets of VN's that have already been hoisted in outer loops.
4977 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4978 typedef VNToBoolMap VNSet;
4980 struct LoopHoistContext
4983 // The set of variables hoisted in the current loop (or nullptr if there are none).
4984 VNSet* m_pHoistedInCurLoop;
4987 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4988 VNSet m_hoistedInParentLoops;
4989 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4990 // Previous decisions on loop-invariance of value numbers in the current loop.
4991 VNToBoolMap m_curLoopVnInvariantCache;
4993 VNSet* GetHoistedInCurLoop(Compiler* comp)
4995 if (m_pHoistedInCurLoop == nullptr)
4997 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4999 return m_pHoistedInCurLoop;
5002 VNSet* ExtractHoistedInCurLoop()
5004 VNSet* res = m_pHoistedInCurLoop;
5005 m_pHoistedInCurLoop = nullptr;
5009 LoopHoistContext(Compiler* comp)
5010 : m_pHoistedInCurLoop(nullptr)
5011 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5012 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5017 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5018 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5019 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5020 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5022 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5023 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5024 // "m_hoistedInParentLoops".
5026 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5028 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5029 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5030 // expressions to "hoistInLoop".
5031 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5033 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5034 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5036 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5037 // that are invariant in loop "lnum" (an index into the optLoopTable)
5038 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5039 // expressions to "hoistInLoop".
5040 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5041 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5042 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5043 bool optHoistLoopExprsForTree(GenTreePtr tree,
5045 LoopHoistContext* hoistCtxt,
5046 bool* firstBlockAndBeforeSideEffect,
5048 bool* pCctorDependent);
5050 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5051 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5053 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5054 // Constants and init values are always loop invariant.
5055 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5056 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5058 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5059 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5060 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5061 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5062 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5064 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5065 // in the loop table.
5066 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5068 // Records the set of "side effects" of all loops: fields (object instance and static)
5069 // written to, and SZ-array element type equivalence classes updated.
5070 void optComputeLoopSideEffects();
5073 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5074 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5075 // static) written to, and SZ-array element type equivalence classes updated.
5076 void optComputeLoopNestSideEffects(unsigned lnum);
5078 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5079 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5081 // Hoist the expression "expr" out of loop "lnum".
5082 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5085 void optOptimizeBools();
5088 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5090 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5093 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5095 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5096 // the loop into a "do-while" loop
5097 // Also finds all natural loops and records them in the loop table
5099 // Optionally clone loops in the loop table.
5100 void optCloneLoops();
5102 // Clone loop "loopInd" in the loop table.
5103 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5105 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5106 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5107 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5109 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5111 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5114 // This enumeration describes what is killed by a call.
5118 CALLINT_NONE, // no interference (most helpers)
5119 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5120 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5121 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5122 CALLINT_ALL, // kills everything (normal method call)
5126 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5127 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5128 // in bbNext order; we use comparisons on the bbNum to decide order.)
5129 // The blocks that define the body are
5130 // first <= top <= entry <= bottom .
5131 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5132 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5133 // Compiler::optFindNaturalLoops().
5136 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5137 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5138 // loop, but not the outer loop.)
5139 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5141 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5142 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5143 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5145 callInterf lpAsgCall; // "callInterf" for calls in the loop
5146 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5147 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5149 unsigned short lpFlags; // Mask of the LPFLG_* constants
5151 unsigned char lpExitCnt; // number of exits from the loop
5153 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5154 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5155 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5156 // (Actually, an "immediately" nested loop --
5157 // no other child of this loop is a parent of lpChild.)
5158 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5159 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5160 // by following "lpChild" then "lpSibling" links.
5162 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5163 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5165 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5166 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5167 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5169 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5170 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5172 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5173 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5174 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5175 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5177 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5178 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5179 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5181 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5182 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5183 // type are assigned to.
5185 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5186 // memory side effects. If this is set, the fields below
5187 // may not be accurate (since they become irrelevant.)
5188 bool lpContainsCall; // True if executing the loop body *may* execute a call
5190 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5191 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5193 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5195 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5196 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5198 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5200 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5201 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5203 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5204 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5206 JitSimplerHashBehavior>
5208 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5209 // instance fields modified
5212 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5213 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5215 JitSimplerHashBehavior>
5217 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5218 // arrays of that type are modified
5221 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5222 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5224 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5225 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5226 // (shifted left, with a low-order bit set to distinguish.)
5227 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5228 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5230 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5232 GenTreePtr lpIterTree; // The "i <op>= const" tree
5233 unsigned lpIterVar(); // iterator variable #
5234 int lpIterConst(); // the constant with which the iterator is incremented
5235 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5236 void VERIFY_lpIterTree();
5238 var_types lpIterOperType(); // For overflow instructions
5241 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5242 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5246 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5248 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5249 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5250 void VERIFY_lpTestTree();
5252 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5253 GenTreePtr lpIterator(); // the iterator node in the loop test
5254 GenTreePtr lpLimit(); // the limit node in the loop test
5256 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5257 // LPFLG_CONST_LIMIT
5258 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5260 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5261 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5262 // LPFLG_ARRLEN_LIMIT
5264 // Returns "true" iff "*this" contains the blk.
5265 bool lpContains(BasicBlock* blk)
5267 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5269 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5270 // to be equal, but requiring bottoms to be different.)
5271 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5273 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5276 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5277 // bottoms to be different.)
5278 bool lpContains(const LoopDsc& lp2)
5280 return lpContains(lp2.lpFirst, lp2.lpBottom);
5283 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5284 // (allowing firsts to be equal, but requiring bottoms to be different.)
5285 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5287 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5290 // Returns "true" iff "*this" is (properly) contained by "lp2"
5291 // (allowing firsts to be equal, but requiring bottoms to be different.)
5292 bool lpContainedBy(const LoopDsc& lp2)
5294 return lpContains(lp2.lpFirst, lp2.lpBottom);
5297 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5298 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5300 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5302 // Returns "true" iff "*this" is disjoint from "lp2".
5303 bool lpDisjoint(const LoopDsc& lp2)
5305 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5307 // Returns "true" iff the loop is well-formed (see code for defn).
5310 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5311 lpEntry->bbNum <= lpBottom->bbNum &&
5312 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5317 bool fgMightHaveLoop(); // returns true if there are any backedges
5318 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5321 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5322 unsigned char optLoopCount; // number of tracked loops
5324 bool optRecordLoop(BasicBlock* head,
5330 unsigned char exitCnt);
5333 unsigned optCallCount; // number of calls made in the method
5334 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5335 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5336 unsigned optLoopsCloned; // number of loops cloned in the current method.
5339 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5340 void optPrintLoopInfo(unsigned loopNum,
5342 BasicBlock* lpFirst,
5344 BasicBlock* lpEntry,
5345 BasicBlock* lpBottom,
5346 unsigned char lpExitCnt,
5348 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5349 void optPrintLoopInfo(unsigned lnum);
5350 void optPrintLoopRecording(unsigned lnum);
5352 void optCheckPreds();
5355 void optSetBlockWeights();
5357 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5359 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5361 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5363 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5364 unsigned optIsLoopIncrTree(GenTreePtr incr);
5365 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5366 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5367 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5368 bool optExtractInitTestIncr(BasicBlock* head,
5373 GenTreePtr* ppIncr);
5375 void optFindNaturalLoops();
5377 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5378 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5379 bool optCanonicalizeLoopNest(unsigned char loopInd);
5381 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5382 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5383 bool optCanonicalizeLoop(unsigned char loopInd);
5385 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5386 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5387 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5388 bool optLoopContains(unsigned l1, unsigned l2);
5390 // Requires "loopInd" to be a valid index into the loop table.
5391 // Updates the loop table by changing loop "loopInd", whose head is required
5392 // to be "from", to be "to". Also performs this transformation for any
5393 // loop nested in "loopInd" that shares the same head as "loopInd".
5394 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5396 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5397 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5398 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5400 // Marks the containsCall information to "lnum" and any parent loops.
5401 void AddContainsCallAllContainingLoops(unsigned lnum);
5402 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5403 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5404 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5405 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5406 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5407 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5409 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5410 // of "from".) Copies the jump destination from "from" to "to".
5411 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5413 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5414 unsigned optLoopDepth(unsigned lnum)
5416 unsigned par = optLoopTable[lnum].lpParent;
5417 if (par == BasicBlock::NOT_IN_LOOP)
5423 return 1 + optLoopDepth(par);
5427 void fgOptWhileLoop(BasicBlock* block);
5429 bool optComputeLoopRep(int constInit,
5432 genTreeOps iterOper,
5434 genTreeOps testOper,
5437 unsigned* iterCount);
5438 #if FEATURE_STACK_FP_X87
5441 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5442 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5443 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5444 #endif // FEATURE_STACK_FP_X87
5447 static fgWalkPreFn optIsVarAssgCB;
5450 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5452 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5454 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5456 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5458 /**************************************************************************
5459 * Optimization conditions
5460 *************************************************************************/
5462 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5463 bool optPentium4(void);
5464 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5465 bool optAvoidIntMult(void);
5470 // The following is the upper limit on how many expressions we'll keep track
5471 // of for the CSE analysis.
5473 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5475 static const int MIN_CSE_COST = 2;
5477 // Keeps tracked cse indices
5478 BitVecTraits* cseTraits;
5481 /* Generic list of nodes - used by the CSE logic */
5489 typedef struct treeLst* treeLstPtr;
5493 treeStmtLst* tslNext;
5494 GenTreePtr tslTree; // tree node
5495 GenTreePtr tslStmt; // statement containing the tree
5496 BasicBlock* tslBlock; // block containing the statement
5499 typedef struct treeStmtLst* treeStmtLstPtr;
5501 // The following logic keeps track of expressions via a simple hash table.
5505 CSEdsc* csdNextInBucket; // used by the hash table
5507 unsigned csdHashValue; // the orginal hashkey
5509 unsigned csdIndex; // 1..optCSECandidateCount
5510 char csdLiveAcrossCall; // 0 or 1
5512 unsigned short csdDefCount; // definition count
5513 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5515 unsigned csdDefWtCnt; // weighted def count
5516 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5518 GenTreePtr csdTree; // treenode containing the 1st occurance
5519 GenTreePtr csdStmt; // stmt containing the 1st occurance
5520 BasicBlock* csdBlock; // block containing the 1st occurance
5522 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5523 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5525 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5526 // number, this will reflect it; otherwise, NoVN.
5529 static const size_t s_optCSEhashSize;
5530 CSEdsc** optCSEhash;
5533 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5535 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5536 // re-numbered with the bound to improve range check elimination
5538 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5539 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5543 CSEdsc* optCSEfindDsc(unsigned index);
5544 void optUnmarkCSE(GenTreePtr tree);
5546 // user defined callback data for the tree walk function optCSE_MaskHelper()
5547 struct optCSE_MaskData
5549 EXPSET_TP CSE_defMask;
5550 EXPSET_TP CSE_useMask;
5553 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5554 static fgWalkPreFn optCSE_MaskHelper;
5556 // This function walks all the node for an given tree
5557 // and return the mask of CSE definitions and uses for the tree
5559 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5561 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5562 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5563 bool optCSE_canSwap(GenTree* tree);
5565 static fgWalkPostFn optPropagateNonCSE;
5566 static fgWalkPreFn optHasNonCSEChild;
5568 static fgWalkPreFn optUnmarkCSEs;
5570 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5571 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5573 void optCleanupCSEs();
5576 void optEnsureClearCSEInfo();
5579 #endif // FEATURE_ANYCSE
5581 #if FEATURE_VALNUM_CSE
5582 /**************************************************************************
5583 * Value Number based CSEs
5584 *************************************************************************/
5587 void optOptimizeValnumCSEs();
5590 void optValnumCSE_Init();
5591 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5592 unsigned optValnumCSE_Locate();
5593 void optValnumCSE_InitDataFlow();
5594 void optValnumCSE_DataFlow();
5595 void optValnumCSE_Availablity();
5596 void optValnumCSE_Heuristic();
5597 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5599 #endif // FEATURE_VALNUM_CSE
5602 bool optDoCSE; // True when we have found a duplicate CSE tree
5603 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5604 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5605 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5606 unsigned optCSEstart; // The first local variable number that is a CSE
5607 unsigned optCSEcount; // The total count of CSE's introduced.
5608 unsigned optCSEweight; // The weight of the current block when we are
5609 // scanning for CSE expressions
5611 bool optIsCSEcandidate(GenTreePtr tree);
5613 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5615 bool lclNumIsTrueCSE(unsigned lclNum) const
5617 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5620 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5622 bool lclNumIsCSE(unsigned lclNum) const
5624 return lvaTable[lclNum].lvIsCSE;
5628 bool optConfigDisableCSE();
5629 bool optConfigDisableCSE2();
5631 void optOptimizeCSEs();
5633 #endif // FEATURE_ANYCSE
5641 unsigned ivaVar; // Variable we are interested in, or -1
5642 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5643 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5644 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5645 callInterf ivaMaskCall; // What kind of calls are there?
5648 static callInterf optCallInterf(GenTreeCall* call);
5651 // VN based copy propagation.
5652 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5653 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5654 LclNumToGenTreePtrStack;
5656 // Kill set to track variables with intervening definitions.
5657 VARSET_TP optCopyPropKillSet;
5659 // Copy propagation functions.
