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 unsigned lvSlotNum; // original slot # (if remapped)
727 typeInfo lvVerTypeInfo; // type info needed for verification
729 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
731 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
733 BYTE* lvGcLayout; // GC layout info for structs
736 BlockSet lvRefBlks; // Set of blocks that contain refs
737 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
738 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
740 var_types TypeGet() const
742 return (var_types)lvType;
744 bool lvStackAligned() const
746 assert(lvIsStructField);
747 return ((lvFldOffset % sizeof(void*)) == 0);
749 bool lvNormalizeOnLoad() const
751 return varTypeIsSmall(TypeGet()) &&
752 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
753 (lvIsParam || lvAddrExposed || lvIsStructField);
756 bool lvNormalizeOnStore()
758 return varTypeIsSmall(TypeGet()) &&
759 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
760 !(lvIsParam || lvAddrExposed || lvIsStructField);
763 void lvaResetSortAgainFlag(Compiler* pComp);
764 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
765 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
766 void setPrefReg(regNumber regNum, Compiler* pComp);
767 void addPrefReg(regMaskTP regMask, Compiler* pComp);
768 bool IsFloatRegType() const
770 return isFloatRegType(lvType) || lvIsHfaRegArg();
772 var_types GetHfaType() const
774 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
776 void SetHfaType(var_types type)
778 assert(varTypeIsFloating(type));
779 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
782 #ifndef LEGACY_BACKEND
783 var_types lvaArgType();
786 PerSsaArray lvPerSsaData;
789 // Keep track of the # of SsaNames, for a bounds check.
790 unsigned lvNumSsaNames;
793 // Returns the address of the per-Ssa data for the given ssaNum (which is required
794 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
795 // not an SSA variable).
796 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
798 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
799 assert(SsaConfig::RESERVED_SSA_NUM == 0);
800 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
801 assert(zeroBased < lvNumSsaNames);
802 return &lvPerSsaData.GetRef(zeroBased);
807 void PrintVarReg() const
809 if (isRegPairType(TypeGet()))
811 printf("%s:%s", getRegName(lvOtherReg), // hi32
812 getRegName(lvRegNum)); // lo32
816 printf("%s", getRegName(lvRegNum));
821 }; // class LclVarDsc
824 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
825 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
829 XX The temporary lclVars allocated by the compiler for code generation XX
831 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
832 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
835 /*****************************************************************************
837 * The following keeps track of temporaries allocated in the stack frame
838 * during code-generation (after register allocation). These spill-temps are
839 * only used if we run out of registers while evaluating a tree.
841 * These are different from the more common temps allocated by lvaGrabTemp().
852 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
860 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
864 0); // temps must have a negative number (so they have a different number from all local variables)
865 tdOffs = BAD_TEMP_OFFSET;
869 IMPL_LIMITATION("too many spill temps");
874 bool tdLegalOffset() const
876 return tdOffs != BAD_TEMP_OFFSET;
880 int tdTempOffs() const
882 assert(tdLegalOffset());
885 void tdSetTempOffs(int offs)
888 assert(tdLegalOffset());
890 void tdAdjustTempOffs(int offs)
893 assert(tdLegalOffset());
896 int tdTempNum() const
901 unsigned tdTempSize() const
905 var_types tdTempType() const
911 // interface to hide linearscan implementation from rest of compiler
912 class LinearScanInterface
915 virtual void doLinearScan() = 0;
916 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
917 virtual bool willEnregisterLocalVars() const = 0;
920 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
922 // Information about arrays: their element type and size, and the offset of the first element.
923 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
924 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
925 // for example, in value numbering of array index expressions.
928 var_types m_elemType;
929 CORINFO_CLASS_HANDLE m_elemStructType;
931 unsigned m_elemOffset;
933 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
937 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
938 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
943 // This enumeration names the phases into which we divide compilation. The phases should completely
944 // partition a compilation.
947 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
948 #include "compphases.h"
952 extern const char* PhaseNames[];
953 extern const char* PhaseEnums[];
954 extern const LPCWSTR PhaseShortNames[];
956 // The following enum provides a simple 1:1 mapping to CLR API's
957 enum API_ICorJitInfo_Names
959 #define DEF_CLR_API(name) API_##name,
960 #include "ICorJitInfo_API_names.h"
964 //---------------------------------------------------------------
968 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
969 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
970 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
971 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
972 // by "m_timerFailure" being true.
973 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
976 #ifdef FEATURE_JIT_METHOD_PERF
977 // The string names of the phases.
978 static const char* PhaseNames[];
980 static bool PhaseHasChildren[];
981 static int PhaseParent[];
982 static bool PhaseReportsIRSize[];
984 unsigned m_byteCodeBytes;
985 unsigned __int64 m_totalCycles;
986 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
987 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
988 #if MEASURE_CLRAPI_CALLS
989 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
990 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
993 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
995 // For better documentation, we call EndPhase on
996 // non-leaf phases. We should also call EndPhase on the
997 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
998 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
999 // We add all such "redundant end phase" intervals to this variable below; we print
1000 // it out in a report, so we can verify that it is, indeed, very small. If it ever
1001 // isn't, this means that we're doing something significant between the end of the last
1002 // declared subphase and the end of its parent.
1003 unsigned __int64 m_parentPhaseEndSlop;
1004 bool m_timerFailure;
1006 #if MEASURE_CLRAPI_CALLS
1007 // The following measures the time spent inside each individual CLR API call.
1008 unsigned m_allClrAPIcalls;
1009 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1010 unsigned __int64 m_allClrAPIcycles;
1011 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1012 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1013 #endif // MEASURE_CLRAPI_CALLS
1015 CompTimeInfo(unsigned byteCodeBytes);
1019 #ifdef FEATURE_JIT_METHOD_PERF
1021 #if MEASURE_CLRAPI_CALLS
1022 struct WrapICorJitInfo;
1025 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1026 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1027 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1028 // threads, so it is protected by a lock.
1029 // This class is intended to be used as a singleton type, with only a single instance.
1030 class CompTimeSummaryInfo
1032 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1033 static CritSecObject s_compTimeSummaryLock;
1037 CompTimeInfo m_total;
1038 CompTimeInfo m_maximum;
1040 int m_numFilteredMethods;
1041 CompTimeInfo m_filtered;
1043 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1044 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1045 // If any OS interaction fails, returns 0.0.
1046 double CyclesPerSecond();
1048 // This can use what ever data you want to determine if the value to be added
1049 // belongs in the filtered section (it's always included in the unfiltered section)
1050 bool IncludedInFilteredData(CompTimeInfo& info);
1053 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1054 static CompTimeSummaryInfo s_compTimeSummary;
1056 CompTimeSummaryInfo()
1057 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1061 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1062 // This is thread safe.
1063 void AddInfo(CompTimeInfo& info, bool includePhases);
1065 // Print the summary information to "f".
1066 // This is not thread-safe; assumed to be called by only one thread.
1067 void Print(FILE* f);
1070 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1071 // and when the current phase started. This is intended to be part of a Compilation object. This is
1072 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1076 unsigned __int64 m_start; // Start of the compilation.
1077 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1078 #if MEASURE_CLRAPI_CALLS
1079 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1080 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1081 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1082 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1083 static double s_cyclesPerSec; // Cached for speedier measurements
1086 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1088 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1090 static CritSecObject s_csvLock; // Lock to protect the time log file.
1091 void PrintCsvMethodStats(Compiler* comp);
1094 void* operator new(size_t);
1095 void* operator new[](size_t);
1096 void operator delete(void*);
1097 void operator delete[](void*);
1100 // Initialized the timer instance
1101 JitTimer(unsigned byteCodeSize);
1103 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1105 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1108 static void PrintCsvHeader();
1110 // Ends the current phase (argument is for a redundant check).
1111 void EndPhase(Compiler* compiler, Phases phase);
1113 #if MEASURE_CLRAPI_CALLS
1114 // Start and end a timed CLR API call.
1115 void CLRApiCallEnter(unsigned apix);
1116 void CLRApiCallLeave(unsigned apix);
1117 #endif // MEASURE_CLRAPI_CALLS
1119 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1120 // and adds it to "sum".
1121 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1123 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1124 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1125 // "m_info" to true.
1126 bool GetThreadCycles(unsigned __int64* cycles)
1128 bool res = CycleTimer::GetThreadCyclesS(cycles);
1131 m_info.m_timerFailure = true;
1136 #endif // FEATURE_JIT_METHOD_PERF
1138 //------------------- Function/Funclet info -------------------------------
1139 enum FuncKind : BYTE
1141 FUNC_ROOT, // The main/root function (always id==0)
1142 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1143 FUNC_FILTER, // a funclet associated with an EH filter
1152 BYTE funFlags; // Currently unused, just here for padding
1153 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1154 // funclet. It is only valid if funKind field indicates this is a
1155 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1157 #if defined(_TARGET_AMD64_)
1159 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1160 emitLocation* startLoc;
1161 emitLocation* endLoc;
1162 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1163 emitLocation* coldEndLoc;
1164 UNWIND_INFO unwindHeader;
1165 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1166 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1167 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1168 unsigned unwindCodeSlot;
1170 #ifdef UNIX_AMD64_ABI
1171 jitstd::vector<CFI_CODE>* cfiCodes;
1172 #endif // UNIX_AMD64_ABI
1174 #elif defined(_TARGET_ARMARCH_)
1176 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1177 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1178 // Note: we only have a pointer here instead of the actual object,
1179 // to save memory in the JIT case (compared to the NGEN case),
1180 // where we don't have any cold section.
1181 // Note 2: we currently don't support hot/cold splitting in functions
1182 // with EH, so uwiCold will be NULL for all funclets.
1184 #endif // _TARGET_ARMARCH_
1186 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1187 // that isn't shared between the main function body and funclets.
1190 struct fgArgTabEntry
1193 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1196 otherRegNum = REG_NA;
1197 isStruct = false; // is this a struct arg
1199 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1201 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1203 // it will point at the actual argument in the gtCallLateArgs list.
1204 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1206 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1208 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1210 unsigned numRegs; // Count of number of registers that this argument uses
1212 // A slot is a pointer sized region in the OutArg area.
1213 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1214 unsigned numSlots; // Count of number of slots that this argument uses
1216 unsigned alignment; // 1 or 2 (slots/registers)
1217 unsigned lateArgInx; // index into gtCallLateArgs list
1218 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1220 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1221 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1222 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1223 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1224 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1225 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1226 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1227 // previous arguments.
1228 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1229 // to be on the stack despite its arg list position.
1231 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1232 bool isStruct : 1; // True if this is a struct arg
1234 regNumber otherRegNum; // The (second) register to use when passing this argument.
1236 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1237 #elif !defined(_TARGET_64BIT_)
1238 __declspec(property(get = getIsStruct)) bool isStruct;
1241 return varTypeIsStruct(node);
1243 #endif // !_TARGET_64BIT_
1246 void SetIsHfaRegArg(bool hfaRegArg)
1248 isHfaRegArg = hfaRegArg;
1251 void SetIsBackFilled(bool backFilled)
1253 isBackFilled = backFilled;
1256 bool IsBackFilled() const
1258 return isBackFilled;
1260 #else // !_TARGET_ARM_
1261 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1263 void SetIsHfaRegArg(bool hfaRegArg)
1267 void SetIsBackFilled(bool backFilled)
1271 bool IsBackFilled() const
1275 #endif // !_TARGET_ARM_
1281 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1283 //-------------------------------------------------------------------------
1285 // The class fgArgInfo is used to handle the arguments
1286 // when morphing a GT_CALL node.