5660 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5661 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5662 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5663 bool optIsSsaLocal(GenTreePtr tree);
5664 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5665 void optVnCopyProp();
5667 /**************************************************************************
5668 * Early value propagation
5669 *************************************************************************/
5675 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5679 static unsigned GetHashCode(SSAName ssaNm)
5681 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5684 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5686 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5690 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5691 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5692 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5693 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5694 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5695 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5697 bool doesMethodHaveFatPointer()
5699 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5702 void setMethodHasFatPointer()
5704 optMethodFlags |= OMF_HAS_FATPOINTER;
5707 void clearMethodHasFatPointer()
5709 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5712 void addFatPointerCandidate(GenTreeCall* call)
5714 setMethodHasFatPointer();
5715 call->SetFatPointerCandidate();
5718 unsigned optMethodFlags;
5720 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5721 // No throughput diff was found with backward walk bound between 3-8.
5722 static const int optEarlyPropRecurBound = 5;
5724 enum class optPropKind
5732 bool gtIsVtableRef(GenTreePtr tree);
5733 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5734 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5735 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5736 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5737 bool optEarlyPropRewriteTree(GenTreePtr tree);
5738 bool optDoEarlyPropForBlock(BasicBlock* block);
5739 bool optDoEarlyPropForFunc();
5740 void optEarlyProp();
5741 void optFoldNullCheck(GenTreePtr tree);
5742 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5745 /**************************************************************************
5746 * Value/Assertion propagation
5747 *************************************************************************/
5749 // Data structures for assertion prop
5750 BitVecTraits* apTraits;
5753 enum optAssertionKind
5770 O1K_CONSTANT_LOOP_BND,
5791 optAssertionKind assertionKind;
5794 unsigned lclNum; // assigned to or property of this local var number
5802 struct AssertionDscOp1
5804 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5811 struct AssertionDscOp2
5813 optOp2Kind kind; // a const or copy assignment
5817 ssize_t iconVal; // integer
5818 unsigned iconFlags; // gtFlags
5820 struct Range // integer subrange
5834 bool IsCheckedBoundArithBound()
5836 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5838 bool IsCheckedBoundBound()
5840 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5842 bool IsConstantBound()
5844 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5845 op1.kind == O1K_CONSTANT_LOOP_BND);
5847 bool IsBoundsCheckNoThrow()
5849 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5852 bool IsCopyAssertion()
5854 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5857 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5859 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5860 a1->op2.kind == a2->op2.kind;
5863 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5865 if (kind == OAK_EQUAL)
5867 return kind2 == OAK_NOT_EQUAL;
5869 else if (kind == OAK_NOT_EQUAL)
5871 return kind2 == OAK_EQUAL;
5876 static ssize_t GetLowerBoundForIntegralType(var_types type)
5896 static ssize_t GetUpperBoundForIntegralType(var_types type)
5920 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5922 if (op1.kind != that->op1.kind)
5926 else if (op1.kind == O1K_ARR_BND)
5929 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5933 return ((vnBased && (op1.vn == that->op1.vn)) ||
5934 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5938 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5940 if (op2.kind != that->op2.kind)
5946 case O2K_IND_CNS_INT:
5948 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5950 case O2K_CONST_LONG:
5951 return (op2.lconVal == that->op2.lconVal);
5953 case O2K_CONST_DOUBLE:
5954 // exact match because of positive and negative zero.
5955 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5957 case O2K_LCLVAR_COPY:
5959 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5960 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5963 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5966 // we will return false
5970 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5976 bool Complementary(AssertionDsc* that, bool vnBased)
5978 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5979 HasSameOp2(that, vnBased);
5982 bool Equals(AssertionDsc* that, bool vnBased)
5984 if (assertionKind != that->assertionKind)
5988 else if (assertionKind == OAK_NO_THROW)
5990 assert(op2.kind == O2K_INVALID);
5991 return HasSameOp1(that, vnBased);
5995 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
6001 static fgWalkPreFn optAddCopiesCallback;
6002 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
6003 unsigned optAddCopyLclNum;
6004 GenTreePtr optAddCopyAsgnNode;
6006 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6007 bool optAssertionPropagated; // set to true if we modified the trees
6008 bool optAssertionPropagatedCurrentStmt;
6010 GenTreePtr optAssertionPropCurrentTree;
6012 AssertionIndex* optComplementaryAssertionMap;
6013 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6014 // using the value of a local var) for each local var
6015 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6016 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6017 AssertionIndex optMaxAssertionCount;
6020 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6021 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6022 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6023 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6024 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6025 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6027 AssertionIndex GetAssertionCount()
6029 return optAssertionCount;
6031 ASSERT_TP* bbJtrueAssertionOut;
6032 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6033 ValueNumToAssertsMap;
6034 ValueNumToAssertsMap* optValueNumToAsserts;
6036 // Assertion prop helpers.
6037 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6038 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6039 void optAssertionInit(bool isLocalProp);
6040 void optAssertionTraitsInit(AssertionIndex assertionCount);
6041 #if LOCAL_ASSERTION_PROP
6042 void optAssertionReset(AssertionIndex limit);
6043 void optAssertionRemove(AssertionIndex index);
6046 // Assertion prop data flow functions.
6047 void optAssertionPropMain();
6048 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6049 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6050 ASSERT_TP* optInitAssertionDataflowFlags();
6051 ASSERT_TP* optComputeAssertionGen();
6053 // Assertion Gen functions.
6054 void optAssertionGen(GenTreePtr tree);
6055 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6056 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6057 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6058 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6059 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6060 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6062 // Assertion creation functions.
6063 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6064 AssertionIndex optCreateAssertion(GenTreePtr op1,
6066 optAssertionKind assertionKind,
6067 AssertionDsc* assertion);
6068 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6070 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6071 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6072 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6074 void optPrintVnAssertionMapping();
6076 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6078 // Used for respective assertion propagations.
6079 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6080 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6081 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6082 bool optAssertionIsNonNull(GenTreePtr op,
6083 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6085 // Used for Relop propagation.
6086 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6087 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6088 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6090 // Assertion prop for lcl var functions.
6091 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6092 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6094 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6095 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6096 const GenTreePtr tree,
6097 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6098 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6100 // Assertion propagation functions.
6101 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6102 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6103 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6104 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6105 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6106 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6107 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6108 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6109 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6110 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6111 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6112 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6114 // Implied assertion functions.
6115 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6116 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6117 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6118 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6121 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6122 void optDebugCheckAssertion(AssertionDsc* assertion);
6123 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6125 void optAddCopies();
6126 #endif // ASSERTION_PROP
6128 /**************************************************************************
6130 *************************************************************************/
6133 struct LoopCloneVisitorInfo
6135 LoopCloneContext* context;
6138 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6139 : context(context), loopNum(loopNum), stmt(nullptr)
6144 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6145 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6146 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6147 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6148 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6149 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6150 void optObtainLoopCloningOpts(LoopCloneContext* context);
6151 bool optIsLoopClonable(unsigned loopInd);
6153 bool optCanCloneLoops();
6156 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6158 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6159 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6160 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6161 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6165 void optInsertLoopCloningStress(BasicBlock* head);
6167 #if COUNT_RANGECHECKS
6168 static unsigned optRangeChkRmv;
6169 static unsigned optRangeChkAll;
6178 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6183 RngChkDsc* rcdNextInBucket; // used by the hash table
6185 unsigned short rcdHashValue; // to make matching faster
6186 unsigned short rcdIndex; // 0..optRngChkCount-1
6188 GenTreePtr rcdTree; // the array index tree
6191 unsigned optRngChkCount;
6192 static const size_t optRngChkHashSize;
6194 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6195 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6197 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6200 bool optLoopsMarked;
6203 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6204 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6208 XX Does the register allocation and puts the remaining lclVars on the stack XX
6210 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6211 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6215 #ifndef LEGACY_BACKEND
6220 #else // LEGACY_BACKEND
6225 #endif // LEGACY_BACKEND
6227 #ifdef LEGACY_BACKEND
6229 void raAssignVars(); // register allocation
6230 #endif // LEGACY_BACKEND
6232 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6234 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6236 void raMarkStkVars();
6239 // Some things are used by both LSRA and regpredict allocators.
6241 FrameType rpFrameType;
6242 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6244 #ifdef LEGACY_BACKEND
6245 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6247 #endif // LEGACY_BACKEND
6249 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6251 #if FEATURE_FP_REGALLOC
6252 enum enumConfigRegisterFP
6254 CONFIG_REGISTER_FP_NONE = 0x0,
6255 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6256 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6257 CONFIG_REGISTER_FP_FULL = 0x3,
6259 enumConfigRegisterFP raConfigRegisterFP();
6260 #endif // FEATURE_FP_REGALLOC
6263 regMaskTP raConfigRestrictMaskFP();
6266 #ifndef LEGACY_BACKEND
6267 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6268 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6269 #else // LEGACY_BACKEND
6270 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6271 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6272 bool raNewBlocks; // True is we added killing blocks for FPU registers
6273 unsigned rpPasses; // Number of passes made by the register predicter
6274 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6275 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6276 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6277 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6278 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6279 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6280 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6281 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6282 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6283 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6284 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6285 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6287 bool rpRegAllocDone; // Set to true after we have completed register allocation
6289 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6291 void raSetupArgMasks(RegState* r);
6293 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6295 void raDumpVarIntf(); // Dump the variable to variable interference graph
6296 void raDumpRegIntf(); // Dump the variable to register interference graph
6298 void raAdjustVarIntf();
6300 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6302 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6304 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6305 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6307 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6309 static fgWalkPreFn rpMarkRegIntf;
6311 regMaskTP rpPredictAddressMode(
6312 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6314 void rpPredictRefAssign(unsigned lclNum);
6316 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6318 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6320 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6322 void rpPredictRegUse(); // Entry point
6324 unsigned raPredictTreeRegUse(GenTreePtr tree);
6325 unsigned raPredictListRegUse(GenTreePtr list);
6327 void raSetRegVarOrder(var_types regType,
6328 regNumber* customVarOrder,
6329 unsigned* customVarOrderSize,
6331 regMaskTP avoidReg);
6333 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6334 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6335 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6336 void raAddToStkPredict(unsigned val)
6338 unsigned newStkPredict = rpStkPredict + val;
6339 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6340 rpStkPredict = UINT_MAX - 1;
6342 rpStkPredict = newStkPredict;
6346 #if !FEATURE_FP_REGALLOC
6347 void raDispFPlifeInfo();
6351 regMaskTP genReturnRegForTree(GenTreePtr tree);
6352 #endif // LEGACY_BACKEND
6354 /* raIsVarargsStackArg is called by raMaskStkVars and by
6355 lvaSortByRefCount. It identifies the special case
6356 where a varargs function has a parameter passed on the
6357 stack, other than the special varargs handle. Such parameters
6358 require special treatment, because they cannot be tracked
6359 by the GC (their offsets in the stack are not known
6363 bool raIsVarargsStackArg(unsigned lclNum)
6367 LclVarDsc* varDsc = &lvaTable[lclNum];
6369 assert(varDsc->lvIsParam);
6371 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6373 #else // _TARGET_X86_
6377 #endif // _TARGET_X86_
6380 #ifdef LEGACY_BACKEND
6381 // Records the current prediction, if it's better than any previous recorded prediction.
6382 void rpRecordPrediction();
6383 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6384 void rpUseRecordedPredictionIfBetter();
6386 // Data members used in the methods above.
6387 unsigned rpBestRecordedStkPredict;
6388 struct VarRegPrediction
6390 bool m_isEnregistered;
6391 regNumberSmall m_regNum;
6392 regNumberSmall m_otherReg;
6394 VarRegPrediction* rpBestRecordedPrediction;
6395 #endif // LEGACY_BACKEND
6398 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6399 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6403 XX Get to the class and method info from the Execution Engine given XX
6404 XX tokens for the class and method XX
6406 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6407 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6411 /* These are the different addressing modes used to access a local var.
6412 * The JIT has to report the location of the locals back to the EE
6413 * for debugging purposes.
6419 VLT_REG_BYREF, // this type is currently only used for value types on X64
6422 VLT_STK_BYREF, // this type is currently only used for value types on X64
6436 siVarLocType vlType;
6439 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6441 // VLT_REG_BYREF -- the specified register contains the address of the variable
6449 // VLT_STK -- Any 32 bit value which is on the stack
6450 // eg. [ESP+0x20], or [EBP-0x28]
6451 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6452 // eg. mov EAX, [ESP+0x20]; [EAX]
6456 regNumber vlsBaseReg;
6457 NATIVE_OFFSET vlsOffset;
6460 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6469 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6470 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6478 regNumber vlrssBaseReg;
6479 NATIVE_OFFSET vlrssOffset;
6483 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6484 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6490 regNumber vlsrsBaseReg;
6491 NATIVE_OFFSET vlsrsOffset;
6497 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6498 // eg 2 DWords at [ESP+0x10]
6502 regNumber vls2BaseReg;
6503 NATIVE_OFFSET vls2Offset;
6506 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6507 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6514 // VLT_FIXED_VA -- fixed argument of a varargs function.
6515 // The argument location depends on the size of the variable
6516 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6517 // location of the first arg. This argument can then be accessed
6518 // relative to the position of the first arg
6522 unsigned vlfvOffset;
6529 void* rpValue; // pointer to the in-process
6530 // location of the value.
6536 bool vlIsInReg(regNumber reg);
6537 bool vlIsOnStk(regNumber reg, signed offset);
6540 /*************************************************************************/
6545 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6546 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6547 CORINFO_CALLINFO_FLAGS flags,
6548 CORINFO_CALL_INFO* pResult);
6549 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6551 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6552 CORINFO_ACCESS_FLAGS flags,
6553 CORINFO_FIELD_INFO* pResult);
6557 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6559 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6561 bool IsSuperPMIException(unsigned code)
6563 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6565 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6566 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6567 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6568 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6569 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6570 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6571 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6572 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6576 case EXCEPTIONCODE_DebugBreakorAV:
6577 case EXCEPTIONCODE_MC:
6578 case EXCEPTIONCODE_LWM:
6579 case EXCEPTIONCODE_SASM:
6580 case EXCEPTIONCODE_SSYM:
6581 case EXCEPTIONCODE_CALLUTILS:
6582 case EXCEPTIONCODE_TYPEUTILS:
6583 case EXCEPTIONCODE_ASSERT:
6590 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6591 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6593 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6594 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6597 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6598 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6599 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6601 // VOM info, method sigs
6603 void eeGetSig(unsigned sigTok,
6604 CORINFO_MODULE_HANDLE scope,
6605 CORINFO_CONTEXT_HANDLE context,
6606 CORINFO_SIG_INFO* retSig);
6608 void eeGetCallSiteSig(unsigned sigTok,
6609 CORINFO_MODULE_HANDLE scope,
6610 CORINFO_CONTEXT_HANDLE context,
6611 CORINFO_SIG_INFO* retSig);
6613 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6615 // Method entry-points, instrs
6617 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6619 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6621 CORINFO_EE_INFO eeInfo;
6622 bool eeInfoInitialized;
6624 CORINFO_EE_INFO* eeGetEEInfo();
6626 // Gets the offset of a SDArray's first element
6627 unsigned eeGetArrayDataOffset(var_types type);
6628 // Gets the offset of a MDArray's first element
6629 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6631 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6633 // Returns the page size for the target machine as reported by the EE.
6634 inline size_t eeGetPageSize()
6636 return eeGetEEInfo()->osPageSize;
6639 // Returns the frame size at which we will generate a loop to probe the stack.
6640 inline size_t getVeryLargeFrameSize()
6643 // The looping probe code is 40 bytes, whereas the straight-line probing for
6644 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6645 // or greater, to generate smaller code.
6646 return 2 * eeGetPageSize();
6648 return 3 * eeGetPageSize();
6652 //------------------------------------------------------------------------
6653 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6655 // It represents Abi and target specific registers for the parameter.
6657 class VirtualStubParamInfo
6660 VirtualStubParamInfo(bool isCoreRTABI)
6662 #if defined(_TARGET_X86_)
6665 #elif defined(_TARGET_AMD64_)
6676 #elif defined(_TARGET_ARM_)
6687 #elif defined(_TARGET_ARM64_)
6691 #error Unsupported or unset target architecture
6694 #ifdef LEGACY_BACKEND
6695 #if defined(_TARGET_X86_)
6696 predict = PREDICT_REG_EAX;
6697 #elif defined(_TARGET_ARM_)
6698 predict = PREDICT_REG_R4;
6700 #error Unsupported or unset target architecture
6702 #endif // LEGACY_BACKEND
6705 regNumber GetReg() const
6710 _regMask_enum GetRegMask() const
6715 #ifdef LEGACY_BACKEND
6716 rpPredictReg GetPredict() const
6724 _regMask_enum regMask;
6726 #ifdef LEGACY_BACKEND
6727 rpPredictReg predict;
6731 VirtualStubParamInfo* virtualStubParamInfo;
6733 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6735 return eeGetEEInfo()->targetAbi == abi;
6738 inline bool generateCFIUnwindCodes()
6740 #ifdef UNIX_AMD64_ABI
6741 return IsTargetAbi(CORINFO_CORERT_ABI);
6749 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6751 // Debugging support - Line number info
6753 void eeGetStmtOffsets();
6755 unsigned eeBoundariesCount;
6757 struct boundariesDsc
6759 UNATIVE_OFFSET nativeIP;
6761 unsigned sourceReason;
6762 } * eeBoundaries; // Boundaries to report to EE
6763 void eeSetLIcount(unsigned count);
6764 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6768 static void eeDispILOffs(IL_OFFSET offs);
6769 static void eeDispLineInfo(const boundariesDsc* line);
6770 void eeDispLineInfos();
6773 // Debugging support - Local var info
6777 unsigned eeVarsCount;
6779 struct VarResultInfo
6781 UNATIVE_OFFSET startOffset;
6782 UNATIVE_OFFSET endOffset;
6786 void eeSetLVcount(unsigned count);
6787 void eeSetLVinfo(unsigned which,
6788 UNATIVE_OFFSET startOffs,
6789 UNATIVE_OFFSET length,
6794 const siVarLoc& loc);
6798 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6799 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6802 // ICorJitInfo wrappers
6804 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6806 void eeAllocUnwindInfo(BYTE* pHotCode,
6812 CorJitFuncKind funcKind);
6814 void eeSetEHcount(unsigned cEH);
6816 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6818 WORD eeGetRelocTypeHint(void* target);
6820 // ICorStaticInfo wrapper functions
6822 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6824 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6826 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6829 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6830 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6831 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6832 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6834 template <typename ParamType>
6835 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6837 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6840 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6842 // Utility functions
6844 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6847 const wchar_t* eeGetCPString(size_t stringHandle);
6850 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6852 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6853 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6855 static fgWalkPreFn CountSharedStaticHelper;
6856 static bool IsSharedStaticHelper(GenTreePtr tree);
6857 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6858 static bool IsGcSafePoint(GenTreePtr tree);
6860 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6861 // returns true/false if 'field' is a Jit Data offset
6862 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6863 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6864 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6866 /*****************************************************************************/
6871 enum TEMP_USAGE_TYPE
6877 static var_types tmpNormalizeType(var_types type);
6878 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6879 void tmpRlsTemp(TempDsc* temp);
6880 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6883 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6884 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6888 bool tmpAllFree() const;
6891 #ifndef LEGACY_BACKEND
6892 void tmpPreAllocateTemps(var_types type, unsigned count);
6893 #endif // !LEGACY_BACKEND
6896 #ifdef LEGACY_BACKEND
6897 unsigned tmpIntSpillMax; // number of int-sized spill temps
6898 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6899 #endif // LEGACY_BACKEND
6901 unsigned tmpCount; // Number of temps
6902 unsigned tmpSize; // Size of all the temps
6905 // Used by RegSet::rsSpillChk()
6906 unsigned tmpGetCount; // Temps which haven't been released yet
6909 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6911 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6912 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6915 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6916 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6920 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6921 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6925 CodeGenInterface* codeGen;
6927 // The following holds information about instr offsets in terms of generated code.
6931 IPmappingDsc* ipmdNext; // next line# record
6932 IL_OFFSETX ipmdILoffsx; // the instr offset
6933 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6934 bool ipmdIsLabel; // Can this code be a branch label?
6937 // Record the instr offset mapping to the generated code
6939 IPmappingDsc* genIPmappingList;
6940 IPmappingDsc* genIPmappingLast;
6942 // Managed RetVal - A side hash table meant to record the mapping from a
6943 // GT_CALL node to its IL offset. This info is used to emit sequence points
6944 // that can be used by debugger to determine the native offset at which the
6945 // managed RetVal will be available.
6947 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6948 // favor of a side table for two reasons: 1) We need IL offset for only those
6949 // GT_CALL nodes (created during importation) that correspond to an IL call and
6950 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6951 // structure and IL offset is needed only when generating debuggable code. Therefore
6952 // it is desirable to avoid memory size penalty in retail scenarios.
6953 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6954 CallSiteILOffsetTable;
6955 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6957 unsigned genReturnLocal; // Local number for the return value when applicable.
6958 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6960 // The following properties are part of CodeGenContext. Getters are provided here for
6961 // convenience and backward compatibility, but the properties can only be set by invoking
6962 // the setter on CodeGenContext directly.
6964 __declspec(property(get = getEmitter)) emitter* genEmitter;
6965 emitter* getEmitter()
6967 return codeGen->getEmitter();
6970 const bool isFramePointerUsed()
6972 return codeGen->isFramePointerUsed();
6975 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6976 bool getInterruptible()
6978 return codeGen->genInterruptible;
6980 void setInterruptible(bool value)
6982 codeGen->setInterruptible(value);
6986 const bool genDoubleAlign()
6988 return codeGen->doDoubleAlign();
6990 DWORD getCanDoubleAlign();
6991 bool shouldDoubleAlign(unsigned refCntStk,
6993 unsigned refCntWtdReg,
6994 unsigned refCntStkParam,
6995 unsigned refCntWtdStkDbl);
6996 #endif // DOUBLE_ALIGN
6998 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6999 bool getFullPtrRegMap()
7001 return codeGen->genFullPtrRegMap;
7003 void setFullPtrRegMap(bool value)
7005 codeGen->setFullPtrRegMap(value);
7008 // Things that MAY belong either in CodeGen or CodeGenContext
7010 #if FEATURE_EH_FUNCLETS
7011 FuncInfoDsc* compFuncInfos;
7012 unsigned short compCurrFuncIdx;
7013 unsigned short compFuncInfoCount;
7015 unsigned short compFuncCount()
7017 assert(fgFuncletsCreated);
7018 return compFuncInfoCount;
7021 #else // !FEATURE_EH_FUNCLETS
7023 // This is a no-op when there are no funclets!
7024 void genUpdateCurrentFunclet(BasicBlock* block)
7029 FuncInfoDsc compFuncInfoRoot;
7031 static const unsigned compCurrFuncIdx = 0;
7033 unsigned short compFuncCount()
7038 #endif // !FEATURE_EH_FUNCLETS
7040 FuncInfoDsc* funCurrentFunc();
7041 void funSetCurrentFunc(unsigned funcIdx);
7042 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7043 unsigned int funGetFuncIdx(BasicBlock* block);
7047 VARSET_TP compCurLife; // current live variables
7048 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7050 template <bool ForCodeGen>
7051 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7053 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7055 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7058 template <bool ForCodeGen>
7059 void compUpdateLife(GenTreePtr tree);
7061 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7062 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7063 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7064 template <bool ForCodeGen>
7065 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7067 template <bool ForCodeGen>
7068 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7070 // Gets a register mask that represent the kill set for a helper call since
7071 // not all JIT Helper calls follow the standard ABI on the target architecture.
7072 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7074 // Gets a register mask that represent the kill set for a NoGC helper call.
7075 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7078 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7079 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7080 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7081 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7082 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7083 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7084 #endif // _TARGET_ARM_
7086 // 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
7088 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7090 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7091 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7092 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7093 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7094 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7095 // for the tracked var indices of the field vars, as in a live var set).
7096 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7098 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7100 if (m_promotedStructDeathVars == nullptr)
7102 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7104 return m_promotedStructDeathVars;
7108 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7109 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7113 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7114 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7117 #if !defined(__GNUC__)
7118 #pragma region Unwind information
7123 // Infrastructure functions: start/stop/reserve/emit.
7126 void unwindBegProlog();
7127 void unwindEndProlog();
7128 void unwindBegEpilog();
7129 void unwindEndEpilog();
7130 void unwindReserve();
7131 void unwindEmit(void* pHotCode, void* pColdCode);
7134 // Specific unwind information functions: called by code generation to indicate a particular
7135 // prolog or epilog unwindable instruction has been generated.
7138 void unwindPush(regNumber reg);
7139 void unwindAllocStack(unsigned size);
7140 void unwindSetFrameReg(regNumber reg, unsigned offset);
7141 void unwindSaveReg(regNumber reg, unsigned offset);
7143 #if defined(_TARGET_ARM_)
7144 void unwindPushMaskInt(regMaskTP mask);
7145 void unwindPushMaskFloat(regMaskTP mask);
7146 void unwindPopMaskInt(regMaskTP mask);
7147 void unwindPopMaskFloat(regMaskTP mask);
7148 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7149 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7150 // called via unwindPadding().
7151 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7152 // instruction and the current location.
7153 #endif // _TARGET_ARM_
7155 #if defined(_TARGET_ARM64_)
7157 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7158 // instruction and the current location.
7159 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7160 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7161 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7162 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7163 void unwindSaveNext(); // unwind code: save_next
7164 void unwindReturn(regNumber reg); // ret lr
7165 #endif // defined(_TARGET_ARM64_)
7168 // Private "helper" functions for the unwind implementation.