1291 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1292 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1293 unsigned argCount; // Updatable arg count value
1294 unsigned nextSlotNum; // Updatable slot count value
1295 unsigned stkLevel; // Stack depth when we make this call (for x86)
1297 #if defined(UNIX_X86_ABI)
1298 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1299 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1300 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1301 // Computed dynamically during codegen, based on stkSizeBytes and the current
1302 // stack level (genStackLevel) when the first stack adjustment is made for
1306 #if FEATURE_FIXED_OUT_ARGS
1307 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1310 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1311 bool hasRegArgs; // true if we have one or more register arguments
1312 bool hasStackArgs; // true if we have one or more stack arguments
1313 bool argsComplete; // marker for state
1314 bool argsSorted; // marker for state
1315 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1318 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1321 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1322 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1324 fgArgTabEntryPtr AddRegArg(
1325 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1327 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1328 fgArgTabEntryPtr AddRegArg(
1335 const bool isStruct,
1336 const regNumber otherRegNum = REG_NA,
1337 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1338 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1340 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1344 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1346 void RemorphReset();
1347 fgArgTabEntryPtr RemorphRegArg(
1348 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1350 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1352 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1354 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1356 void ArgsComplete();
1360 void EvalArgsToTemps();
1362 void RecordStkLevel(unsigned stkLvl);
1363 unsigned RetrieveStkLevel();
1369 fgArgTabEntryPtr* ArgTable()
1373 unsigned GetNextSlotNum()
1383 return hasStackArgs;
1385 bool AreArgsComplete() const
1387 return argsComplete;
1389 #if FEATURE_FIXED_OUT_ARGS
1390 unsigned GetOutArgSize() const
1394 void SetOutArgSize(unsigned newVal)
1396 outArgSize = newVal;
1398 #endif // FEATURE_FIXED_OUT_ARGS
1400 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1402 #if defined(UNIX_X86_ABI)
1403 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1404 #endif // defined(UNIX_X86_ABI)
1407 void SetStkSizeBytes(unsigned newStkSizeBytes)
1409 #if defined(UNIX_X86_ABI)
1410 stkSizeBytes = newStkSizeBytes;
1411 #endif // defined(UNIX_X86_ABI)
1414 #if defined(UNIX_X86_ABI)
1415 unsigned GetStkAlign()
1419 unsigned GetStkSizeBytes() const
1421 return stkSizeBytes;
1423 bool IsStkAlignmentDone() const
1425 return alignmentDone;
1427 void SetStkAlignmentDone()
1429 alignmentDone = true;
1431 #endif // defined(UNIX_X86_ABI)
1433 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1434 GenTreePtr GetLateArg(unsigned argIndex);
1438 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1439 // We have the ability to mark source expressions with "Test Labels."
1440 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1441 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1443 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1446 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1447 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1448 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1449 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1450 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1453 struct TestLabelAndNum
1458 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1463 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1465 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1468 // This class implements the "IAllocator" interface, so that we can use
1469 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1471 class CompAllocator : public IAllocator
1474 #if MEASURE_MEM_ALLOC
1478 CompAllocator(Compiler* comp, CompMemKind cmk)
1480 #if MEASURE_MEM_ALLOC
1486 inline void* Alloc(size_t sz);
1488 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1490 // For the compiler's no-release allocator, free operations are no-ops.
1497 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1498 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1500 XX The big guy. The sections are currently organized as : XX
1502 XX o GenTree and BasicBlock XX
1514 XX o PrologScopeInfo XX
1515 XX o CodeGenerator XX
1520 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1521 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1526 friend class emitter;
1527 friend class UnwindInfo;
1528 friend class UnwindFragmentInfo;
1529 friend class UnwindEpilogInfo;
1530 friend class JitTimer;
1531 friend class LinearScan;
1532 friend class fgArgInfo;
1533 friend class Rationalizer;
1535 friend class Lowering;
1536 friend class CSE_DataFlow;
1537 friend class CSE_Heuristic;
1538 friend class CodeGenInterface;
1539 friend class CodeGen;
1540 friend class LclVarDsc;
1541 friend class TempDsc;
1543 friend class ObjectAllocator;
1545 #ifndef _TARGET_64BIT_
1546 friend class DecomposeLongs;
1547 #endif // !_TARGET_64BIT_
1550 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1551 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1553 XX Misc structs definitions XX
1555 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1556 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1560 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1579 bool dumpIRDataflow;
1580 bool dumpIRBlockHeaders;
1582 LPCWSTR dumpIRPhase;
1583 LPCWSTR dumpIRFormat;
1585 bool shouldUseVerboseTrees();
1586 bool asciiTrees; // If true, dump trees using only ASCII characters
1587 bool shouldDumpASCIITrees();
1588 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1589 bool shouldUseVerboseSsa();
1590 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1591 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1593 const char* VarNameToStr(VarName name)
1598 DWORD expensiveDebugCheckLevel;
1601 #if FEATURE_MULTIREG_RET
1602 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1603 #endif // FEATURE_MULTIREG_RET
1606 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1607 #endif // ARM_SOFTFP
1609 //-------------------------------------------------------------------------
1610 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1611 // HFAs are one to four element structs where each element is the same
1612 // type, either all float or all double. They are treated specially
1613 // in the ARM Procedure Call Standard, specifically, they are passed in
1614 // floating-point registers instead of the general purpose registers.
1617 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1618 bool IsHfa(GenTreePtr tree);
1620 var_types GetHfaType(GenTreePtr tree);
1621 unsigned GetHfaCount(GenTreePtr tree);
1623 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1624 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1626 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1627 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1629 //-------------------------------------------------------------------------
1630 // The following is used for validating format of EH table
1634 typedef struct EHNodeDsc* pEHNodeDsc;
1636 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1637 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1650 EHBlockType ehnBlockType; // kind of EH block
1651 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1652 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1653 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1655 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1656 pEHNodeDsc ehnChild; // leftmost nested block
1658 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1659 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1661 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1662 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1664 inline void ehnSetTryNodeType()
1666 ehnBlockType = TryNode;
1668 inline void ehnSetFilterNodeType()
1670 ehnBlockType = FilterNode;
1672 inline void ehnSetHandlerNodeType()
1674 ehnBlockType = HandlerNode;
1676 inline void ehnSetFinallyNodeType()
1678 ehnBlockType = FinallyNode;
1680 inline void ehnSetFaultNodeType()
1682 ehnBlockType = FaultNode;
1685 inline BOOL ehnIsTryBlock()
1687 return ehnBlockType == TryNode;
1689 inline BOOL ehnIsFilterBlock()
1691 return ehnBlockType == FilterNode;
1693 inline BOOL ehnIsHandlerBlock()
1695 return ehnBlockType == HandlerNode;
1697 inline BOOL ehnIsFinallyBlock()
1699 return ehnBlockType == FinallyNode;
1701 inline BOOL ehnIsFaultBlock()
1703 return ehnBlockType == FaultNode;
1706 // returns true if there is any overlap between the two nodes
1707 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1709 if (node1->ehnStartOffset < node2->ehnStartOffset)
1711 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1715 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1719 // fails with BADCODE if inner is not completely nested inside outer
1720 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1722 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1726 //-------------------------------------------------------------------------
1727 // Exception handling functions
1730 #if !FEATURE_EH_FUNCLETS
1732 bool ehNeedsShadowSPslots()
1734 return (info.compXcptnsCount || opts.compDbgEnC);
1737 // 0 for methods with no EH
1738 // 1 for methods with non-nested EH, or where only the try blocks are nested
1739 // 2 for a method with a catch within a catch
1741 unsigned ehMaxHndNestingCount;
1743 #endif // !FEATURE_EH_FUNCLETS
1745 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1746 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1748 bool bbInCatchHandlerILRange(BasicBlock* blk);
1749 bool bbInFilterILRange(BasicBlock* blk);
1750 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1751 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1752 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1753 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1754 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1756 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1757 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1759 // Returns true if "block" is the start of a try region.
1760 bool bbIsTryBeg(BasicBlock* block);
1762 // Returns true if "block" is the start of a handler or filter region.
1763 bool bbIsHandlerBeg(BasicBlock* block);
1765 // Returns true iff "block" is where control flows if an exception is raised in the
1766 // try region, and sets "*regionIndex" to the index of the try for the handler.
1767 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1768 // block of the filter, but not for the filter's handler.
1769 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1771 bool ehHasCallableHandlers();
1773 // Return the EH descriptor for the given region index.
1774 EHblkDsc* ehGetDsc(unsigned regionIndex);
1776 // Return the EH index given a region descriptor.
1777 unsigned ehGetIndex(EHblkDsc* ehDsc);
1779 // Return the EH descriptor index of the enclosing try, for the given region index.
1780 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1782 // Return the EH descriptor index of the enclosing handler, for the given region index.
1783 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1785 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1786 // block is not in a 'try' region).
1787 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1789 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1790 // if this block is not in a filter or handler region).
1791 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1793 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1794 // nullptr if this block's exceptions propagate to caller).
1795 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1797 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1798 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1799 bool ehIsBlockEHLast(BasicBlock* block);
1801 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1803 // Return the region index of the most nested EH region this block is in.
1804 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1806 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1807 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1809 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1810 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1811 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1812 // (It can never be a filter.)
1813 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1815 // A block has been deleted. Update the EH table appropriately.
1816 void ehUpdateForDeletedBlock(BasicBlock* block);
1818 // Determine whether a block can be deleted while preserving the EH normalization rules.
1819 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1821 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1822 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1824 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1825 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1826 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1827 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1828 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1829 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1830 // lives in a filter.)
1831 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1833 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1834 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1835 // (nullptr if the last block is the last block in the program).
1836 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1837 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1840 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1841 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1842 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1845 #if FEATURE_EH_FUNCLETS
1846 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1847 // if there is a filter that protects a region with a nested EH clause (such as a
1848 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1849 // genFuncletProlog() for more details. However, the VM seems to use it for more
1850 // purposes, maybe including debugging. Until we are sure otherwise, always create
1851 // a PSPSym for functions with any EH.
1852 bool ehNeedsPSPSym() const
1856 #else // _TARGET_X86_
1857 return compHndBBtabCount > 0;
1858 #endif // _TARGET_X86_
1861 bool ehAnyFunclets(); // Are there any funclets in this function?
1862 unsigned ehFuncletCount(); // Return the count of funclets in the function
1864 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1865 #else // !FEATURE_EH_FUNCLETS
1866 bool ehAnyFunclets()
1870 unsigned ehFuncletCount()
1875 unsigned bbThrowIndex(BasicBlock* blk)
1877 return blk->bbTryIndex;
1878 } // Get the index to use as the cache key for sharing throw blocks
1879 #endif // !FEATURE_EH_FUNCLETS
1881 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1882 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1883 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1884 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1885 // convenient to also consider it a predecessor.)