7172 #if FEATURE_EH_FUNCLETS
7173 void unwindGetFuncLocations(FuncInfoDsc* func,
7174 bool getHotSectionData,
7175 /* OUT */ emitLocation** ppStartLoc,
7176 /* OUT */ emitLocation** ppEndLoc);
7177 #endif // FEATURE_EH_FUNCLETS
7179 void unwindReserveFunc(FuncInfoDsc* func);
7180 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7182 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7184 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7185 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7187 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7189 #if defined(_TARGET_AMD64_)
7191 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7193 void unwindBegPrologWindows();
7194 void unwindPushWindows(regNumber reg);
7195 void unwindAllocStackWindows(unsigned size);
7196 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7197 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7199 #ifdef UNIX_AMD64_ABI
7200 void unwindBegPrologCFI();
7201 void unwindPushCFI(regNumber reg);
7202 void unwindAllocStackCFI(unsigned size);
7203 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7204 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7205 int mapRegNumToDwarfReg(regNumber reg);
7206 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7207 #endif // UNIX_AMD64_ABI
7208 #elif defined(_TARGET_ARM_)
7210 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7211 void unwindPushPopMaskFloat(regMaskTP mask);
7212 void unwindSplit(FuncInfoDsc* func);
7214 #endif // _TARGET_ARM_
7216 #if !defined(__GNUC__)
7217 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7221 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7222 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7226 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7227 XX that contains the distinguished, well-known SIMD type definitions). XX
7229 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7230 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7233 // Get highest available instruction set for floating point codegen
7234 InstructionSet getFloatingPointInstructionSet()
7236 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7239 return InstructionSet_AVX;
7244 return InstructionSet_SSE3_4;
7248 assert(canUseSSE2());
7249 return InstructionSet_SSE2;
7251 assert(!"getFPInstructionSet() is not implemented for target arch");
7253 return InstructionSet_NONE;
7257 // Get highest available instruction set for SIMD codegen
7258 InstructionSet getSIMDInstructionSet()
7260 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7261 return getFloatingPointInstructionSet();
7263 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7265 return InstructionSet_NONE;
7271 // Should we support SIMD intrinsics?
7274 // Have we identified any SIMD types?
7275 // This is currently used by struct promotion to avoid getting type information for a struct
7276 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7278 bool _usesSIMDTypes;
7279 bool usesSIMDTypes()
7281 return _usesSIMDTypes;
7283 void setUsesSIMDTypes(bool value)
7285 _usesSIMDTypes = value;
7288 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7289 // that require indexed access to the individual fields of the vector, which is not well supported
7290 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7291 unsigned lvaSIMDInitTempVarNum;
7294 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7295 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7296 CORINFO_CLASS_HANDLE SIMDIntHandle;
7297 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7298 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7299 CORINFO_CLASS_HANDLE SIMDShortHandle;
7300 CORINFO_CLASS_HANDLE SIMDByteHandle;
7301 CORINFO_CLASS_HANDLE SIMDLongHandle;
7302 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7303 CORINFO_CLASS_HANDLE SIMDULongHandle;
7304 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7305 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7306 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7307 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7309 // Get the handle for a SIMD type.
7310 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7312 if (simdBaseType == TYP_FLOAT)
7317 return SIMDVector2Handle;
7319 return SIMDVector3Handle;
7321 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7323 return SIMDVector4Handle;
7332 assert(simdType == getSIMDVectorType());
7333 switch (simdBaseType)
7336 return SIMDFloatHandle;
7338 return SIMDDoubleHandle;
7340 return SIMDIntHandle;
7342 return SIMDUShortHandle;
7344 return SIMDUShortHandle;
7346 return SIMDUByteHandle;
7348 return SIMDShortHandle;
7350 return SIMDByteHandle;
7352 return SIMDLongHandle;
7354 return SIMDUIntHandle;
7356 return SIMDULongHandle;
7358 assert(!"Didn't find a class handle for simdType");
7360 return NO_CLASS_HANDLE;
7364 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7365 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7366 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7368 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7369 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7370 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7371 bool isSIMDTypeLocal(GenTree* tree)
7373 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7376 // Returns true if the type of the tree is a byref of TYP_SIMD
7377 bool isAddrOfSIMDType(GenTree* tree)
7379 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7381 switch (tree->OperGet())
7384 return varTypeIsSIMD(tree->gtGetOp1());
7386 case GT_LCL_VAR_ADDR:
7387 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7390 return isSIMDTypeLocal(tree);
7397 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7399 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7400 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7401 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7404 // Returns base type of a TYP_SIMD local.
7405 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7406 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7408 if (isSIMDTypeLocal(tree))
7410 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7416 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7418 return info.compCompHnd->isInSIMDModule(clsHnd);
7421 bool isSIMDClass(typeInfo* pTypeInfo)
7423 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7426 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7427 // if it is not a SIMD type or is an unsupported base type.
7428 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7430 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7432 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7435 // Get SIMD Intrinsic info given the method handle.
7436 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7437 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7438 CORINFO_METHOD_HANDLE methodHnd,
7439 CORINFO_SIG_INFO* sig,
7442 var_types* baseType,
7443 unsigned* sizeBytes);
7445 // Pops and returns GenTree node from importers type stack.
7446 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7447 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7449 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7450 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7452 // Creates a GT_SIMD tree for Select operation
7453 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7455 unsigned simdVectorSize,
7460 // Creates a GT_SIMD tree for Min/Max operation
7461 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7462 CORINFO_CLASS_HANDLE typeHnd,
7464 unsigned simdVectorSize,
7468 // Transforms operands and returns the SIMD intrinsic to be applied on
7469 // transformed operands to obtain given relop result.
7470 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7471 CORINFO_CLASS_HANDLE typeHnd,
7472 unsigned simdVectorSize,
7473 var_types* baseType,
7477 // Creates a GT_SIMD tree for Abs intrinsic.
7478 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7480 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7481 // Transforms operands and returns the SIMD intrinsic to be applied on
7482 // transformed operands to obtain == comparison result.
7483 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7484 unsigned simdVectorSize,
7488 // Transforms operands and returns the SIMD intrinsic to be applied on
7489 // transformed operands to obtain > comparison result.
7490 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7491 unsigned simdVectorSize,
7495 // Transforms operands and returns the SIMD intrinsic to be applied on
7496 // transformed operands to obtain >= comparison result.
7497 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7498 unsigned simdVectorSize,
7502 // Transforms operands and returns the SIMD intrinsic to be applied on
7503 // transformed operands to obtain >= comparison result in case of int32
7504 // and small int base type vectors.
7505 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7506 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7507 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7509 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7510 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7511 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7512 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7513 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7515 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7516 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7517 GenTreePtr newobjThis,
7518 CORINFO_CLASS_HANDLE clsHnd,
7519 CORINFO_METHOD_HANDLE method,
7520 CORINFO_SIG_INFO* sig,
7523 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7525 // Whether SIMD vector occupies part of SIMD register.
7526 // SSE2: vector2f/3f are considered sub register SIMD types.
7527 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7528 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7530 unsigned sizeBytes = 0;
7531 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7532 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7535 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7537 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7540 // Get the type for the hardware SIMD vector.
7541 // This is the maximum SIMD type supported for this target.
7542 var_types getSIMDVectorType()
7544 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7551 assert(canUseSSE2());
7555 assert(!"getSIMDVectorType() unimplemented on target arch");
7560 // Get the size of the SIMD type in bytes
7561 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7563 unsigned sizeBytes = 0;
7564 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7568 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7569 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7571 // Get the the number of elements of basetype of SIMD vector given by its type handle
7572 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7574 // Get preferred alignment of SIMD type.
7575 int getSIMDTypeAlignment(var_types simdType);
7577 // Get the number of bytes in a SIMD Vector for the current compilation.
7578 unsigned getSIMDVectorRegisterByteLength()
7580 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7583 return YMM_REGSIZE_BYTES;
7587 assert(canUseSSE2());
7588 return XMM_REGSIZE_BYTES;
7591 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7596 // The minimum and maximum possible number of bytes in a SIMD vector.
7597 unsigned int maxSIMDStructBytes()
7599 return getSIMDVectorRegisterByteLength();
7601 unsigned int minSIMDStructBytes()
7603 return emitTypeSize(TYP_SIMD8);
7606 #ifdef FEATURE_AVX_SUPPORT
7607 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7608 static const unsigned maxPossibleSIMDStructBytes = 32;
7609 #else // !FEATURE_AVX_SUPPORT
7610 static const unsigned maxPossibleSIMDStructBytes = 16;
7611 #endif // !FEATURE_AVX_SUPPORT
7613 // Returns the codegen type for a given SIMD size.
7614 var_types getSIMDTypeForSize(unsigned size)
7616 var_types simdType = TYP_UNDEF;
7619 simdType = TYP_SIMD8;
7621 else if (size == 12)
7623 simdType = TYP_SIMD12;
7625 else if (size == 16)
7627 simdType = TYP_SIMD16;
7629 #ifdef FEATURE_AVX_SUPPORT
7630 else if (size == 32)
7632 simdType = TYP_SIMD32;
7634 #endif // FEATURE_AVX_SUPPORT
7637 noway_assert(!"Unexpected size for SIMD type");
7642 unsigned getSIMDInitTempVarNum()
7644 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7646 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7647 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7649 return lvaSIMDInitTempVarNum;
7652 #endif // FEATURE_SIMD
7655 //------------------------------------------------------------------------
7656 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7658 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7659 // candidate for enregistration.
7661 unsigned largestEnregisterableStructSize()
7664 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7665 if (vectorRegSize > TARGET_POINTER_SIZE)
7667 return vectorRegSize;
7670 #endif // FEATURE_SIMD
7672 return TARGET_POINTER_SIZE;
7677 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7678 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7679 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7681 // Is this var is of type simd struct?
7682 bool lclVarIsSIMDType(unsigned varNum)
7684 LclVarDsc* varDsc = lvaTable + varNum;
7685 return varDsc->lvIsSIMDType();
7688 // Is this Local node a SIMD local?
7689 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7691 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7694 // Returns true if the TYP_SIMD locals on stack are aligned at their
7695 // preferred byte boundary specified by getSIMDTypeAlignment().
7697 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7698 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7699 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7700 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7701 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7702 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7703 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7706 bool isSIMDTypeLocalAligned(unsigned varNum)
7708 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7709 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7712 int off = lvaFrameAddress(varNum, &ebpBased);
7713 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7714 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7715 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7718 #endif // FEATURE_SIMD
7723 // Whether SSE2 is available
7724 bool canUseSSE2() const
7726 #ifdef _TARGET_XARCH_
7727 return opts.compCanUseSSE2;
7733 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7734 bool CanUseSSE3_4() const
7736 #ifdef _TARGET_XARCH_
7737 return opts.compCanUseSSE3_4;
7743 bool canUseAVX() const
7745 #ifdef FEATURE_AVX_SUPPORT
7746 return opts.compCanUseAVX;
7753 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7754 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7758 XX Generic info about the compilation and the method being compiled. XX
7759 XX It is responsible for driving the other phases. XX
7760 XX It is also responsible for all the memory management. XX
7762 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7763 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7767 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7769 InlineResult* compInlineResult; // The result of importing the inlinee method.
7771 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7772 bool compJmpOpUsed; // Does the method do a JMP
7773 bool compLongUsed; // Does the method use TYP_LONG
7774 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7775 bool compTailCallUsed; // Does the method do a tailcall
7776 bool compLocallocUsed; // Does the method use localloc.
7777 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7778 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7779 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7781 // NOTE: These values are only reliable after
7782 // the importing is completely finished.
7784 #ifdef LEGACY_BACKEND
7785 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7786 // we can iterate over these efficiently.
7789 #if CPU_USES_BLOCK_MOVE
7790 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7794 // State information - which phases have completed?
7795 // These are kept together for easy discoverability
7797 bool bRangeAllowStress;
7798 bool compCodeGenDone;
7799 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7800 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7801 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7802 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7805 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7806 bool fgLocalVarLivenessChanged;
7808 bool compStackProbePrologDone;
7810 #ifndef LEGACY_BACKEND
7812 #endif // !LEGACY_BACKEND
7813 bool compRationalIRForm;
7815 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7817 bool compGeneratingProlog;
7818 bool compGeneratingEpilog;
7819 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7820 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7821 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7822 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7823 bool getNeedsGSSecurityCookie() const
7825 return compNeedsGSSecurityCookie;
7827 void setNeedsGSSecurityCookie()
7829 compNeedsGSSecurityCookie = true;
7832 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7833 // frame layout calculations, this is the level we are currently
7836 //---------------------------- JITing options -----------------------------
7849 JitFlags* jitFlags; // all flags passed from the EE
7850 unsigned compFlags; // method attributes
7852 codeOptimize compCodeOpt; // what type of code optimizations
7856 #ifdef _TARGET_XARCH_
7857 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7858 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7860 #ifdef FEATURE_AVX_SUPPORT
7861 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7862 #endif // FEATURE_AVX_SUPPORT
7863 #endif // _TARGET_XARCH_
7865 // optimize maximally and/or favor speed over size?
7867 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7868 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7869 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7870 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7871 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7873 // Maximun number of locals before turning off the inlining
7874 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7877 unsigned instrCount;
7878 unsigned lvRefCount;
7879 bool compMinOptsIsSet;
7881 bool compMinOptsIsUsed;
7883 inline bool MinOpts()
7885 assert(compMinOptsIsSet);
7886 compMinOptsIsUsed = true;
7889 inline bool IsMinOptsSet()
7891 return compMinOptsIsSet;
7894 inline bool MinOpts()
7898 inline bool IsMinOptsSet()
7900 return compMinOptsIsSet;
7903 inline void SetMinOpts(bool val)
7905 assert(!compMinOptsIsUsed);
7906 assert(!compMinOptsIsSet || (compMinOpts == val));
7908 compMinOptsIsSet = true;
7911 // true if the CLFLG_* for an optimization is set.
7912 inline bool OptEnabled(unsigned optFlag)
7914 return !!(compFlags & optFlag);
7917 #ifdef FEATURE_READYTORUN_COMPILER
7918 inline bool IsReadyToRun()
7920 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7923 inline bool IsReadyToRun()
7929 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7930 // PInvoke transitions inline (e.g. when targeting CoreRT).
7931 inline bool ShouldUsePInvokeHelpers()
7933 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7936 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7938 inline bool IsReversePInvoke()
7940 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7943 // true if we must generate code compatible with JIT32 quirks
7944 inline bool IsJit32Compat()
7946 #if defined(_TARGET_X86_)
7947 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7953 // true if we must generate code compatible with Jit64 quirks
7954 inline bool IsJit64Compat()
7956 #if defined(_TARGET_AMD64_)
7957 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7958 #elif !defined(FEATURE_CORECLR)
7965 bool compScopeInfo; // Generate the LocalVar info ?
7966 bool compDbgCode; // Generate debugger-friendly code?
7967 bool compDbgInfo; // Gather debugging info?
7970 #ifdef PROFILING_SUPPORTED
7971 bool compNoPInvokeInlineCB;
7973 static const bool compNoPInvokeInlineCB;
7977 bool compGcChecks; // Check arguments and return values to ensure they are sane
7978 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7979 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7983 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7984 // to be allocated on the stack.
7985 // It will be set to true in the following cases:
7986 // 1. When the method being compiled has a declarative security
7987 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7988 // This is also the case when we inject a prolog and epilog in the method.
7990 // 2. When the method being compiled has imperative security (i.e. the method
7991 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7993 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7995 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7996 // which gets reported as a GC root to stackwalker.
7997 // (See also ICodeManager::GetAddrOfSecurityObject.)
8002 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
8003 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
8007 #ifdef UNIX_AMD64_ABI
8008 // This flag is indicating if there is a need to align the frame.
8009 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8010 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8011 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8012 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8013 // there are calls and making sure the frame alignment logic is executed.
8014 bool compNeedToAlignFrame;
8015 #endif // UNIX_AMD64_ABI
8017 bool compProcedureSplitting; // Separate cold code from hot code
8019 bool genFPorder; // Preserve FP order (operations are non-commutative)
8020 bool genFPopt; // Can we do frame-pointer-omission optimization?
8021 bool altJit; // True if we are an altjit and are compiling this method
8024 bool optRepeat; // Repeat optimizer phases k times
8028 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8029 bool dspCode; // Display native code generated
8030 bool dspEHTable; // Display the EH table reported to the VM
8031 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8032 bool dspEmit; // Display emitter output
8033 bool dspLines; // Display source-code lines intermixed with native code output
8034 bool dmpHex; // Display raw bytes in hex of native code output
8035 bool varNames; // Display variables names in native code output
8036 bool disAsm; // Display native code as it is generated
8037 bool disAsmSpilled; // Display native code when any register spilling occurs
8038 bool disDiffable; // Makes the Disassembly code 'diff-able'
8039 bool disAsm2; // Display native code after it is generated using external disassembler
8040 bool dspOrder; // Display names of each of the methods that we ngen/jit
8041 bool dspUnwind; // Display the unwind info output
8042 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8043 bool compLongAddress; // Force using large pseudo instructions for long address
8044 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8045 bool dspGCtbls; // Display the GC tables
8049 bool doLateDisasm; // Run the late disassembler
8050 #endif // LATE_DISASM
8052 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8053 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8054 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8055 static const bool dspGCtbls = true;
8058 // We need stack probes to guarantee that we won't trigger a stack overflow
8059 // when calling unmanaged code until they get a chance to set up a frame, because
8060 // the EE will have no idea where it is.
8062 // We will only be doing this currently for hosted environments. Unfortunately
8063 // we need to take care of stubs, so potentially, we will have to do the probes
8064 // for any call. We have a plan for not needing for stubs though
8065 bool compNeedStackProbes;
8067 #ifdef PROFILING_SUPPORTED
8068 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8069 // This option helps make the JIT behave as if it is running under a profiler.
8070 bool compJitELTHookEnabled;
8071 #endif // PROFILING_SUPPORTED
8073 #if FEATURE_TAILCALL_OPT
8074 // Whether opportunistic or implicit tail call optimization is enabled.
8075 bool compTailCallOpt;
8076 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8077 bool compTailCallLoopOpt;
8081 static const bool compUseSoftFP = true;
8082 #else // !ARM_SOFTFP
8083 static const bool compUseSoftFP = false;
8086 GCPollType compGCPollType;
8090 static bool s_pAltJitExcludeAssembliesListInitialized;
8091 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8096 template <typename T>
8099 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8102 template <typename T>
8105 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8108 static int dspTreeID(GenTree* tree)
8110 return tree->gtTreeID;
8112 static void printTreeID(GenTree* tree)
8114 if (tree == nullptr)
8120 printf("[%06d]", dspTreeID(tree));
8127 #define STRESS_MODES \
8131 /* "Variations" stress areas which we try to mix up with each other. */ \
8132 /* These should not be exhaustively used as they might */ \
8133 /* hide/trivialize other areas */ \
8136 STRESS_MODE(DBL_ALN) \
8137 STRESS_MODE(LCL_FLDS) \
8138 STRESS_MODE(UNROLL_LOOPS) \
8139 STRESS_MODE(MAKE_CSE) \
8140 STRESS_MODE(LEGACY_INLINE) \
8141 STRESS_MODE(CLONE_EXPR) \
8142 STRESS_MODE(USE_FCOMI) \
8143 STRESS_MODE(USE_CMOV) \
8145 STRESS_MODE(BB_PROFILE) \
8146 STRESS_MODE(OPT_BOOLS_GC) \
8147 STRESS_MODE(REMORPH_TREES) \
8148 STRESS_MODE(64RSLT_MUL) \
8149 STRESS_MODE(DO_WHILE_LOOPS) \
8150 STRESS_MODE(MIN_OPTS) \
8151 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8152 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8153 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8154 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8155 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8156 STRESS_MODE(NULL_OBJECT_CHECK) \
8157 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8158 STRESS_MODE(RANDOM_INLINE) \
8159 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8160 STRESS_MODE(GENERIC_VARN) \
8162 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8164 STRESS_MODE(COUNT_VARN) \
8166 /* "Check" stress areas that can be exhaustively used if we */ \
8167 /* dont care about performance at all */ \
8169 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8170 STRESS_MODE(CHK_FLOW_UPDATE) \
8171 STRESS_MODE(EMITTER) \
8172 STRESS_MODE(CHK_REIMPORT) \
8173 STRESS_MODE(FLATFP) \
8174 STRESS_MODE(GENERIC_CHECK) \
8179 #define STRESS_MODE(mode) STRESS_##mode,
8186 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8187 BYTE compActiveStressModes[STRESS_COUNT];
8190 #define MAX_STRESS_WEIGHT 100
8192 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8196 bool compInlineStress()
8198 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8201 bool compRandomInlineStress()
8203 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8208 bool compTailCallStress()
8211 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8217 codeOptimize compCodeOpt()
8220 // Switching between size & speed has measurable throughput impact
8221 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8222 // DEBUG, but should generate identical code between CHK & RET builds,
8223 // so that's not acceptable.
8224 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8225 // Investigate the cause of the throughput regression.
8227 return opts.compCodeOpt;
8229 return BLENDED_CODE;
8233 //--------------------- Info about the procedure --------------------------
8237 COMP_HANDLE compCompHnd;
8238 CORINFO_MODULE_HANDLE compScopeHnd;
8239 CORINFO_CLASS_HANDLE compClassHnd;
8240 CORINFO_METHOD_HANDLE compMethodHnd;
8241 CORINFO_METHOD_INFO* compMethodInfo;
8243 BOOL hasCircularClassConstraints;
8244 BOOL hasCircularMethodConstraints;
8246 #if defined(DEBUG) || defined(LATE_DISASM)
8247 const char* compMethodName;
8248 const char* compClassName;
8249 const char* compFullName;
8250 #endif // defined(DEBUG) || defined(LATE_DISASM)
8252 #if defined(DEBUG) || defined(INLINE_DATA)
8253 // Method hash is logcally const, but computed
8255 mutable unsigned compMethodHashPrivate;
8256 unsigned compMethodHash() const;
8257 #endif // defined(DEBUG) || defined(INLINE_DATA)
8259 #ifdef PSEUDORANDOM_NOP_INSERTION
8260 // things for pseudorandom nop insertion
8261 unsigned compChecksum;
8265 // The following holds the FLG_xxxx flags for the method we're compiling.
8268 // The following holds the class attributes for the method we're compiling.
8269 unsigned compClassAttr;
8271 const BYTE* compCode;
8272 IL_OFFSET compILCodeSize; // The IL code size
8273 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8274 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8275 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8276 // (2) the code is hot/cold split, and we issued less code than we expected
8277 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8279 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8280 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8281 bool compIsContextful : 1; // contextful method
8282 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8283 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8284 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8285 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8286 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8288 var_types compRetType; // Return type of the method as declared in IL
8289 var_types compRetNativeType; // Normalized return type as per target arch ABI
8290 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8291 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8293 #if FEATURE_FASTTAILCALL
8294 unsigned compArgRegCount; // Number of incoming integer argument registers used for incoming arguments
8295 unsigned compFloatArgRegCount; // Number of incoming floating argument registers used for incoming arguments
8296 size_t compArgStackSize; // Incoming argument stack size in bytes
8297 #endif // FEATURE_FASTTAILCALL
8299 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8300 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8301 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8302 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8303 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8304 unsigned compMaxStack;
8305 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8306 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8308 unsigned compCallUnmanaged; // count of unmanaged calls
8309 unsigned compLvFrameListRoot; // lclNum for the Frame root
8310 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8311 // You should generally use compHndBBtabCount instead: it is the
8312 // current number of EH clauses (after additions like synchronized
8313 // methods and funclets, and removals like unreachable code deletion).
8315 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8316 // and the VM expects that, or the JIT is a "self-host" compiler
8317 // (e.g., x86 hosted targeting x86) and the VM expects that.
8319 /* The following holds IL scope information about local variables.
8322 unsigned compVarScopesCount;
8323 VarScopeDsc* compVarScopes;
8325 /* The following holds information about instr offsets for
8326 * which we need to report IP-mappings
8329 IL_OFFSET* compStmtOffsets; // sorted
8330 unsigned compStmtOffsetsCount;
8331 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8333 #define CPU_X86 0x0100 // The generic X86 CPU
8334 #define CPU_X86_PENTIUM_4 0x0110
8336 #define CPU_X64 0x0200 // The generic x64 CPU
8337 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8338 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8340 #define CPU_ARM 0x0300 // The generic ARM CPU
8342 unsigned genCPU; // What CPU are we running on
8345 // Returns true if the method being compiled returns a non-void and non-struct value.
8346 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8347 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8348 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8349 // Methods returning such structs are considered to return non-struct return value and
8350 // this method returns true in that case.
8351 bool compMethodReturnsNativeScalarType()
8353 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8356 // Returns true if the method being compiled returns RetBuf addr as its return value
8357 bool compMethodReturnsRetBufAddr()
8359 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8360 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8362 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8363 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8364 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8365 // methods with hidden RetBufArg.
8367 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8368 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8369 // returning the address of RetBuf.
8371 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8372 // to be returned in RAX.
8373 CLANG_FORMAT_COMMENT_ANCHOR;
8375 #ifdef _TARGET_AMD64_
8376 return (info.compRetBuffArg != BAD_VAR_NUM);
8377 #else // !_TARGET_AMD64_
8378 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8379 #endif // !_TARGET_AMD64_
8382 // Returns true if the method returns a value in more than one return register
8383 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8384 // TODO-ARM64: Does this apply for ARM64 too?
8385 bool compMethodReturnsMultiRegRetType()
8387 #if FEATURE_MULTIREG_RET
8388 #if defined(_TARGET_X86_)
8389 // On x86 only 64-bit longs are returned in multiple registers
8390 return varTypeIsLong(info.compRetNativeType);
8391 #else // targets: X64-UNIX, ARM64 or ARM32
8392 // On all other targets that support multireg return values:
8393 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8394 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8395 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8396 #endif // TARGET_XXX
8398 #else // not FEATURE_MULTIREG_RET
8400 // For this architecture there are no multireg returns
8403 #endif // FEATURE_MULTIREG_RET
8406 #if FEATURE_MULTIREG_ARGS
8407 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8408 // return the gcPtr layout for the pointers sized fields
8409 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8410 #endif // FEATURE_MULTIREG_ARGS
8412 // Returns true if the method being compiled returns a value
8413 bool compMethodHasRetVal()
8415 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8416 compMethodReturnsMultiRegRetType();
8421 void compDispLocalVars();
8425 //-------------------------- Global Compiler Data ------------------------------------
8428 static unsigned s_compMethodsCount; // to produce unique label names
8429 unsigned compGenTreeID;
8430 unsigned compBasicBlockID;
8433 BasicBlock* compCurBB; // the current basic block in process
8434 GenTreePtr compCurStmt; // the current statement in process
8436 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8439 // The following is used to create the 'method JIT info' block.