1886 flowList* BlockPredsWithEH(BasicBlock* blk);
1888 // This table is useful for memoization of the method above.
1889 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1891 BlockToFlowListMap* m_blockToEHPreds;
1892 BlockToFlowListMap* GetBlockToEHPreds()
1894 if (m_blockToEHPreds == nullptr)
1896 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1898 return m_blockToEHPreds;
1901 void* ehEmitCookie(BasicBlock* block);
1902 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1904 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1906 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1908 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1910 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1912 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1914 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1916 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1918 void fgAllocEHTable();
1920 void fgRemoveEHTableEntry(unsigned XTnum);
1922 #if FEATURE_EH_FUNCLETS
1924 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1926 #endif // FEATURE_EH_FUNCLETS
1930 #endif // !FEATURE_EH
1932 void fgSortEHTable();
1934 // Causes the EH table to obey some well-formedness conditions, by inserting
1935 // empty BB's when necessary:
1936 // * No block is both the first block of a handler and the first block of a try.
1937 // * No block is the first block of multiple 'try' regions.
1938 // * No block is the last block of multiple EH regions.
1939 void fgNormalizeEH();
1940 bool fgNormalizeEHCase1();
1941 bool fgNormalizeEHCase2();
1942 bool fgNormalizeEHCase3();
1945 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1946 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1947 void fgVerifyHandlerTab();
1948 void fgDispHandlerTab();
1951 bool fgNeedToSortEHTable;
1953 void verInitEHTree(unsigned numEHClauses);
1954 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1955 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1956 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1957 void verCheckNestingLevel(EHNodeDsc* initRoot);
1960 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1961 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1963 XX GenTree and BasicBlock XX
1965 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1967 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1968 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1971 // Functions to create nodes
1972 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1975 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1977 // For binary opers.
1978 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1980 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1982 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1983 var_types type = TYP_I_IMPL,
1984 GenTreePtr op1 = nullptr,
1985 GenTreePtr op2 = nullptr);
1987 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1989 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1991 GenTreePtr gtNewJmpTableNode();
1992 GenTreePtr gtNewIconHandleNode(
1993 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1995 unsigned gtTokenToIconFlags(unsigned token);
1997 GenTreePtr gtNewIconEmbHndNode(void* value,
2000 unsigned handle1 = 0,
2001 void* handle2 = nullptr,
2002 void* compileTimeHandle = nullptr);
2004 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2005 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2006 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2007 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2009 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2011 GenTreePtr gtNewLconNode(__int64 value);
2013 GenTreePtr gtNewDconNode(double value);
2015 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2017 GenTreePtr gtNewZeroConNode(var_types type);
2019 GenTreePtr gtNewOneConNode(var_types type);
2022 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2023 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2026 GenTreeBlk* gtNewBlkOpNode(
2027 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
2029 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2031 GenTree* gtNewPutArgReg(var_types type, GenTreePtr arg, regNumber argReg);
2034 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
2037 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2038 void gtSetObjGcInfo(GenTreeObj* objNode);
2039 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2040 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
2042 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2044 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
2046 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2047 CORINFO_METHOD_HANDLE handle,
2049 GenTreeArgList* args,
2050 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2052 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
2054 GenTreeArgList* args,
2055 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2057 GenTreeCall* gtNewHelperCallNode(unsigned helper,
2060 GenTreeArgList* args = nullptr);
2062 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2065 GenTreeSIMD* gtNewSIMDNode(
2066 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2067 GenTreeSIMD* gtNewSIMDNode(var_types type,
2070 SIMDIntrinsicID simdIntrinsicID,
2073 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2076 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2077 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2078 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2080 GenTreePtr gtNewCodeRef(BasicBlock* block);
2082 GenTreePtr gtNewFieldRef(
2083 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2085 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2087 GenTreeArgList* gtNewArgList(GenTreePtr op);
2088 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2089 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2091 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2092 static fgArgTabEntryPtr gtArgEntryByNode(GenTreeCall* call, GenTreePtr node);
2093 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2094 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2096 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2098 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2100 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2101 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2102 CORINFO_ACCESS_FLAGS access,
2103 CORINFO_FIELD_INFO* pFieldInfo,
2105 CORINFO_CLASS_HANDLE structType,
2108 GenTreePtr gtNewNothingNode();
2110 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2112 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2114 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2116 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2118 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2120 //------------------------------------------------------------------------
2121 // Other GenTree functions
2123 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2125 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2126 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2127 // IntCnses with value `deepVarVal`.
2128 GenTreePtr gtCloneExpr(
2129 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2131 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2132 // `varNum` to int constants with value `varVal`.
2133 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2135 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2138 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2140 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2142 // Returns "true" iff the complexity (not formally defined, but first interpretation
2143 // is #of nodes in subtree) of "tree" is greater than "limit".
2144 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2145 // before they have been set.)
2146 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2148 bool gtCompareTree(GenTree* op1, GenTree* op2);
2150 GenTreePtr gtReverseCond(GenTree* tree);
2152 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2154 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2156 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2158 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2161 unsigned gtHashValue(GenTree* tree);
2163 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2166 void gtPrepareCost(GenTree* tree);
2167 bool gtIsLikelyRegVar(GenTree* tree);
2169 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2171 // Returns true iff the secondNode can be swapped with firstNode.
2172 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2174 unsigned gtSetEvalOrder(GenTree* tree);
2176 #if FEATURE_STACK_FP_X87
2178 void gtComputeFPlvls(GenTreePtr tree);
2179 #endif // FEATURE_STACK_FP_X87
2181 void gtSetStmtInfo(GenTree* stmt);
2183 // Returns "true" iff "node" has any of the side effects in "flags".
2184 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2186 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2187 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2189 // Appends 'expr' in front of 'list'
2190 // 'list' will typically start off as 'nullptr'
2191 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2192 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2194 void gtExtractSideEffList(GenTreePtr expr,
2196 unsigned flags = GTF_SIDE_EFFECT,
2197 bool ignoreRoot = false);
2199 GenTreePtr gtGetThisArg(GenTreeCall* call);
2201 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2202 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2203 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2204 // the given "fldHnd", is such an object pointer.
2205 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2207 // Return true if call is a recursive call; return false otherwise.
2208 // Note when inlining, this looks for calls back to the root method.
2209 bool gtIsRecursiveCall(GenTreeCall* call)
2211 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2214 //-------------------------------------------------------------------------
2216 GenTreePtr gtFoldExpr(GenTreePtr tree);
2219 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2220 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2221 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2222 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2223 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2224 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2225 // optimizations for now.
2226 __attribute__((optnone))
2228 gtFoldExprConst(GenTreePtr tree);
2229 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2230 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2231 bool gtTryRemoveBoxUpstreamEffects(GenTreePtr tree);
2233 //-------------------------------------------------------------------------
2234 // Get the handle, if any.
2235 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2236 // Get the handle, and assert if not found.
2237 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2238 // Get the handle for a ref type.
2239 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2241 //-------------------------------------------------------------------------
2242 // Functions to display the trees
2245 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2247 void gtDispVN(GenTreePtr tree);
2248 void gtDispConst(GenTreePtr tree);
2249 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2250 void gtDispNodeName(GenTreePtr tree);
2251 void gtDispRegVal(GenTreePtr tree);
2263 void gtDispChild(GenTreePtr child,
2264 IndentStack* indentStack,
2266 __in_opt const char* msg = nullptr,
2267 bool topOnly = false);
2268 void gtDispTree(GenTreePtr tree,
2269 IndentStack* indentStack = nullptr,
2270 __in_opt const char* msg = nullptr,
2271 bool topOnly = false,
2272 bool isLIR = false);
2273 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2274 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2275 char* gtGetLclVarName(unsigned lclNum);
2276 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2277 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2278 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2279 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2280 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2281 void gtDispFieldSeq(FieldSeqNode* pfsn);
2283 void gtDispRange(LIR::ReadOnlyRange const& range);
2285 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2287 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2299 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2300 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2303 static fgWalkPreFn gtAssertColonCond;
2305 static fgWalkPreFn gtMarkColonCond;
2306 static fgWalkPreFn gtClearColonCond;
2308 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2309 bool gtHasCatchArg(GenTreePtr tree);
2310 bool gtHasUnmanagedCall(GenTreePtr tree);
2312 typedef ArrayStack<GenTree*> GenTreeStack;
2314 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2315 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2317 //=========================================================================
2318 // BasicBlock functions
2320 // This is a debug flag we will use to assert when creating block during codegen
2321 // as this interferes with procedure splitting. If you know what you're doing, set
2322 // it to true before creating the block. (DEBUG only)
2323 bool fgSafeBasicBlockCreation;
2326 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2329 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2330 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2334 XX The variables to be used by the code generator. XX
2336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2337 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2341 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2342 // be placed in the stack frame and it's fields must be laid out sequentially.
2344 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2345 // a local variable that can be enregistered or placed in the stack frame.
2346 // The fields do not need to be laid out sequentially
2348 enum lvaPromotionType
2350 PROMOTION_TYPE_NONE, // The struct local is not promoted
2351 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2352 // and its field locals are independent of its parent struct local.
2353 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2354 // but its field locals depend on its parent struct local.
2357 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2358 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2360 /*****************************************************************************/
2362 enum FrameLayoutState
2365 INITIAL_FRAME_LAYOUT,
2366 PRE_REGALLOC_FRAME_LAYOUT,
2367 REGALLOC_FRAME_LAYOUT,
2368 TENTATIVE_FRAME_LAYOUT,
2373 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2374 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2375 bool lvaSortAgain; // true: We need to sort the lvaTable
2376 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2377 unsigned lvaCount; // total number of locals
2379 unsigned lvaRefCount; // total number of references to locals
2380 LclVarDsc* lvaTable; // variable descriptor table
2381 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2383 LclVarDsc** lvaRefSorted; // table sorted by refcount
2385 unsigned short lvaTrackedCount; // actual # of locals being tracked
2386 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2388 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2389 // Only for AMD64 System V cache the first caller stack homed argument.
2390 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2391 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2394 VARSET_TP lvaTrackedVars; // set of tracked variables
2396 #ifndef _TARGET_64BIT_
2397 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2399 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2401 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2402 // It that changes, this changes. VarSets from different epochs
2403 // cannot be meaningfully combined.
2405 unsigned GetCurLVEpoch()
2410 // reverse map of tracked number to var number
2411 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2413 #ifdef LEGACY_BACKEND
2414 // variable interference graph
2415 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2418 // variable preference graph
2419 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2423 // # of procs compiled a with double-aligned stack
2424 static unsigned s_lvaDoubleAlignedProcsCount;
2428 // Getters and setters for address-exposed and do-not-enregister local var properties.