8440 size_t compInfoBlkSize;
8441 BYTE* compInfoBlkAddr;
8443 EHblkDsc* compHndBBtab; // array of EH data
8444 unsigned compHndBBtabCount; // element count of used elements in EH data array
8445 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8447 #if defined(_TARGET_X86_)
8449 //-------------------------------------------------------------------------
8450 // Tracking of region covered by the monitor in synchronized methods
8451 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8452 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8454 #endif // !_TARGET_X86_
8456 Phases previousCompletedPhase; // the most recently completed phase
8458 //-------------------------------------------------------------------------
8459 // The following keeps track of how many bytes of local frame space we've
8460 // grabbed so far in the current function, and how many argument bytes we
8461 // need to pop when we return.
8464 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8466 // Count of callee-saved regs we pushed in the prolog.
8467 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8468 // In case of Amd64 this doesn't include float regs saved on stack.
8469 unsigned compCalleeRegsPushed;
8471 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8472 // Mask of callee saved float regs on stack.
8473 regMaskTP compCalleeFPRegsSavedMask;
8475 #ifdef _TARGET_AMD64_
8476 // Quirk for VS debug-launch scenario to work:
8477 // Bytes of padding between save-reg area and locals.
8478 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8479 unsigned compVSQuirkStackPaddingNeeded;
8480 bool compQuirkForPPPflag;
8483 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8485 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8486 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8487 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8489 //-------------------------------------------------------------------------
8491 static void compStartup(); // One-time initialization
8492 static void compShutdown(); // One-time finalization
8494 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8497 static void compDisplayStaticSizes(FILE* fout);
8499 //------------ Some utility functions --------------
8501 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8502 void** ppIndirection); /* OUT */
8504 // Several JIT/EE interface functions return a CorInfoType, and also return a
8505 // class handle as an out parameter if the type is a value class. Returns the
8506 // size of the type these describe.
8507 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8510 // Components used by the compiler may write unit test suites, and
8511 // have them run within this method. They will be run only once per process, and only
8512 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8513 // These should fail by asserting.
8514 void compDoComponentUnitTestsOnce();
8517 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8518 CORINFO_MODULE_HANDLE classPtr,
8519 COMP_HANDLE compHnd,
8520 CORINFO_METHOD_INFO* methodInfo,
8521 void** methodCodePtr,
8522 ULONG* methodCodeSize,
8523 JitFlags* compileFlags);
8524 void compCompileFinish();
8525 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8526 COMP_HANDLE compHnd,
8527 CORINFO_METHOD_INFO* methodInfo,
8528 void** methodCodePtr,
8529 ULONG* methodCodeSize,
8530 JitFlags* compileFlags,
8531 CorInfoInstantiationVerification instVerInfo);
8533 ArenaAllocator* compGetAllocator();
8535 #if MEASURE_MEM_ALLOC
8537 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8541 unsigned allocCnt; // # of allocs
8542 UINT64 allocSz; // total size of those alloc.
8543 UINT64 allocSzMax; // Maximum single allocation.
8544 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8545 UINT64 nraTotalSizeAlloc;
8546 UINT64 nraTotalSizeUsed;
8548 static const char* s_CompMemKindNames[]; // Names of the kinds.
8550 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8552 for (int i = 0; i < CMK_Count; i++)
8554 allocSzByKind[i] = 0;
8557 MemStats(const MemStats& ms)
8558 : allocCnt(ms.allocCnt)
8559 , allocSz(ms.allocSz)
8560 , allocSzMax(ms.allocSzMax)
8561 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8562 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8564 for (int i = 0; i < CMK_Count; i++)
8566 allocSzByKind[i] = ms.allocSzByKind[i];
8570 // Until we have ubiquitous constructors.
8573 this->MemStats::MemStats();
8576 void AddAlloc(size_t sz, CompMemKind cmk)
8580 if (sz > allocSzMax)
8584 allocSzByKind[cmk] += sz;
8587 void Print(FILE* f); // Print these stats to f.
8588 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8590 MemStats genMemStats;
8592 struct AggregateMemStats : public MemStats
8596 AggregateMemStats() : MemStats(), nMethods(0)
8600 void Add(const MemStats& ms)
8603 allocCnt += ms.allocCnt;
8604 allocSz += ms.allocSz;
8605 allocSzMax = max(allocSzMax, ms.allocSzMax);
8606 for (int i = 0; i < CMK_Count; i++)
8608 allocSzByKind[i] += ms.allocSzByKind[i];
8610 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8611 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8614 void Print(FILE* f); // Print these stats to jitstdout.
8617 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8618 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8619 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8621 #endif // MEASURE_MEM_ALLOC
8623 #if LOOP_HOIST_STATS
8624 unsigned m_loopsConsidered;
8625 bool m_curLoopHasHoistedExpression;
8626 unsigned m_loopsWithHoistedExpressions;
8627 unsigned m_totalHoistedExpressions;
8629 void AddLoopHoistStats();
8630 void PrintPerMethodLoopHoistStats();
8632 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8633 static unsigned s_loopsConsidered;
8634 static unsigned s_loopsWithHoistedExpressions;
8635 static unsigned s_totalHoistedExpressions;
8637 static void PrintAggregateLoopHoistStats(FILE* f);
8638 #endif // LOOP_HOIST_STATS
8640 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8641 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8642 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8643 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8644 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8645 void compFreeMem(void*);
8647 bool compIsForImportOnly();
8648 bool compIsForInlining();
8649 bool compDonotInline();
8652 const char* compLocalVarName(unsigned varNum, unsigned offs);
8653 VarName compVarName(regNumber reg, bool isFloatReg = false);
8654 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8655 const char* compRegPairName(regPairNo regPair);
8656 const char* compRegNameForSize(regNumber reg, size_t size);
8657 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8658 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8659 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8662 //-------------------------------------------------------------------------
8664 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8666 struct VarScopeMapInfo
8668 VarScopeListNode* head;
8669 VarScopeListNode* tail;
8670 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8672 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8679 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8680 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8682 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8683 VarNumToScopeDscMap;
8685 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8686 VarNumToScopeDscMap* compVarScopeMap;
8688 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8690 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8692 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8694 void compInitVarScopeMap();
8696 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8697 // enter scope, sorted by instr offset
8698 unsigned compNextEnterScope;
8700 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8701 // go out of scope, sorted by instr offset
8702 unsigned compNextExitScope;
8704 void compInitScopeLists();
8706 void compResetScopeLists();
8708 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8710 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8712 void compProcessScopesUntil(unsigned offset,
8714 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8715 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8718 void compDispScopeLists();
8721 bool compIsProfilerHookNeeded();
8723 //-------------------------------------------------------------------------
8724 /* Statistical Data Gathering */
8726 void compJitStats(); // call this function and enable
8727 // various ifdef's below for statistical data
8730 void compCallArgStats();
8731 static void compDispCallArgStats(FILE* fout);
8734 //-------------------------------------------------------------------------
8741 ArenaAllocator* compAllocator;
8744 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8745 // suitable for use by utilcode collection types.
8746 IAllocator* compAsIAllocator;
8748 #if MEASURE_MEM_ALLOC
8749 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8750 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8751 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8753 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8755 #endif // MEASURE_MEM_ALLOC
8757 void compFunctionTraceStart();
8758 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8761 size_t compMaxUncheckedOffsetForNullObject;
8763 void compInitOptions(JitFlags* compileFlags);
8765 void compSetProcessor();
8766 void compInitDebuggingInfo();
8767 void compSetOptimizationLevel();
8768 #ifdef _TARGET_ARMARCH_
8769 bool compRsvdRegCheck(FrameLayoutState curState);
8771 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8773 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8774 void ResetOptAnnotations();
8776 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8777 void RecomputeLoopInfo();
8779 #ifdef PROFILING_SUPPORTED
8780 // Data required for generating profiler Enter/Leave/TailCall hooks
8782 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8783 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8784 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8787 #ifdef _TARGET_AMD64_
8788 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8791 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8792 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8794 IAllocator* getAllocator()
8796 return compAsIAllocator;
8799 #if MEASURE_MEM_ALLOC
8800 IAllocator* getAllocatorBitset()
8802 return compAsIAllocatorBitset;
8804 IAllocator* getAllocatorGC()
8806 return compAsIAllocatorGC;
8808 IAllocator* getAllocatorLoopHoist()
8810 return compAsIAllocatorLoopHoist;
8812 #else // !MEASURE_MEM_ALLOC
8813 IAllocator* getAllocatorBitset()
8815 return compAsIAllocator;
8817 IAllocator* getAllocatorGC()
8819 return compAsIAllocator;
8821 IAllocator* getAllocatorLoopHoist()
8823 return compAsIAllocator;
8825 #endif // !MEASURE_MEM_ALLOC
8828 IAllocator* getAllocatorDebugOnly()
8830 #if MEASURE_MEM_ALLOC
8831 return compAsIAllocatorDebugOnly;
8832 #else // !MEASURE_MEM_ALLOC
8833 return compAsIAllocator;
8834 #endif // !MEASURE_MEM_ALLOC
8839 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8840 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8844 XX Checks for type compatibility and merges types XX
8846 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8847 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8851 // Set to TRUE if verification cannot be skipped for this method
8852 // If we detect unverifiable code, we will lazily check
8853 // canSkipMethodVerification() to see if verification is REALLY needed.
8854 BOOL tiVerificationNeeded;
8856 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8857 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8858 BOOL tiIsVerifiableCode;
8860 // Set to TRUE if runtime callout is needed for this method
8861 BOOL tiRuntimeCalloutNeeded;
8863 // Set to TRUE if security prolog/epilog callout is needed for this method
8864 // Note: This flag is different than compNeedSecurityCheck.
8865 // compNeedSecurityCheck means whether or not a security object needs
8866 // to be allocated on the stack, which is currently true for EnC as well.
8867 // tiSecurityCalloutNeeded means whether or not security callouts need
8868 // to be inserted in the jitted code.
8869 BOOL tiSecurityCalloutNeeded;
8871 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8872 // This support is necessary to suport attributes that are not described in
8873 // for example, signatures. For example, the permanent home byref (byref that
8874 // points to the gc heap), isn't a property of method signatures, therefore,
8875 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8876 // but when deciding if we need to reimport a block, we need to take these
8878 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8880 // Returns TRUE if child is equal to or a subtype of parent.
8881 // normalisedForStack indicates that both types are normalised for the stack
8882 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8884 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8885 // *pDest is modified to represent the merged type. Sets "*changed" to true
8886 // if this changes "*pDest".
8887 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8889 // Set pDest from the primitive value type.
8890 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8892 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8895 // <BUGNUM> VSW 471305
8896 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8897 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8898 // We use a "short" as we need to push/pop this scope.
8900 short compRegSetCheckLevel;
8904 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8905 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8907 XX IL verification stuff XX
8910 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8911 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8915 // The following is used to track liveness of local variables, initialization
8916 // of valueclass constructors, and type safe use of IL instructions.
8918 // dynamic state info needed for verification
8919 EntryState verCurrentState;
8921 // this ptr of object type .ctors are considered intited only after
8922 // the base class ctor is called, or an alternate ctor is called.
8923 // An uninited this ptr can be used to access fields, but cannot
8924 // be used to call a member function.
8925 BOOL verTrackObjCtorInitState;
8927 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8929 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8930 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8931 void verInitCurrentState();
8932 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8934 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8935 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8936 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8938 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8939 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8940 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8941 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8942 typeInfo verMakeTypeInfo(CorInfoType ciType,
8943 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8944 BOOL verIsSDArray(typeInfo ti);
8945 typeInfo verGetArrayElemType(typeInfo ti);
8947 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8948 BOOL verNeedsVerification();
8949 BOOL verIsByRefLike(const typeInfo& ti);
8950 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8952 // generic type variables range over types that satisfy IsBoxable
8953 BOOL verIsBoxable(const typeInfo& ti);
8955 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8956 DEBUGARG(unsigned line));
8957 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8958 DEBUGARG(unsigned line));
8959 bool verCheckTailCallConstraint(OPCODE opcode,
8960 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8961 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8962 // on a type parameter?
8963 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8964 // return false to the caller.
8965 // If false, it will throw.
8967 bool verIsBoxedValueType(typeInfo ti);
8969 void verVerifyCall(OPCODE opcode,
8970 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8971 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8973 bool readonlyCall, // is this a "readonly." call?
8974 const BYTE* delegateCreateStart,
8975 const BYTE* codeAddr,
8976 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8978 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8980 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8981 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8982 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8983 const CORINFO_FIELD_INFO& fieldInfo,
8984 const typeInfo* tiThis,
8986 BOOL allowPlainStructAsThis = FALSE);
8987 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8988 void verVerifyThisPtrInitialised();
8989 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8991 // Register allocator
8992 void raInitStackFP();
8993 void raEnregisterVarsPrePassStackFP();
8994 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8995 void raEnregisterVarsPostPassStackFP();
8996 void raGenerateFPRefCounts();
8997 void raEnregisterVarsStackFP();
8998 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
9000 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
9001 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
9003 // returns true if enregistering v1 would save more mem accesses than v2
9004 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
9007 void raDumpHeightsStackFP();
9008 void raDumpVariableRegIntfFloat();
9011 #if FEATURE_STACK_FP_X87
9013 // Currently, we use FP transition blocks in only 2 situations:
9015 // -conditional jump on longs where FP stack differs with target: it's not strictly
9016 // necessary, but its low frequency and the code would get complicated if we try to
9017 // inline the FP stack adjustment, as we have a lot of special casing going on to try
9018 // minimize the way we generate the jump code.
9019 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
9020 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
9022 // However, transition blocks have 2 problems
9024 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9025 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9026 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9027 // in the right place without preordering them), this causes us to have to generate the transition
9028 // blocks in the cold area if we want procedure splitting.