2429 bool lvaVarAddrExposed(unsigned varNum);
2430 void lvaSetVarAddrExposed(unsigned varNum);
2431 bool lvaVarDoNotEnregister(unsigned varNum);
2433 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2434 enum DoNotEnregisterReason
2439 DNER_VMNeedsStackAddr,
2440 DNER_LiveInOutOfHandler,
2441 DNER_LiveAcrossUnmanagedCall,
2442 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2443 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2444 DNER_DepField, // It is a field of a dependently promoted struct
2445 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2446 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2447 #if !defined(LEGACY_BACKEND) && !defined(_TARGET_64BIT_)
2448 DNER_LongParamField, // It is a decomposed field of a long parameter.
2450 #ifdef JIT32_GCENCODER
2455 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2457 unsigned lvaVarargsHandleArg;
2459 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2461 #endif // _TARGET_X86_
2463 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2464 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2465 #if FEATURE_FIXED_OUT_ARGS
2466 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2468 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2469 // that tracks whether the lock has been taken
2471 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2472 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2473 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2475 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2476 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2478 #if FEATURE_FIXED_OUT_ARGS
2479 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2480 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2481 #endif // FEATURE_FIXED_OUT_ARGS
2484 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2485 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2486 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2487 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2488 // this variable to be this scratch word whenever struct promotion occurs.
2489 unsigned lvaPromotedStructAssemblyScratchVar;
2490 #endif // _TARGET_ARM_
2493 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2494 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2497 unsigned lvaGenericsContextUseCount;
2499 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2500 // CORINFO_GENERICS_CTXT_FROM_THIS?
2501 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2503 //-------------------------------------------------------------------------
2504 // All these frame offsets are inter-related and must be kept in sync
2506 #if !FEATURE_EH_FUNCLETS
2507 // This is used for the callable handlers
2508 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2509 #endif // FEATURE_EH_FUNCLETS
2511 unsigned lvaCachedGenericContextArgOffs;
2512 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2515 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2517 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2519 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2520 // after the reg predict we will use a computed maxTmpSize
2521 // which is based upon the number of spill temps predicted by reg predict
2522 // All this is necessary because if we under-estimate the size of the spill
2523 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2525 // Pre codegen max spill temp size.
2526 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2528 //-------------------------------------------------------------------------
2530 unsigned lvaGetMaxSpillTempSize();
2532 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2533 #endif // _TARGET_ARM_
2534 void lvaAssignFrameOffsets(FrameLayoutState curState);
2535 void lvaFixVirtualFrameOffsets();
2537 #ifndef LEGACY_BACKEND
2538 void lvaUpdateArgsWithInitialReg();
2539 #endif // !LEGACY_BACKEND
2541 void lvaAssignVirtualFrameOffsetsToArgs();
2542 #ifdef UNIX_AMD64_ABI
2543 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2544 #else // !UNIX_AMD64_ABI
2545 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2546 #endif // !UNIX_AMD64_ABI
2547 void lvaAssignVirtualFrameOffsetsToLocals();
2548 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2549 #ifdef _TARGET_AMD64_
2550 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2551 bool lvaIsCalleeSavedIntRegCountEven();
2553 void lvaAlignFrame();
2554 void lvaAssignFrameOffsetsToPromotedStructs();
2555 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2558 void lvaDumpRegLocation(unsigned lclNum);
2559 void lvaDumpFrameLocation(unsigned lclNum);
2560 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2561 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2562 // layout state defined by lvaDoneFrameLayout
2565 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2566 // to avoid bugs from borderline cases.
2567 #define MAX_FrameSize 0x3FFFFFFF
2568 void lvaIncrementFrameSize(unsigned size);
2570 unsigned lvaFrameSize(FrameLayoutState curState);
2572 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2573 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2575 // Returns the caller-SP-relative offset for the local variable "varNum."
2576 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2578 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2579 int lvaGetSPRelativeOffset(unsigned varNum);
2581 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2582 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2584 //------------------------ For splitting types ----------------------------
2586 void lvaInitTypeRef();
2588 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2589 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2590 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2591 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2592 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2593 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2595 void lvaInitVarDsc(LclVarDsc* varDsc,
2597 CorInfoType corInfoType,
2598 CORINFO_CLASS_HANDLE typeHnd,
2599 CORINFO_ARG_LIST_HANDLE varList,
2600 CORINFO_SIG_INFO* varSig);
2602 static unsigned lvaTypeRefMask(var_types type);
2604 var_types lvaGetActualType(unsigned lclNum);
2605 var_types lvaGetRealType(unsigned lclNum);
2607 //-------------------------------------------------------------------------
2611 unsigned lvaLclSize(unsigned varNum);
2612 unsigned lvaLclExactSize(unsigned varNum);
2614 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2616 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2617 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2618 // the return result.
2619 bool lvaLclVarRefsAccum(
2620 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2622 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2623 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2624 // and (destructively) unions "trkedVars" into "*result".
2625 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2627 ALLVARSET_VALARG_TP allVars,
2628 VARSET_VALARG_TP trkdVars);
2630 bool lvaHaveManyLocals() const;
2632 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2633 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2634 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2637 void lvaSortByRefCount();
2638 void lvaDumpRefCounts();
2640 void lvaMarkLocalVars(BasicBlock* block);
2642 void lvaMarkLocalVars(); // Local variable ref-counting
2644 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2646 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2648 void lvaIncRefCnts(GenTreePtr tree);
2649 void lvaDecRefCnts(GenTreePtr tree);
2651 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2652 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2653 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2656 struct lvaStressLclFldArgs
2658 Compiler* m_pCompiler;
2662 static fgWalkPreFn lvaStressLclFldCB;
2663 void lvaStressLclFld();
2665 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2666 void lvaDispVarSet(VARSET_VALARG_TP set);
2671 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2673 int lvaFrameAddress(int varNum, bool* pFPbased);
2676 bool lvaIsParameter(unsigned varNum);
2677 bool lvaIsRegArgument(unsigned varNum);
2678 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2679 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2680 // that writes to arg0
2682 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2683 // (this is an overload of lvIsTemp because there are no temp parameters).
2684 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2685 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2686 bool lvaIsImplicitByRefLocal(unsigned varNum)
2688 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2689 LclVarDsc* varDsc = &(lvaTable[varNum]);
2690 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2692 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2695 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2699 // Returns true if this local var is a multireg struct
2700 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2702 // If the local is a TYP_STRUCT, get/set a class handle describing it
2703 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2704 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2706 // If the local is TYP_REF, set or update the associated class information.
2707 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2708 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2709 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2710 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2712 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2714 // Info about struct fields
2715 struct lvaStructFieldInfo
2717 CORINFO_FIELD_HANDLE fldHnd;
2718 unsigned char fldOffset;
2719 unsigned char fldOrdinal;
2722 CORINFO_CLASS_HANDLE fldTypeHnd;
2725 // Info about struct to be promoted.
2726 struct lvaStructPromotionInfo
2728 CORINFO_CLASS_HANDLE typeHnd;
2730 bool requiresScratchVar;
2733 unsigned char fieldCnt;
2734 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2736 lvaStructPromotionInfo()
2737 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2742 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2743 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2744 lvaStructPromotionInfo* StructPromotionInfo,
2746 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2747 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2748 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2749 #if !defined(_TARGET_64BIT_)
2750 void lvaPromoteLongVars();
2751 #endif // !defined(_TARGET_64BIT_)
2752 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2753 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2754 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2755 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2756 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2757 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2758 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2760 #if defined(FEATURE_SIMD)
2761 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2763 assert(varDsc->lvType == TYP_SIMD12);
2764 assert(varDsc->lvExactSize == 12);
2766 #if defined(_TARGET_64BIT_)
2767 assert(varDsc->lvSize() == 16);
2768 #endif // defined(_TARGET_64BIT_)
2770 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2771 // already does this calculation. However, we also need to prevent mapping types if the var is a
2772 // dependently promoted struct field, which must remain its exact size within its parent struct.
2773 // However, we don't know this until late, so we may have already pretended the field is bigger
2775 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2784 #endif // defined(FEATURE_SIMD)
2786 BYTE* lvaGetGcLayout(unsigned varNum);
2787 bool lvaTypeIsGC(unsigned varNum);
2788 unsigned lvaGSSecurityCookie; // LclVar number
2789 bool lvaTempsHaveLargerOffsetThanVars();
2791 unsigned lvaSecurityObject; // variable representing the security object on the stack
2792 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2794 #if FEATURE_EH_FUNCLETS
2795 unsigned lvaPSPSym; // variable representing the PSPSym
2798 InlineInfo* impInlineInfo;
2799 InlineStrategy* m_inlineStrategy;
2801 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2802 Compiler* impInlineRoot();
2804 #if defined(DEBUG) || defined(INLINE_DATA)
2805 unsigned __int64 getInlineCycleCount()
2807 return m_compCycles;
2809 #endif // defined(DEBUG) || defined(INLINE_DATA)
2811 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2812 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2814 //=========================================================================
2816 //=========================================================================
2819 //---------------- Local variable ref-counting ----------------------------
2822 BasicBlock* lvaMarkRefsCurBlock;
2823 GenTreePtr lvaMarkRefsCurStmt;
2825 BasicBlock::weight_t lvaMarkRefsWeight;
2827 void lvaMarkLclRefs(GenTreePtr tree);
2829 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2830 void SetVolatileHint(LclVarDsc* varDsc);
2832 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2833 PerSsaArray lvMemoryPerSsaData;
2834 unsigned lvMemoryNumSsaNames;
2837 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2838 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2839 // not an SSA variable).
2840 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2842 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2843 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2845 assert(ssaNum < lvMemoryNumSsaNames);
2846 return &lvMemoryPerSsaData.GetRef(ssaNum);
2850 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2851 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2855 XX Imports the given method and converts it to semantic trees XX
2857 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2858 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2864 void impImport(BasicBlock* method);
2866 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2867 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2868 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2869 CORINFO_CLASS_HANDLE impGetStringClass();
2870 CORINFO_CLASS_HANDLE impGetObjectClass();
2872 //=========================================================================
2874 //=========================================================================
2877 //-------------------- Stack manipulation ---------------------------------
2879 unsigned impStkSize; // Size of the full stack
2881 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2883 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2885 struct SavedStack // used to save/restore stack contents.
2887 unsigned ssDepth; // number of values on stack
2888 StackEntry* ssTrees; // saved tree values
2891 bool impIsPrimitive(CorInfoType type);
2892 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2894 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2896 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2897 void impPushNullObjRefOnStack();
2898 StackEntry impPopStack();
2899 StackEntry& impStackTop(unsigned n = 0);
2900 unsigned impStackHeight();
2902 void impSaveStackState(SavedStack* savePtr, bool copy);
2903 void impRestoreStackState(SavedStack* savePtr);
2905 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2906 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2907 CORINFO_CALL_INFO* pCallInfo);
2909 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2911 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2913 bool impCanPInvokeInline();
2914 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2915 void impCheckForPInvokeCall(
2916 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2917 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2918 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2920 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2921 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2922 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2924 var_types impImportCall(OPCODE opcode,
2925 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2926 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2928 GenTreePtr newobjThis,
2930 CORINFO_CALL_INFO* callInfo,
2931 IL_OFFSET rawILOffset);
2933 void impDevirtualizeCall(GenTreeCall* call,
2935 CORINFO_METHOD_HANDLE* method,
2936 unsigned* methodFlags,
2937 CORINFO_CONTEXT_HANDLE* contextHandle,
2938 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2940 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2942 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2944 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2947 var_types impImportJitTestLabelMark(int numArgs);
2950 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2952 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2954 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2955 CORINFO_ACCESS_FLAGS access,
2956 CORINFO_FIELD_INFO* pFieldInfo,
2959 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2961 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2963 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2965 void impImportLeave(BasicBlock* block);
2966 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2967 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2968 CORINFO_CLASS_HANDLE clsHnd,
2969 CORINFO_METHOD_HANDLE method,
2970 CORINFO_SIG_INFO* sig,
2974 CorInfoIntrinsics* pIntrinsicID);
2975 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2976 CORINFO_SIG_INFO* sig,
2979 CorInfoIntrinsics intrinsicID);
2980 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2982 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2984 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2985 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2986 CORINFO_THIS_TRANSFORM transform);
2988 //----------------- Manipulating the trees and stmts ----------------------
2990 GenTreePtr impTreeList; // Trees for the BB being imported
2991 GenTreePtr impTreeLast; // The last tree for the current BB
2995 CHECK_SPILL_ALL = -1,
2996 CHECK_SPILL_NONE = -2
3000 void impBeginTreeList();
3001 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
3002 void impEndTreeList(BasicBlock* block);
3003 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3004 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3005 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3006 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3007 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3008 void impAssignTempGen(unsigned tmp,
3011 GenTreePtr* pAfterStmt = nullptr,
3012 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3013 BasicBlock* block = nullptr);
3014 void impAssignTempGen(unsigned tmpNum,
3016 CORINFO_CLASS_HANDLE structHnd,
3018 GenTreePtr* pAfterStmt = nullptr,
3019 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3020 BasicBlock* block = nullptr);
3021 GenTreePtr impCloneExpr(GenTreePtr tree,
3023 CORINFO_CLASS_HANDLE structHnd,
3025 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3026 GenTreePtr impAssignStruct(GenTreePtr dest,
3028 CORINFO_CLASS_HANDLE structHnd,
3030 GenTreePtr* pAfterStmt = nullptr,
3031 BasicBlock* block = nullptr);
3032 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3034 CORINFO_CLASS_HANDLE structHnd,
3036 GenTreePtr* pAfterStmt = nullptr,
3037 BasicBlock* block = nullptr);
3039 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3040 CORINFO_CLASS_HANDLE structHnd,
3044 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3045 BYTE* gcLayout = nullptr,
3046 unsigned* numGCVars = nullptr,
3047 var_types* simdBaseType = nullptr);
3049 GenTreePtr impNormStructVal(GenTreePtr structVal,
3050 CORINFO_CLASS_HANDLE structHnd,
3052 bool forceNormalization = false);
3054 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3055 BOOL* pRuntimeLookup = nullptr,
3056 BOOL mustRestoreHandle = FALSE,
3057 BOOL importParent = FALSE);
3059 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3060 BOOL* pRuntimeLookup = nullptr,
3061 BOOL mustRestoreHandle = FALSE)
3063 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3066 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3067 CORINFO_LOOKUP* pLookup,
3069 void* compileTimeHandle);
3071 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3073 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3074 CORINFO_LOOKUP* pLookup,
3075 void* compileTimeHandle);
3077 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3079 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3080 CorInfoHelpFunc helper,
3082 GenTreeArgList* arg = nullptr,
3083 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3085 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3087 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3090 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3091 CORINFO_CLASS_HANDLE typeClass,
3095 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3096 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3097 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3098 static bool IsMathIntrinsic(GenTreePtr tree);
3101 //----------------- Importing the method ----------------------------------
3103 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3106 unsigned impCurOpcOffs;
3107 const char* impCurOpcName;
3108 bool impNestedStackSpill;
3110 // For displaying instrs with generated native code (-n:B)
3111 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3112 void impNoteLastILoffs();
3115 /* IL offset of the stmt currently being imported. It gets set to
3116 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3117 updated at IL offsets for which we have to report mapping info.
3118 It also includes flag bits, so use jitGetILoffs()
3119 to get the actual IL offset value.
3122 IL_OFFSETX impCurStmtOffs;
3123 void impCurStmtOffsSet(IL_OFFSET offs);
3125 void impNoteBranchOffs();
3127 unsigned impInitBlockLineInfo();
3129 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3130 bool impIsThis(GenTreePtr obj);
3131 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3132 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3133 bool impIsAnySTLOC(OPCODE opcode)
3135 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3136 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3139 GenTreeArgList* impPopList(unsigned count,
3141 CORINFO_SIG_INFO* sig,
3142 GenTreeArgList* prefixTree = nullptr);
3144 GenTreeArgList* impPopRevList(unsigned count,
3146 CORINFO_SIG_INFO* sig,
3147 unsigned skipReverseCount = 0);
3150 * Get current IL offset with stack-empty info incoporated
3152 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3154 //---------------- Spilling the importer stack ----------------------------
3156 // The maximum number of bytes of IL processed without clean stack state.
3157 // It allows to limit the maximum tree size and depth.
3158 static const unsigned MAX_TREE_SIZE = 200;
3159 bool impCanSpillNow(OPCODE prevOpcode);
3165 SavedStack pdSavedStack;
3166 ThisInitState pdThisPtrInit;
3169 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3170 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3172 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3173 ExpandArray<BYTE> impPendingBlockMembers;
3175 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3176 // Operates on the map in the top-level ancestor.
3177 BYTE impGetPendingBlockMember(BasicBlock* blk)
3179 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3182 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3183 // Operates on the map in the top-level ancestor.
3184 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3186 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3189 bool impCanReimport;
3191 bool impSpillStackEntry(unsigned level,
3195 bool bAssertOnRecursion,
3200 void impSpillStackEnsure(bool spillLeaves = false);
3201 void impEvalSideEffects();
3202 void impSpillSpecialSideEff();
3203 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3204 void impSpillValueClasses();
3205 void impSpillEvalStack();
3206 static fgWalkPreFn impFindValueClasses;
3207 void impSpillLclRefs(ssize_t lclNum);
3209 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3211 void impImportBlockCode(BasicBlock* block);
3213 void impReimportMarkBlock(BasicBlock* block);
3214 void impReimportMarkSuccessors(BasicBlock* block);
3216 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3218 void impImportBlockPending(BasicBlock* block);
3220 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3221 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3222 // for the block, but instead, just re-uses the block's existing EntryState.
3223 void impReimportBlockPending(BasicBlock* block);
3225 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3227 void impImportBlock(BasicBlock* block);
3229 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3230 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3231 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3232 // the variables that will be used -- and for all the predecessors of those successors, and the
3233 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3234 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3235 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3236 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3237 // of local variable numbers, so we represent them with the base local variable number), returns that.
3238 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3239 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3240 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3241 // on which kind of member of the clique the block is).
3242 unsigned impGetSpillTmpBase(BasicBlock* block);
3244 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3245 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3246 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3247 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3248 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3249 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3250 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3251 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3252 // successors receive a native int. Similarly float and double are unified to double.
3253 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3254 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3255 // predecessors, so they insert an upcast if needed).
3256 void impReimportSpillClique(BasicBlock* block);
3258 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3259 // block, and represent the predecessor and successor members of the clique currently being computed.
3260 // *** Access to these will need to be locked in a parallel compiler.
3261 ExpandArray<BYTE> impSpillCliquePredMembers;
3262 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3270 // Abstract class for receiving a callback while walking a spill clique
3271 class SpillCliqueWalker
3274 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3277 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3278 class SetSpillTempsBase : public SpillCliqueWalker
3283 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3286 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3289 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3290 class ReimportSpillClique : public SpillCliqueWalker
3295 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3298 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3301 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3302 // predecessor or successor within the spill clique
3303 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3305 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3306 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3307 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3308 void impRetypeEntryStateTemps(BasicBlock* blk);
3310 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3311 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3313 void impPushVar(GenTree* op, typeInfo tiRetVal);
3314 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3315 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3317 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3319 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3320 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3321 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3324 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3327 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3328 struct BlockListNode
3331 BlockListNode* m_next;
3332 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3335 void* operator new(size_t sz, Compiler* comp);
3337 BlockListNode* impBlockListNodeFreeList;
3339 BlockListNode* AllocBlockListNode();
3340 void FreeBlockListNode(BlockListNode* node);
3342 bool impIsValueType(typeInfo* pTypeInfo);
3343 var_types mangleVarArgsType(var_types type);
3346 regNumber getCallArgIntRegister(regNumber floatReg);
3347 regNumber getCallArgFloatRegister(regNumber intReg);
3348 #endif // FEATURE_VARARG
3351 static unsigned jitTotalMethodCompiled;
3355 static LONG jitNestingLevel;
3358 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3360 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3362 // STATIC inlining decision based on the IL code.
3363 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3364 CORINFO_METHOD_INFO* methInfo,
3366 InlineResult* inlineResult);
3368 void impCheckCanInline(GenTreePtr call,
3369 CORINFO_METHOD_HANDLE fncHandle,
3371 CORINFO_CONTEXT_HANDLE exactContextHnd,
3372 InlineCandidateInfo** ppInlineCandidateInfo,
3373 InlineResult* inlineResult);
3375 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3376 GenTreePtr curArgVal,
3378 InlineResult* inlineResult);
3380 void impInlineInitVars(InlineInfo* pInlineInfo);
3382 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3384 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3386 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3388 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3389 GenTreePtr variableBeingDereferenced,
3390 InlArgInfo* inlArgInfo);
3392 void impMarkInlineCandidate(GenTreePtr call,
3393 CORINFO_CONTEXT_HANDLE exactContextHnd,
3394 bool exactContextNeedsRuntimeLookup,
3395 CORINFO_CALL_INFO* callInfo);
3397 bool impTailCallRetTypeCompatible(var_types callerRetType,
3398 CORINFO_CLASS_HANDLE callerRetTypeClass,
3399 var_types calleeRetType,
3400 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3402 bool impIsTailCallILPattern(bool tailPrefixed,
3404 const BYTE* codeAddrOfNextOpcode,
3405 const BYTE* codeEnd,
3407 bool* IsCallPopRet = nullptr);
3409 bool impIsImplicitTailCallCandidate(
3410 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3412 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3416 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3420 XX Info about the basic-blocks, their contents and the flow analysis XX
3422 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3423 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3427 BasicBlock* fgFirstBB; // Beginning of the basic block list
3428 BasicBlock* fgLastBB; // End of the basic block list
3429 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3430 #if FEATURE_EH_FUNCLETS
3431 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3433 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3435 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3436 unsigned fgEdgeCount; // # of control flow edges between the BBs
3437 unsigned fgBBcount; // # of BBs in the method
3439 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3441 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3442 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3443 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3444 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3446 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3447 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3448 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3449 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3450 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3451 // index). The arrays are of size fgBBNumMax + 1.
3452 unsigned* fgDomTreePreOrder;
3453 unsigned* fgDomTreePostOrder;
3455 bool fgBBVarSetsInited;
3457 // Allocate array like T* a = new T[fgBBNumMax + 1];
3458 // Using helper so we don't keep forgetting +1.