9031 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9032 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9033 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9034 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9035 // a big change in the exception.
9037 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9038 // optimizations. For these 2 cases:
9040 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9041 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9042 // a switch statement.
9044 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9045 // current procedure splitting and exception code have.
9046 bool compMayHaveTransitionBlocks;
9048 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9050 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9052 unsigned raCntStkStackFP;
9053 unsigned raCntWtdStkDblStackFP;
9054 unsigned raCntStkParamDblStackFP;
9056 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9057 // TODO: Do we want to put this in LclVarDsc?
9058 unsigned raPayloadStackFP[lclMAX_TRACKED];
9059 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9061 // Useful for debugging
9062 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9064 #endif // FEATURE_STACK_FP_X87
9067 // One line log function. Default level is 0. Increasing it gives you
9068 // more log information
9070 // levels are currently unused: #define JITDUMP(level,...) ();
9071 void JitLogEE(unsigned level, const char* fmt, ...);
9073 bool compDebugBreak;
9075 bool compJitHaltMethod();
9080 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9081 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9083 XX GS Security checks for unsafe buffers XX
9085 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9086 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9089 struct ShadowParamVarInfo
9091 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9092 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9094 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9096 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9097 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9098 // slots and update all trees to refer to shadow slots is done immediately after
9099 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9100 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9101 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9102 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9103 // creating a shadow slot even though this routine returns true.
9105 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9106 // required. There are two cases under which a reg arg could potentially be used from its
9108 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9109 // b) LSRA spills it
9111 // Possible solution to address case (a)
9112 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9113 // in this routine. Note that live out of exception handler is something we may not be
9114 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9115 // Therefore, for methods with exception handling and need GS cookie check we might have
9116 // to take conservative approach.
9118 // Possible solution to address case (b)
9119 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9120 // create a new spill temp if the method needs GS cookie check.
9121 return varDsc->lvIsParam;
9122 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9123 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9130 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9135 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9136 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9137 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9139 void gsGSChecksInitCookie(); // Grabs cookie variable
9140 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9141 bool gsFindVulnerableParams(); // Shadow param analysis code
9142 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9144 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9145 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9147 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9148 // This can be overwritten by setting complus_JITInlineSize env variable.
9150 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9153 #ifdef FEATURE_JIT_METHOD_PERF
9154 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9155 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9157 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9158 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9160 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9162 #if MEASURE_CLRAPI_CALLS
9163 // Thin wrappers that call into JitTimer (if present).
9164 inline void CLRApiCallEnter(unsigned apix);
9165 inline void CLRApiCallLeave(unsigned apix);
9168 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9169 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9174 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9175 // These variables are associated with maintaining SQM data about compile time.
9176 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9177 // in the current compilation.
9178 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9179 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9180 // the inlining phase in the current compilation.
9181 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9183 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9184 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9185 // type-loading and class initialization).
9186 void RecordStateAtEndOfInlining();
9187 // Assumes being called at the end of compilation. Update the SQM state.
9188 void RecordStateAtEndOfCompilation();
9190 #ifdef FEATURE_CLRSQM
9191 // Does anything SQM related necessary at process shutdown time.
9192 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9193 #endif // FEATURE_CLRSQM
9196 #if FUNC_INFO_LOGGING
9197 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9198 // filename to write it to.
9199 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9200 #endif // FUNC_INFO_LOGGING
9202 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9204 // Is the compilation in a full trust context?
9205 bool compIsFullTrust();
9208 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9209 #endif // MEASURE_NOWAY
9211 #ifndef FEATURE_TRACELOGGING
9212 // Should we actually fire the noway assert body and the exception handler?
9213 bool compShouldThrowOnNoway();
9214 #else // FEATURE_TRACELOGGING
9215 // Should we actually fire the noway assert body and the exception handler?
9216 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9218 // Telemetry instance to use per method compilation.
9219 JitTelemetry compJitTelemetry;
9221 // Get common parameters that have to be logged with most telemetry data.
9222 void compGetTelemetryDefaults(const char** assemblyName,
9223 const char** scopeName,
9224 const char** methodName,
9225 unsigned* methodHash);
9226 #endif // !FEATURE_TRACELOGGING
9230 NodeToTestDataMap* m_nodeTestData;
9232 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9233 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9234 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9235 // Current kept in this.
9237 NodeToTestDataMap* GetNodeTestData()
9239 Compiler* compRoot = impInlineRoot();
9240 if (compRoot->m_nodeTestData == nullptr)
9242 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9244 return compRoot->m_nodeTestData;
9247 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9249 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9250 // currently occur in the AST graph.
9251 NodeToIntMap* FindReachableNodesInNodeTestData();
9253 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9254 // test data, associate that data with "to".
9255 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9257 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9258 // have annotations, attach similar annotations to the corresponding nodes in "to".
9259 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9261 // These are the methods that test that the various conditions implied by the
9262 // test attributes are satisfied.
9263 void JitTestCheckSSA(); // SSA builder tests.
9264 void JitTestCheckVN(); // Value numbering tests.
9267 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9269 FieldSeqStore* m_fieldSeqStore;
9271 FieldSeqStore* GetFieldSeqStore()
9273 Compiler* compRoot = impInlineRoot();
9274 if (compRoot->m_fieldSeqStore == nullptr)
9276 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9277 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9278 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9280 return compRoot->m_fieldSeqStore;
9283 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9285 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9286 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9287 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9288 // attach the field sequence directly to the address node.
9289 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9291 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9293 // Don't need to worry about inlining here
9294 if (m_zeroOffsetFieldMap == nullptr)
9296 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9298 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9299 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9301 return m_zeroOffsetFieldMap;
9304 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9305 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9306 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9307 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9308 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9309 // record the the field sequence using the ZeroOffsetFieldMap described above.
9311 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9312 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9313 // CoreRT. Such case is handled same as the default case.
9314 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9316 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9318 NodeToArrayInfoMap* m_arrayInfoMap;
9320 NodeToArrayInfoMap* GetArrayInfoMap()
9322 Compiler* compRoot = impInlineRoot();
9323 if (compRoot->m_arrayInfoMap == nullptr)
9325 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9326 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9327 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9329 return compRoot->m_arrayInfoMap;
9332 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9334 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9335 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9336 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9337 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9339 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9341 // Use the same map for GCHeap and ByrefExposed when their states match.
9342 memoryKind = ByrefExposed;
9345 assert(memoryKind < MemoryKindCount);
9346 Compiler* compRoot = impInlineRoot();
9347 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9349 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9350 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9351 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9353 return compRoot->m_memorySsaMap[memoryKind];
9356 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9357 CORINFO_CLASS_HANDLE m_refAnyClass;
9358 CORINFO_FIELD_HANDLE GetRefanyDataField()
9360 if (m_refAnyClass == nullptr)
9362 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9364 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9366 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9368 if (m_refAnyClass == nullptr)
9370 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9372 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9376 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9378 #if ALLVARSET_COUNTOPS
9379 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9382 static HelperCallProperties s_helperCallProperties;
9384 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9385 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9386 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9388 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9391 unsigned __int8* offset0,
9392 unsigned __int8* offset1);
9393 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9394 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9396 void fgMorphMultiregStructArgs(GenTreeCall* call);
9397 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9399 }; // end of class Compiler
9401 // Inline methods of CompAllocator.
9402 void* CompAllocator::Alloc(size_t sz)
9404 #if MEASURE_MEM_ALLOC
9405 return m_comp->compGetMem(sz, m_cmk);
9407 return m_comp->compGetMem(sz);
9411 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9413 #if MEASURE_MEM_ALLOC
9414 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9416 return m_comp->compGetMemArray(elems, elemSize);
9420 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9421 inline LclVarDsc::LclVarDsc(Compiler* comp)
9422 : // Initialize the ArgRegs to REG_STK.
9423 // The morph will do the right thing to change
9424 // to the right register if passed in register.
9427 #if FEATURE_MULTIREG_ARGS
9428 _lvOtherArgReg(REG_STK)
9430 #endif // FEATURE_MULTIREG_ARGS
9432 lvRefBlks(BlockSetOps::UninitVal())
9434 #endif // ASSERTION_PROP
9435 lvPerSsaData(comp->getAllocator())
9439 //---------------------------------------------------------------------------------------------------------------------
9440 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9442 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9443 // shown in parentheses):
9445 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9446 // of a misnomer, as the first entry will always be the current node.
9448 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9449 // argument before visiting the node's operands.
9451 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9452 // argument after visiting the node's operands.
9454 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9455 // `DoPreOrder` must be true if this option is true.
9457 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9458 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9459 // visited before the first).
9461 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9463 // A simple pre-order visitor might look something like the following:
9465 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9470 // DoPreOrder = true
9473 // unsigned m_count;
9475 // CountingVisitor(Compiler* compiler)
9476 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9480 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9486 // This visitor would then be used like so:
9488 // CountingVisitor countingVisitor(compiler);
9489 // countingVisitor.WalkTree(root);
9491 template <typename TVisitor>
9492 class GenTreeVisitor
9495 typedef Compiler::fgWalkResult fgWalkResult;
9499 ComputeStack = false,
9501 DoPostOrder = false,
9502 DoLclVarsOnly = false,
9503 UseExecutionOrder = false,
9506 Compiler* m_compiler;
9507 ArrayStack<GenTree*> m_ancestors;
9509 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9511 assert(compiler != nullptr);
9513 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9514 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9517 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9519 return fgWalkResult::WALK_CONTINUE;
9522 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9524 return fgWalkResult::WALK_CONTINUE;
9528 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9530 assert(use != nullptr);
9532 GenTree* node = *use;
9534 if (TVisitor::ComputeStack)
9536 m_ancestors.Push(node);
9539 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9540 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9542 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9543 if (result == fgWalkResult::WALK_ABORT)
9549 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9555 switch (node->OperGet())
9560 case GT_LCL_VAR_ADDR:
9561 case GT_LCL_FLD_ADDR:
9562 if (TVisitor::DoLclVarsOnly)
9564 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9565 if (result == fgWalkResult::WALK_ABORT)
9581 case GT_MEMORYBARRIER:
9586 case GT_START_NONGC:
9588 #if !FEATURE_EH_FUNCLETS
9590 #endif // !FEATURE_EH_FUNCLETS
9592 #ifndef LEGACY_BACKEND
9594 #endif // LEGACY_BACKEND
9597 case GT_CLS_VAR_ADDR:
9601 case GT_PINVOKE_PROLOG:
9602 case GT_PINVOKE_EPILOG:
9606 // Lclvar unary operators
9607 case GT_STORE_LCL_VAR:
9608 case GT_STORE_LCL_FLD:
9609 if (TVisitor::DoLclVarsOnly)
9611 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9612 if (result == fgWalkResult::WALK_ABORT)
9619 // Standard unary operators
9647 GenTreeUnOp* const unOp = node->AsUnOp();
9648 if (unOp->gtOp1 != nullptr)
9650 result = WalkTree(&unOp->gtOp1, unOp);
9651 if (result == fgWalkResult::WALK_ABORT)
9662 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9664 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9665 if (result == fgWalkResult::WALK_ABORT)
9669 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9670 if (result == fgWalkResult::WALK_ABORT)
9674 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9675 if (result == fgWalkResult::WALK_ABORT)
9682 case GT_ARR_BOUNDS_CHECK:
9685 #endif // FEATURE_SIMD
9687 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9689 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9690 if (result == fgWalkResult::WALK_ABORT)
9694 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9695 if (result == fgWalkResult::WALK_ABORT)
9704 GenTreeField* const field = node->AsField();
9706 if (field->gtFldObj != nullptr)
9708 result = WalkTree(&field->gtFldObj, field);
9709 if (result == fgWalkResult::WALK_ABORT)
9719 GenTreeArrElem* const arrElem = node->AsArrElem();
9721 result = WalkTree(&arrElem->gtArrObj, arrElem);
9722 if (result == fgWalkResult::WALK_ABORT)
9727 const unsigned rank = arrElem->gtArrRank;
9728 for (unsigned dim = 0; dim < rank; dim++)
9730 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9731 if (result == fgWalkResult::WALK_ABORT)
9741 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9743 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9744 if (result == fgWalkResult::WALK_ABORT)
9748 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9749 if (result == fgWalkResult::WALK_ABORT)
9753 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9754 if (result == fgWalkResult::WALK_ABORT)
9763 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9765 GenTree** op1Use = &dynBlock->gtOp1;
9766 GenTree** op2Use = &dynBlock->gtDynamicSize;
9768 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9770 std::swap(op1Use, op2Use);
9773 result = WalkTree(op1Use, dynBlock);
9774 if (result == fgWalkResult::WALK_ABORT)
9778 result = WalkTree(op2Use, dynBlock);
9779 if (result == fgWalkResult::WALK_ABORT)
9786 case GT_STORE_DYN_BLK:
9788 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9790 GenTree** op1Use = &dynBlock->gtOp1;
9791 GenTree** op2Use = &dynBlock->gtOp2;
9792 GenTree** op3Use = &dynBlock->gtDynamicSize;
9794 if (TVisitor::UseExecutionOrder)
9796 if (dynBlock->IsReverseOp())
9798 std::swap(op1Use, op2Use);
9800 if (dynBlock->gtEvalSizeFirst)
9802 std::swap(op3Use, op2Use);
9803 std::swap(op2Use, op1Use);
9807 result = WalkTree(op1Use, dynBlock);
9808 if (result == fgWalkResult::WALK_ABORT)
9812 result = WalkTree(op2Use, dynBlock);
9813 if (result == fgWalkResult::WALK_ABORT)
9817 result = WalkTree(op3Use, dynBlock);
9818 if (result == fgWalkResult::WALK_ABORT)
9827 GenTreeCall* const call = node->AsCall();
9829 if (call->gtCallObjp != nullptr)
9831 result = WalkTree(&call->gtCallObjp, call);
9832 if (result == fgWalkResult::WALK_ABORT)
9838 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9840 result = WalkTree(args->pCurrent(), call);
9841 if (result == fgWalkResult::WALK_ABORT)
9847 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9849 result = WalkTree(args->pCurrent(), call);
9850 if (result == fgWalkResult::WALK_ABORT)
9856 if (call->gtCallType == CT_INDIRECT)
9858 if (call->gtCallCookie != nullptr)
9860 result = WalkTree(&call->gtCallCookie, call);
9861 if (result == fgWalkResult::WALK_ABORT)
9867 result = WalkTree(&call->gtCallAddr, call);
9868 if (result == fgWalkResult::WALK_ABORT)
9874 if (call->gtControlExpr != nullptr)
9876 result = WalkTree(&call->gtControlExpr, call);
9877 if (result == fgWalkResult::WALK_ABORT)
9889 assert(node->OperIsBinary());
9891 GenTreeOp* const op = node->AsOp();
9893 GenTree** op1Use = &op->gtOp1;
9894 GenTree** op2Use = &op->gtOp2;
9896 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9898 std::swap(op1Use, op2Use);
9901 if (*op1Use != nullptr)
9903 result = WalkTree(op1Use, op);
9904 if (result == fgWalkResult::WALK_ABORT)
9910 if (*op2Use != nullptr)
9912 result = WalkTree(op2Use, op);
9913 if (result == fgWalkResult::WALK_ABORT)
9923 // Finally, visit the current node
9924 if (TVisitor::DoPostOrder)
9926 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9929 if (TVisitor::ComputeStack)
9938 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9939 class GenericTreeWalker final
9940 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
9945 ComputeStack = computeStack,
9946 DoPreOrder = doPreOrder,
9947 DoPostOrder = doPostOrder,
9948 DoLclVarsOnly = doLclVarsOnly,
9949 UseExecutionOrder = useExecutionOrder,
9953 Compiler::fgWalkData* m_walkData;
9956 GenericTreeWalker(Compiler::fgWalkData* walkData)
9957 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
9959 , m_walkData(walkData)
9961 assert(walkData != nullptr);
9965 walkData->parentStack = &this->m_ancestors;
9969 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9971 m_walkData->parent = user;
9972 return m_walkData->wtprVisitorFn(use, m_walkData);
9975 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9977 m_walkData->parent = user;
9978 return m_walkData->wtpoVisitorFn(use, m_walkData);
9982 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
9988 DoLclVarsOnly = true
9991 IncLclVarRefCountsVisitor(Compiler* compiler);
9992 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9994 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9997 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
10003 DoLclVarsOnly = true
10006 DecLclVarRefCountsVisitor(Compiler* compiler);
10007 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
10009 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10013 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10014 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10016 XX Miscellaneous Compiler stuff XX
10018 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10019 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10022 // Values used to mark the types a stack slot is used for
10024 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10025 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10026 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10027 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10028 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10029 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10030 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10031 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10033 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10035 /*****************************************************************************
10037 * Variables to keep track of total code amounts.