3459 template <typename T>
3460 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3462 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3465 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3466 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3467 // cannot be meaningfully combined. Note that new blocks can be created with higher
3468 // block numbers without changing the basic block epoch. These blocks *cannot*
3469 // participate in a block set until the blocks are all renumbered, causing the epoch
3470 // to change. This is useful if continuing to use previous block sets is valuable.
3471 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3472 unsigned fgCurBBEpoch;
3474 unsigned GetCurBasicBlockEpoch()
3476 return fgCurBBEpoch;
3479 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3480 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3481 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3482 unsigned fgCurBBEpochSize;
3484 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3485 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3486 unsigned fgBBSetCountInSizeTUnits;
3488 void NewBasicBlockEpoch()
3490 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3492 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3494 fgCurBBEpochSize = fgBBNumMax + 1;
3495 fgBBSetCountInSizeTUnits =
3496 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3499 // All BlockSet objects are now invalid!
3500 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3501 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3505 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3506 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3507 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3508 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3510 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3511 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3512 // array of size_t bitsets), then print that out.
3513 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3520 void EnsureBasicBlockEpoch()
3522 if (fgCurBBEpochSize != fgBBNumMax + 1)
3524 NewBasicBlockEpoch();
3528 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3529 void fgEnsureFirstBBisScratch();
3530 bool fgFirstBBisScratch();
3531 bool fgBBisScratch(BasicBlock* block);
3533 void fgExtendEHRegionBefore(BasicBlock* block);
3534 void fgExtendEHRegionAfter(BasicBlock* block);
3536 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3538 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3540 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3543 BasicBlock* nearBlk,
3544 bool putInFilter = false,
3545 bool runRarely = false,
3546 bool insertAtEnd = false);
3548 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3550 bool runRarely = false,
3551 bool insertAtEnd = false);
3553 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3555 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3556 BasicBlock* afterBlk,
3557 unsigned xcptnIndex,
3558 bool putInTryRegion);
3560 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3561 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3562 void fgUnlinkBlock(BasicBlock* block);
3564 unsigned fgMeasureIR();
3566 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3567 bool fgMultipleNots;
3570 bool fgModified; // True if the flow graph has been modified recently
3571 bool fgComputePredsDone; // Have we computed the bbPreds list
3572 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3573 bool fgDomsComputed; // Have we computed the dominator sets?
3574 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3576 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3577 bool fgHasPostfix; // any postfix ++/-- found?
3578 unsigned fgIncrCount; // number of increment nodes found
3580 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3584 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3585 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3588 bool fgRemoveRestOfBlock; // true if we know that we will throw
3589 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3591 // There are two modes for ordering of the trees.
3592 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3593 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3594 // by traversing the tree according to the order of the operands.
3595 // - In FGOrderLinear, the dominant ordering is the linear order.
3602 FlowGraphOrder fgOrder;
3604 // The following are boolean flags that keep track of the state of internal data structures
3606 bool fgStmtListThreaded;
3607 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3608 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3609 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3610 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3611 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3612 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3613 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3614 // This is derived from the profile data
3615 // or is BB_UNITY_WEIGHT when we don't have profile data
3617 #if FEATURE_EH_FUNCLETS
3618 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3619 #endif // FEATURE_EH_FUNCLETS
3621 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3622 // since fgMorphTree can be called from several places
3624 bool impBoxTempInUse; // the temp below is valid and available
3625 unsigned impBoxTemp; // a temporary that is used for boxing
3628 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3629 // and we are trying to compile again in a "safer", minopts mode?
3633 unsigned impInlinedCodeSize;
3636 //-------------------------------------------------------------------------
3642 void fgTransformFatCalli();
3646 void fgRemoveEmptyTry();
3648 void fgRemoveEmptyFinally();
3650 void fgMergeFinallyChains();
3652 void fgCloneFinally();
3654 void fgCleanupContinuation(BasicBlock* continuation);
3656 void fgUpdateFinallyTargetFlags();
3658 void fgClearAllFinallyTargetBits();
3660 void fgAddFinallyTargetFlags();
3662 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3663 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3664 // when this is necessary.
3665 bool fgNeedToAddFinallyTargetBits;
3666 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3668 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3669 BasicBlock* handler,
3670 BlockToBlockMap& continuationMap);
3672 GenTreePtr fgGetCritSectOfStaticMethod();
3674 #if FEATURE_EH_FUNCLETS
3676 void fgAddSyncMethodEnterExit();
3678 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3680 void fgConvertSyncReturnToLeave(BasicBlock* block);
3682 #endif // FEATURE_EH_FUNCLETS
3684 void fgAddReversePInvokeEnterExit();
3686 bool fgMoreThanOneReturnBlock();
3688 // The number of separate return points in the method.
3689 unsigned fgReturnCount;
3691 void fgAddInternal();
3693 bool fgFoldConditional(BasicBlock* block);
3695 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3696 void fgMorphBlocks();
3698 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3700 void fgCheckArgCnt();
3701 void fgSetOptions();
3704 static fgWalkPreFn fgAssertNoQmark;
3705 void fgPreExpandQmarkChecks(GenTreePtr expr);
3706 void fgPostExpandQmarkChecks();
3707 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3710 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3712 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3713 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3714 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3715 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3716 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3718 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3719 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3720 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3721 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3723 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3724 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3725 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3726 void fgExpandQmarkNodes();
3730 // Do "simple lowering." This functionality is (conceptually) part of "general"
3731 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3732 void fgSimpleLowering();
3734 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3736 GenTreePtr fgInitThisClass();
3738 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3740 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3742 inline bool backendRequiresLocalVarLifetimes()
3744 #if defined(LEGACY_BACKEND)
3747 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3751 void fgLocalVarLiveness();
3753 void fgLocalVarLivenessInit();
3755 #ifdef LEGACY_BACKEND
3756 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3758 void fgPerNodeLocalVarLiveness(GenTree* node);
3760 void fgPerBlockLocalVarLiveness();
3762 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3764 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3766 // This is used in the liveness computation, as a temporary. When we use the
3767 // arbitrary-length VarSet representation, it is better not to allocate a new one
3769 VARSET_TP fgMarkIntfUnionVS;
3771 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3773 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3775 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3777 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3779 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3781 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3783 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3785 void fgComputeLife(VARSET_TP& life,
3786 GenTreePtr startNode,
3788 VARSET_VALARG_TP volatileVars,
3789 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3791 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3793 bool fgRemoveDeadStore(GenTree** pTree,
3795 VARSET_VALARG_TP life,
3797 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3799 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3801 // For updating liveset during traversal AFTER fgComputeLife has completed
3802 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3803 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3805 // Returns the set of live variables after endTree,
3806 // assuming that liveSet is the set of live variables BEFORE tree.
3807 // Requires that fgComputeLife has completed, and that tree is in the same
3808 // statement as endTree, and that it comes before endTree in execution order
3810 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3812 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3813 while (tree != nullptr && tree != endTree->gtNext)
3815 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3816 tree = tree->gtNext;
3818 assert(tree == endTree->gtNext);
3822 void fgInterBlockLocalVarLiveness();
3824 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3825 // "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
3826 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3827 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3828 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3829 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3830 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3832 if (m_opAsgnVarDefSsaNums == nullptr)
3834 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3836 return m_opAsgnVarDefSsaNums;
3839 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3840 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3841 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3843 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3845 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3846 // Except: assumes that lcl is a def, and if it is
3847 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3848 // rather than the "use" SSA number recorded in the tree "lcl".
3849 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3851 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3852 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3853 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3854 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3855 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3857 // (byref addrS1 = &s1,
3858 // *(addrS1 * offsetof(f0)) = s2f0,
3860 // *(addrS1 * offsetof(fn)) = s2fn)
3862 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3863 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3864 // give it SSA names and value numbers?
3866 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3867 // end with an instance of the structure below, whose fields are described in the declaration.
3868 struct IndirectAssignmentAnnotation
3870 unsigned m_lclNum; // The local num that is being indirectly assigned.
3871 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3872 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3873 // be the singleton field sequence "g". The individual assignments would
3874 // further append the fields of "s.g" to that.
3875 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3876 // structure has a single field).
3877 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3878 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3881 IndirectAssignmentAnnotation(unsigned lclNum,
3882 FieldSeqNode* fldSeq,
3884 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3885 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3886 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3890 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3891 NodeToIndirAssignMap;
3892 NodeToIndirAssignMap* m_indirAssignMap;
3893 NodeToIndirAssignMap* GetIndirAssignMap()
3895 if (m_indirAssignMap == nullptr)
3897 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3898 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3899 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3901 return m_indirAssignMap;
3904 // Performs SSA conversion.
3907 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3908 void fgResetForSsa();
3910 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3912 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3913 inline bool fgExcludeFromSsa(unsigned lclNum);
3915 // The value numbers for this compilation.
3916 ValueNumStore* vnStore;
3919 ValueNumStore* GetValueNumStore()
3924 // Do value numbering (assign a value number to each
3926 void fgValueNumber();
3928 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3929 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3930 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3931 // match the element type of the array or fldSeq. When this type doesn't match
3932 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3934 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3937 FieldSeqNode* fldSeq,
3941 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3942 // has been parsed to yield the other input arguments. If evaluation of the address
3943 // can raise exceptions, those should be captured in the exception set "excVN."
3944 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3945 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3946 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3947 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3948 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3950 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3951 CORINFO_CLASS_HANDLE elemTypeEq,
3955 FieldSeqNode* fldSeq);
3957 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3958 // by evaluating the array index expression "tree". Returns the value number resulting from
3959 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3960 // "GT_IND" that does the dereference, and it is given the returned value number.
3961 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3963 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3964 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3966 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3968 // Utility functions for fgValueNumber.
3970 // Perform value-numbering for the trees in "blk".
3971 void fgValueNumberBlock(BasicBlock* blk);
3973 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3974 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3975 // assumed for the memoryKind at the start "entryBlk".
3976 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3978 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3979 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3980 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3982 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3984 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3986 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3987 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3988 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3990 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3991 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3993 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3994 // value in that SSA #.
3995 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3997 // The input 'tree' is a leaf node that is a constant
3998 // Assign the proper value number to the tree
3999 void fgValueNumberTreeConst(GenTreePtr tree);
4001 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
4002 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4004 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4006 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
4008 // Does value-numbering for a block assignment.
4009 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4011 // Does value-numbering for a cast tree.
4012 void fgValueNumberCastTree(GenTreePtr tree);
4014 // Does value-numbering for an intrinsic tree.
4015 void fgValueNumberIntrinsic(GenTreePtr tree);
4017 // Does value-numbering for a call. We interpret some helper calls.
4018 void fgValueNumberCall(GenTreeCall* call);
4020 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4021 void fgUpdateArgListVNs(GenTreeArgList* args);
4023 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4024 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4026 // Requires "helpCall" to be a helper call. Assigns it a value number;
4027 // we understand the semantics of some of the calls. Returns "true" if
4028 // the call may modify the heap (we assume arbitrary memory side effects if so).
4029 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4031 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4032 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4034 // These are the current value number for the memory implicit variables while
4035 // doing value numbering. These are the value numbers under the "liberal" interpretation
4036 // of memory values; the "conservative" interpretation needs no VN, since every access of
4037 // memory yields an unknown value.