10042 extern size_t grossVMsize;
10043 extern size_t grossNCsize;
10044 extern size_t totalNCsize;
10046 extern unsigned genMethodICnt;
10047 extern unsigned genMethodNCnt;
10048 extern size_t gcHeaderISize;
10049 extern size_t gcPtrMapISize;
10050 extern size_t gcHeaderNSize;
10051 extern size_t gcPtrMapNSize;
10053 #endif // DISPLAY_SIZES
10055 /*****************************************************************************
10057 * Variables to keep track of basic block counts (more data on 1 BB methods)
10060 #if COUNT_BASIC_BLOCKS
10061 extern Histogram bbCntTable;
10062 extern Histogram bbOneBBSizeTable;
10065 /*****************************************************************************
10067 * Used by optFindNaturalLoops to gather statistical information such as
10068 * - total number of natural loops
10069 * - number of loops with 1, 2, ... exit conditions
10070 * - number of loops that have an iterator (for like)
10071 * - number of loops that have a constant iterator
10076 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10077 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10078 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10079 extern unsigned totalLoopCount; // counts the total number of natural loops
10080 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10081 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10082 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10083 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10085 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10086 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10087 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10088 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10089 extern Histogram loopCountTable; // Histogram of loop counts
10090 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10092 #endif // COUNT_LOOPS
10094 /*****************************************************************************
10095 * variables to keep track of how many iterations we go in a dataflow pass
10100 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10101 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10103 #endif // DATAFLOW_ITER
10105 #if MEASURE_BLOCK_SIZE
10106 extern size_t genFlowNodeSize;
10107 extern size_t genFlowNodeCnt;
10108 #endif // MEASURE_BLOCK_SIZE
10110 #if MEASURE_NODE_SIZE
10111 struct NodeSizeStats
10115 genTreeNodeCnt = 0;
10116 genTreeNodeSize = 0;
10117 genTreeNodeActualSize = 0;
10120 // Count of tree nodes allocated.
10121 unsigned __int64 genTreeNodeCnt;
10123 // The size we allocate.
10124 unsigned __int64 genTreeNodeSize;
10126 // The actual size of the node. Note that the actual size will likely be smaller
10127 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10128 // a smaller node to a larger one. TODO-Cleanup: add stats on
10129 // SetOper()/ChangeOper() usage to quantify this.
10130 unsigned __int64 genTreeNodeActualSize;
10132 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10133 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10134 extern Histogram genTreeNcntHist;
10135 extern Histogram genTreeNsizHist;
10136 #endif // MEASURE_NODE_SIZE
10138 /*****************************************************************************
10139 * Count fatal errors (including noway_asserts).
10143 extern unsigned fatal_badCode;
10144 extern unsigned fatal_noWay;
10145 extern unsigned fatal_NOMEM;
10146 extern unsigned fatal_noWayAssertBody;
10148 extern unsigned fatal_noWayAssertBodyArgs;
10150 extern unsigned fatal_NYI;
10151 #endif // MEASURE_FATAL
10153 /*****************************************************************************
10157 #ifdef _TARGET_XARCH_
10159 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10160 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10161 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10163 const instruction INS_AND = INS_and;
10164 const instruction INS_OR = INS_or;
10165 const instruction INS_XOR = INS_xor;
10166 const instruction INS_NEG = INS_neg;
10167 const instruction INS_TEST = INS_test;
10168 const instruction INS_MUL = INS_imul;
10169 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10170 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10171 const instruction INS_BREAKPOINT = INS_int3;
10172 const instruction INS_ADDC = INS_adc;
10173 const instruction INS_SUBC = INS_sbb;
10174 const instruction INS_NOT = INS_not;
10178 #ifdef _TARGET_ARM_
10180 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10181 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10182 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10184 const instruction INS_AND = INS_and;
10185 const instruction INS_OR = INS_orr;
10186 const instruction INS_XOR = INS_eor;
10187 const instruction INS_NEG = INS_rsb;
10188 const instruction INS_TEST = INS_tst;
10189 const instruction INS_MUL = INS_mul;
10190 const instruction INS_MULADD = INS_mla;
10191 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10192 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10193 const instruction INS_BREAKPOINT = INS_bkpt;
10194 const instruction INS_ADDC = INS_adc;
10195 const instruction INS_SUBC = INS_sbc;
10196 const instruction INS_NOT = INS_mvn;
10198 const instruction INS_ABS = INS_vabs;
10199 const instruction INS_ROUND = INS_invalid;
10200 const instruction INS_SQRT = INS_vsqrt;
10204 #ifdef _TARGET_ARM64_
10206 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10207 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10208 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10210 const instruction INS_AND = INS_and;
10211 const instruction INS_OR = INS_orr;
10212 const instruction INS_XOR = INS_eor;
10213 const instruction INS_NEG = INS_neg;
10214 const instruction INS_TEST = INS_tst;
10215 const instruction INS_MUL = INS_mul;
10216 const instruction INS_MULADD = INS_madd;
10217 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10218 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10219 const instruction INS_BREAKPOINT = INS_bkpt;
10220 const instruction INS_ADDC = INS_adc;
10221 const instruction INS_SUBC = INS_sbc;
10222 const instruction INS_NOT = INS_mvn;
10224 const instruction INS_ABS = INS_fabs;
10225 const instruction INS_ROUND = INS_frintn;
10226 const instruction INS_SQRT = INS_fsqrt;
10230 /*****************************************************************************/
10232 extern const BYTE genTypeSizes[];
10233 extern const BYTE genTypeAlignments[];
10234 extern const BYTE genTypeStSzs[];
10235 extern const BYTE genActualTypes[];
10237 /*****************************************************************************/
10239 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10240 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10242 #ifdef _TARGET_ARM_
10243 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10244 #elif defined(_TARGET_ARM64_)
10245 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10248 /*****************************************************************************/
10250 #define REG_CORRUPT regNumber(REG_NA + 1)
10251 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10252 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10254 /*****************************************************************************/
10256 extern BasicBlock dummyBB;
10258 /*****************************************************************************/
10259 /*****************************************************************************/
10261 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10262 // nodes of a statement in execution order.
10263 // __stmt: a GT_STMT type GenTree*
10264 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10266 #define foreach_treenode_execution_order(__node, __stmt) \
10267 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10269 // foreach_block: An iterator over all blocks in the function.
10270 // __compiler: the Compiler* object
10271 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10273 #define foreach_block(__compiler, __block) \
10274 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10276 /*****************************************************************************/
10277 /*****************************************************************************/
10281 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10283 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10284 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10286 XX Debugging helpers XX
10288 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10289 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10292 /*****************************************************************************/
10293 /* The following functions are intended to be called from the debugger, to dump
10294 * various data structures. The can be used in the debugger Watch or Quick Watch
10295 * windows. They are designed to be short to type and take as few arguments as
10296 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10297 * See the function definition comment for more details.
10300 void cBlock(Compiler* comp, BasicBlock* block);
10301 void cBlocks(Compiler* comp);
10302 void cBlocksV(Compiler* comp);
10303 void cTree(Compiler* comp, GenTree* tree);
10304 void cTrees(Compiler* comp);
10305 void cEH(Compiler* comp);
10306 void cVar(Compiler* comp, unsigned lclNum);
10307 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10308 void cVars(Compiler* comp);
10309 void cVarsFinal(Compiler* comp);
10310 void cBlockPreds(Compiler* comp, BasicBlock* block);
10311 void cReach(Compiler* comp);
10312 void cDoms(Compiler* comp);
10313 void cLiveness(Compiler* comp);
10314 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10316 void cFuncIR(Compiler* comp);
10317 void cBlockIR(Compiler* comp, BasicBlock* block);
10318 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10319 void cTreeIR(Compiler* comp, GenTree* tree);
10320 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10321 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10322 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10323 int cOperandIR(Compiler* comp, GenTree* operand);
10324 int cLeafIR(Compiler* comp, GenTree* tree);
10325 int cIndirIR(Compiler* comp, GenTree* tree);
10326 int cListIR(Compiler* comp, GenTree* list);
10327 int cSsaNumIR(Compiler* comp, GenTree* tree);
10328 int cValNumIR(Compiler* comp, GenTree* tree);
10329 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10331 void dBlock(BasicBlock* block);
10334 void dTree(GenTree* tree);
10337 void dVar(unsigned lclNum);
10338 void dVarDsc(LclVarDsc* varDsc);
10341 void dBlockPreds(BasicBlock* block);
10345 void dCVarSet(VARSET_VALARG_TP vars);
10347 void dVarSet(VARSET_VALARG_TP vars);
10348 void dRegMask(regMaskTP mask);
10351 void dBlockIR(BasicBlock* block);
10352 void dTreeIR(GenTree* tree);
10353 void dLoopIR(Compiler::LoopDsc* loop);
10354 void dLoopNumIR(unsigned loopNum);
10355 int dTabStopIR(int curr, int tabstop);
10356 int dTreeTypeIR(GenTree* tree);
10357 int dTreeKindsIR(GenTree* tree);
10358 int dTreeFlagsIR(GenTree* tree);
10359 int dOperandIR(GenTree* operand);
10360 int dLeafIR(GenTree* tree);
10361 int dIndirIR(GenTree* tree);
10362 int dListIR(GenTree* list);
10363 int dSsaNumIR(GenTree* tree);
10364 int dValNumIR(GenTree* tree);
10365 int dDependsIR(GenTree* comma);
10368 GenTree* dFindTree(GenTree* tree, unsigned id);
10369 GenTree* dFindTree(unsigned id);
10370 GenTreeStmt* dFindStmt(unsigned id);
10371 BasicBlock* dFindBlock(unsigned bbNum);
10375 #include "compiler.hpp" // All the shared inline functions
10377 /*****************************************************************************/
10378 #endif //_COMPILER_H_
10379 /*****************************************************************************/