4038 ValueNum fgCurMemoryVN[MemoryKindCount];
4040 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4041 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4042 // is 1, and the rest is an encoding of "elemTyp".
4043 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4045 if (elemStructType != nullptr)
4047 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4048 varTypeIsIntegral(elemTyp));
4049 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4050 return elemStructType;
4054 elemTyp = varTypeUnsignedToSigned(elemTyp);
4055 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4058 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4059 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4060 // the struct type of the element).
4061 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4063 size_t clsHndVal = size_t(clsHnd);
4064 if (clsHndVal & 0x1)
4066 return var_types(clsHndVal >> 1);
4074 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4075 var_types getJitGCType(BYTE gcType);
4077 enum structPassingKind
4079 SPK_Unknown, // Invalid value, never returned
4080 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4081 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4082 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4083 // parameters registers are used, then the stack will be used)
4084 // for X86 passed on the stack, for ARM32 passed in registers
4085 // or the stack or split between registers and the stack.
4086 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4088 }; // The struct is passed/returned by reference to a copy/buffer.
4090 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4091 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4092 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4093 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4095 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4097 // Get the type that is used to pass values of the given struct type.
4098 // If you have already retrieved the struct size then pass it as the optional third argument
4100 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4101 structPassingKind* wbPassStruct,
4102 unsigned structSize = 0);
4104 // Get the type that is used to return values of the given struct type.
4105 // If you have already retrieved the struct size then pass it as the optional third argument
4107 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4108 structPassingKind* wbPassStruct = nullptr,
4109 unsigned structSize = 0);
4112 // Print a representation of "vnp" or "vn" on standard output.
4113 // If "level" is non-zero, we also print out a partial expansion of the value.
4114 void vnpPrint(ValueNumPair vnp, unsigned level);
4115 void vnPrint(ValueNum vn, unsigned level);
4118 // Dominator computation member functions
4119 // Not exposed outside Compiler
4121 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4123 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4125 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4126 // flow graph. We first assume the fields bbIDom on each
4127 // basic block are invalid. This computation is needed later
4128 // by fgBuildDomTree to build the dominance tree structure.
4129 // Based on: A Simple, Fast Dominance Algorithm
4130 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4132 void fgCompDominatedByExceptionalEntryBlocks();
4134 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4135 // Note: this is relatively slow compared to calling fgDominate(),
4136 // especially if dealing with a single block versus block check.
4138 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4140 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4142 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4144 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4146 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4148 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4149 // processed in topological sort, this function takes care of that.
4151 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4153 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4154 // Returns this as a set.
4156 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4157 // root nodes. Returns this as a set.
4160 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4163 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4164 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4167 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4168 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4169 // && postOrder(A) >= postOrder(B) making the computation O(1).
4170 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4172 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4174 void fgUpdateChangedFlowGraph();
4177 // Compute the predecessors of the blocks in the control flow graph.
4178 void fgComputePreds();
4180 // Remove all predecessor information.
4181 void fgRemovePreds();
4183 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4184 // before the full predecessors lists are computed.
4185 void fgComputeCheapPreds();
4188 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4190 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4200 // Initialize the per-block variable sets (used for liveness analysis).
4201 void fgInitBlockVarSets();
4203 // true if we've gone through and created GC Poll calls.
4204 bool fgGCPollsCreated;
4205 void fgMarkGCPollBlocks();
4206 void fgCreateGCPolls();
4207 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4209 // Requires that "block" is a block that returns from
4210 // a finally. Returns the number of successors (jump targets of
4211 // of blocks in the covered "try" that did a "LEAVE".)
4212 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4214 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4215 // a finally. Returns its "i"th successor (jump targets of
4216 // of blocks in the covered "try" that did a "LEAVE".)
4217 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4218 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4221 // Factor out common portions of the impls of the methods above.
4222 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4225 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4226 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4227 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4228 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4229 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4230 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4231 // we leave the entry associated with the block, but it will no longer be accessed.)
4232 struct SwitchUniqueSuccSet
4234 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4235 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4238 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4239 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4240 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4241 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4244 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4245 BlockToSwitchDescMap;
4248 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4249 // iteration over only the distinct successors.
4250 BlockToSwitchDescMap* m_switchDescMap;
4253 BlockToSwitchDescMap* GetSwitchDescMap()
4255 if (m_switchDescMap == nullptr)
4257 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4259 return m_switchDescMap;
4262 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4263 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4264 // we don't accidentally look up and return the wrong switch data.
4265 void InvalidateUniqueSwitchSuccMap()
4267 m_switchDescMap = nullptr;
4270 // Requires "switchBlock" to be a block that ends in a switch. Returns
4271 // the corresponding SwitchUniqueSuccSet.
4272 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4274 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4275 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4276 // remove it from "this", and ensure that "to" is a member.
4277 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4279 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4280 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4282 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4284 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4286 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4288 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4290 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4292 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4294 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4296 void fgRemoveBlockAsPred(BasicBlock* block);
4298 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4300 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4302 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4304 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4306 flowList* fgAddRefPred(BasicBlock* block,
4307 BasicBlock* blockPred,
4308 flowList* oldEdge = nullptr,
4309 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4312 void fgFindBasicBlocks();
4314 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4316 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4318 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4319 bool putInTryRegion,
4320 BasicBlock* startBlk,
4322 BasicBlock* nearBlk,
4323 BasicBlock* jumpBlk,
4326 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4328 void fgRemoveEmptyBlocks();
4330 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4332 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4334 void fgCreateLoopPreHeader(unsigned lnum);
4336 void fgUnreachableBlock(BasicBlock* block);
4338 void fgRemoveConditionalJump(BasicBlock* block);
4340 BasicBlock* fgLastBBInMainFunction();
4342 BasicBlock* fgEndBBAfterMainFunction();
4344 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4346 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4348 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4350 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4352 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4354 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4356 bool fgRenumberBlocks();
4358 bool fgExpandRarelyRunBlocks();
4360 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4362 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4364 enum FG_RELOCATE_TYPE
4366 FG_RELOCATE_TRY, // relocate the 'try' region
4367 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4369 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4371 #if FEATURE_EH_FUNCLETS
4372 #if defined(_TARGET_ARM_)
4373 void fgClearFinallyTargetBit(BasicBlock* block);
4374 #endif // defined(_TARGET_ARM_)
4375 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4376 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4377 void fgInsertFuncletPrologBlock(BasicBlock* block);
4378 void fgCreateFuncletPrologBlocks();
4379 void fgCreateFunclets();
4380 #else // !FEATURE_EH_FUNCLETS
4381 bool fgRelocateEHRegions();
4382 #endif // !FEATURE_EH_FUNCLETS
4384 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4386 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4388 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4390 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4392 bool fgOptimizeEmptyBlock(BasicBlock* block);
4394 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4396 bool fgOptimizeBranch(BasicBlock* bJump);
4398 bool fgOptimizeSwitchBranches(BasicBlock* block);
4400 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4402 bool fgOptimizeSwitchJumps();
4404 void fgPrintEdgeWeights();
4406 void fgComputeEdgeWeights();
4408 void fgReorderBlocks();
4410 void fgDetermineFirstColdBlock();
4412 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4414 bool fgUpdateFlowGraph(bool doTailDup = false);
4416 void fgFindOperOrder();
4418 // method that returns if you should split here
4419 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4421 void fgSetBlockOrder();
4423 void fgRemoveReturnBlock(BasicBlock* block);
4425 /* Helper code that has been factored out */
4426 inline void fgConvertBBToThrowBB(BasicBlock* block);
4428 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4429 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4430 GenTreePtr fgMakeTmpArgNode(
4431 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4433 // The following check for loops that don't execute calls
4434 bool fgLoopCallMarked;
4436 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4437 void fgLoopCallMark();
4439 void fgMarkLoopHead(BasicBlock* block);
4441 unsigned fgGetCodeEstimate(BasicBlock* block);
4444 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4445 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4446 bool fgDumpFlowGraph(Phases phase);
4448 #endif // DUMP_FLOWGRAPHS
4453 void fgDispBBLiveness(BasicBlock* block);
4454 void fgDispBBLiveness();
4455 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4456 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4457 void fgDispBasicBlocks(bool dumpTrees = false);
4458 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4459 void fgDumpBlock(BasicBlock* block);
4460 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4462 static fgWalkPreFn fgStress64RsltMulCB;
4463 void fgStress64RsltMul();
4464 void fgDebugCheckUpdate();
4465 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4466 void fgDebugCheckBlockLinks();
4467 void fgDebugCheckLinks(bool morphTrees = false);
4468 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4469 void fgDebugCheckFlags(GenTreePtr tree);
4470 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4471 void fgDebugCheckTryFinallyExits();
4474 #ifdef LEGACY_BACKEND
4475 static void fgOrderBlockOps(GenTreePtr tree,
4479 GenTreePtr* opsPtr, // OUT
4480 regMaskTP* regsPtr); // OUT
4481 #endif // LEGACY_BACKEND
4483 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4484 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4485 void fgTraverseRPO();
4487 //--------------------- Walking the trees in the IR -----------------------
4492 fgWalkPreFn* wtprVisitorFn;
4493 fgWalkPostFn* wtpoVisitorFn;
4494 void* pCallbackData; // user-provided data
4495 bool wtprLclsOnly; // whether to only visit lclvar nodes
4496 GenTreePtr parent; // parent of current node, provided to callback
4497 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4499 bool printModified; // callback can use this
4503 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4504 fgWalkPreFn* visitor,
4505 void* pCallBackData = nullptr,
4506 bool lclVarsOnly = false,
4507 bool computeStack = false);
4509 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4510 fgWalkPreFn* preVisitor,
4511 fgWalkPostFn* postVisitor,
4512 void* pCallBackData = nullptr);
4514 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4518 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4519 fgWalkPostFn* visitor,
4520 void* pCallBackData = nullptr,
4521 bool computeStack = false);
4523 // An fgWalkPreFn that looks for expressions that have inline throws in
4524 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4525 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4526 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4527 // properly propagated to parent trees). It returns WALK_CONTINUE
4529 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4530 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4531 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4533 /**************************************************************************
4535 *************************************************************************/
4538 friend class SsaBuilder;
4539 friend struct ValueNumberState;
4541 //--------------------- Detect the basic blocks ---------------------------
4543 BasicBlock** fgBBs; // Table of pointers to the BBs
4545 void fgInitBBLookup();
4546 BasicBlock* fgLookupBB(unsigned addr);
4548 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4550 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4552 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4554 void fgLinkBasicBlocks();
4556 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4558 void fgCheckBasicBlockControlFlow();
4560 void fgControlFlowPermitted(BasicBlock* blkSrc,
4561 BasicBlock* blkDest,
4562 BOOL IsLeave = false /* is the src a leave block */);
4564 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4566 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4568 void fgAdjustForAddressExposedOrWrittenThis();
4570 bool fgProfileData_ILSizeMismatch;
4571 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4572 ULONG fgProfileBufferCount;
4573 ULONG fgNumProfileRuns;
4575 unsigned fgStressBBProf()
4578 unsigned result = JitConfig.JitStressBBProf();
4581 if (compStressCompile(STRESS_BB_PROFILE, 15))
4592 bool fgHaveProfileData();
4593 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4594 void fgInstrumentMethod();
4597 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4598 // or if we have some fake profile data for the stress mode
4599 bool fgIsUsingProfileWeights()
4601 return (fgHaveProfileData() || fgStressBBProf());
4604 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4605 // or BB_UNITY_WEIGHT when we aren't using profile data.
4606 unsigned fgProfileRunsCount()
4608 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4611 //-------- Insert a statement at the start or end of a basic block --------
4615 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4619 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4621 public: // Used by linear scan register allocation
4622 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4625 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4626 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4628 public: // Used by linear scan register allocation
4629 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4632 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4634 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4636 // Create a new temporary variable to hold the result of *ppTree,
4637 // and transform the graph accordingly.
4638 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4639 GenTree* fgMakeMultiUse(GenTree** ppTree);
4642 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4643 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4644 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4646 //-------- Determine the order in which the trees will be evaluated -------
4648 unsigned fgTreeSeqNum;
4649 GenTree* fgTreeSeqLst;
4650 GenTree* fgTreeSeqBeg;
4652 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4653 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4654 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4655 void fgSetStmtSeq(GenTree* tree);
4656 void fgSetBlockOrder(BasicBlock* block);
4658 //------------------------- Morphing --------------------------------------
4660 unsigned fgPtrArgCntCur;
4661 unsigned fgPtrArgCntMax;
4662 hashBv* fgOutgoingArgTemps;
4663 hashBv* fgCurrentlyInUseArgTemps;
4665 bool compCanEncodePtrArgCntMax();
4667 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4670 void fgMoveOpsLeft(GenTreePtr tree);
4673 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4675 bool fgIsThrow(GenTreePtr tree);
4677 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4678 bool fgIsBlockCold(BasicBlock* block);
4680 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4682 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4684 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4686 bool fgMorphRelopToQmark(GenTreePtr tree);
4688 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4689 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4690 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4691 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4692 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4693 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4694 // small; hence the other fields of MorphAddrContext.
4695 enum MorphAddrContextKind
4700 struct MorphAddrContext
4702 MorphAddrContextKind m_kind;
4703 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4704 // top-level indirection and here have been constants.
4705 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4706 // In that case, is the sum of those constant offsets.
4708 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4713 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4714 static MorphAddrContext s_CopyBlockMAC;
4717 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4718 var_types* baseTypeOut,
4720 unsigned* simdSizeOut,
4721 bool ignoreUsedInSIMDIntrinsic = false);
4722 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4723 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4724 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4725 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4727 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4728 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4729 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4731 #endif // FEATURE_SIMD
4732 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4733 GenTreePtr fgMorphCast(GenTreePtr tree);
4734 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4735 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4737 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4740 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4741 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4743 void fgFixupStructReturn(GenTreePtr call);
4744 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4745 bool fgAddrCouldBeNull(GenTreePtr addr);
4746 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4747 bool fgCanFastTailCall(GenTreeCall* call);
4748 void fgMorphTailCall(GenTreeCall* call);
4749 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4750 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4751 fgArgTabEntryPtr argTabEntry,
4753 IL_OFFSETX callILOffset,
4754 GenTreePtr tmpAssignmentInsertionPoint,
4755 GenTreePtr paramAssignmentInsertionPoint);
4756 static int fgEstimateCallStackSize(GenTreeCall* call);
4757 GenTreePtr fgMorphCall(GenTreeCall* call);
4758 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4759 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4761 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4762 static fgWalkPreFn fgFindNonInlineCandidate;
4764 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4765 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4766 CORINFO_RESOLVED_TOKEN* ldftnToken);
4767 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4768 void fgAssignSetVarDef(GenTreePtr tree);
4769 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4770 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4771 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4772 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4773 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4774 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4775 void fgMorphUnsafeBlk(GenTreeObj* obj);
4776 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4777 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4778 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4779 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4780 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4781 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4782 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4784 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4785 GenTreePtr fgMorphConst(GenTreePtr tree);
4788 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4791 #if LOCAL_ASSERTION_PROP
4792 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4793 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4795 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4797 GenTreeStmt* fgMorphStmt;
4799 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4800 // used when morphing big offset.
4802 //----------------------- Liveness analysis -------------------------------
4804 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4805 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4807 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4808 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4809 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4811 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4813 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4815 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4816 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4818 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4819 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4821 void fgExtendDbgScopes();
4822 void fgExtendDbgLifetimes();
4825 void fgDispDebugScopes();
4828 //-------------------------------------------------------------------------
4830 // The following keeps track of any code we've added for things like array
4831 // range checking or explicit calls to enable GC, and so on.
4836 AddCodeDsc* acdNext;
4837 BasicBlock* acdDstBlk; // block to which we jump
4839 SpecialCodeKind acdKind; // what kind of a special block is this?
4840 unsigned short acdStkLvl;
4844 static unsigned acdHelper(SpecialCodeKind codeKind);
4846 AddCodeDsc* fgAddCodeList;
4848 bool fgRngChkThrowAdded;
4849 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4851 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4853 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4856 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4859 bool fgIsCodeAdded();
4861 bool fgIsThrowHlpBlk(BasicBlock* block);
4862 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4864 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4866 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4867 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4868 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4869 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4870 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4872 #if FEATURE_MULTIREG_RET
4873 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4874 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4875 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4876 #endif // FEATURE_MULTIREG_RET
4878 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4881 static fgWalkPreFn fgDebugCheckInlineCandidates;
4883 void CheckNoFatPointerCandidatesLeft();
4884 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4887 void fgPromoteStructs();
4888 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4889 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4891 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4892 void fgMarkImplicitByRefArgs();
4894 // Change implicit byrefs' types from struct to pointer, and for any that were
4895 // promoted, create new promoted struct temps.
4896 void fgRetypeImplicitByRefArgs();
4898 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4899 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4900 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4902 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4903 void fgMarkDemotedImplicitByRefArgs();
4905 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4906 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4907 void fgMarkAddressExposedLocals();
4908 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4910 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4912 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4914 // The given local variable, required to be a struct variable, is being assigned via
4915 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4916 // the variable is not enregistered, and is therefore not promoted independently.
4917 void fgLclFldAssign(unsigned lclNum);
4919 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4920 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4921 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4922 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4925 bool fgPrintInlinedMethods;
4928 bool fgIsBigOffset(size_t offset);
4930 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4931 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4932 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4933 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4934 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4937 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4938 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4942 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4943 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4950 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4953 void optRemoveRangeCheck(
4954 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4955 bool optIsRangeCheckRemovable(GenTreePtr tree);
4958 static fgWalkPreFn optValidRangeCheckIndex;
4959 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4962 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4964 /**************************************************************************
4966 *************************************************************************/
4969 // Do hoisting for all loops.
4970 void optHoistLoopCode();
4972 // To represent sets of VN's that have already been hoisted in outer loops.
4973 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4974 typedef VNToBoolMap VNSet;
4976 struct LoopHoistContext
4979 // The set of variables hoisted in the current loop (or nullptr if there are none).
4980 VNSet* m_pHoistedInCurLoop;
4983 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4984 VNSet m_hoistedInParentLoops;
4985 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4986 // Previous decisions on loop-invariance of value numbers in the current loop.
4987 VNToBoolMap m_curLoopVnInvariantCache;
4989 VNSet* GetHoistedInCurLoop(Compiler* comp)
4991 if (m_pHoistedInCurLoop == nullptr)
4993 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4995 return m_pHoistedInCurLoop;
4998 VNSet* ExtractHoistedInCurLoop()
5000 VNSet* res = m_pHoistedInCurLoop;
5001 m_pHoistedInCurLoop = nullptr;
5005 LoopHoistContext(Compiler* comp)
5006 : m_pHoistedInCurLoop(nullptr)
5007 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5008 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5013 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5014 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5015 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5016 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5018 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5019 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5020 // "m_hoistedInParentLoops".
5022 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5024 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5025 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5026 // expressions to "hoistInLoop".
5027 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5029 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5030 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5032 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5033 // that are invariant in loop "lnum" (an index into the optLoopTable)
5034 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5035 // expressions to "hoistInLoop".
5036 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5037 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5038 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5039 bool optHoistLoopExprsForTree(GenTreePtr tree,
5041 LoopHoistContext* hoistCtxt,
5042 bool* firstBlockAndBeforeSideEffect,
5044 bool* pCctorDependent);
5046 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5047 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5049 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5050 // Constants and init values are always loop invariant.
5051 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5052 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5054 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5055 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5056 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5057 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5058 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5060 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5061 // in the loop table.
5062 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5064 // Records the set of "side effects" of all loops: fields (object instance and static)
5065 // written to, and SZ-array element type equivalence classes updated.
5066 void optComputeLoopSideEffects();
5069 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5070 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5071 // static) written to, and SZ-array element type equivalence classes updated.
5072 void optComputeLoopNestSideEffects(unsigned lnum);
5074 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5075 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5077 // Hoist the expression "expr" out of loop "lnum".
5078 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5081 void optOptimizeBools();
5084 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5086 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5089 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5091 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5092 // the loop into a "do-while" loop
5093 // Also finds all natural loops and records them in the loop table
5095 // Optionally clone loops in the loop table.
5096 void optCloneLoops();
5098 // Clone loop "loopInd" in the loop table.
5099 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5101 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5102 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5103 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5105 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5107 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5110 // This enumeration describes what is killed by a call.
5114 CALLINT_NONE, // no interference (most helpers)
5115 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5116 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5117 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5118 CALLINT_ALL, // kills everything (normal method call)
5122 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5123 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5124 // in bbNext order; we use comparisons on the bbNum to decide order.)
5125 // The blocks that define the body are
5126 // first <= top <= entry <= bottom .
5127 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5128 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5129 // Compiler::optFindNaturalLoops().
5132 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5133 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5134 // loop, but not the outer loop.)
5135 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5137 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5138 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5139 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5141 callInterf lpAsgCall; // "callInterf" for calls in the loop
5142 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5143 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5145 unsigned short lpFlags; // Mask of the LPFLG_* constants
5147 unsigned char lpExitCnt; // number of exits from the loop
5149 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5150 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5151 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5152 // (Actually, an "immediately" nested loop --
5153 // no other child of this loop is a parent of lpChild.)
5154 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5155 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5156 // by following "lpChild" then "lpSibling" links.
5158 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5159 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5161 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5162 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5163 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5165 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5166 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5168 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5169 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5170 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5171 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5173 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5174 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5175 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5177 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5178 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5179 // type are assigned to.
5181 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5182 // memory side effects. If this is set, the fields below
5183 // may not be accurate (since they become irrelevant.)
5184 bool lpContainsCall; // True if executing the loop body *may* execute a call
5186 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5187 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5189 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5191 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5192 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5194 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5196 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5197 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5199 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5200 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5202 JitSimplerHashBehavior>
5204 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5205 // instance fields modified
5208 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5209 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5211 JitSimplerHashBehavior>
5213 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5214 // arrays of that type are modified
5217 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5218 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5220 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5221 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5222 // (shifted left, with a low-order bit set to distinguish.)
5223 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5224 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5226 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5228 GenTreePtr lpIterTree; // The "i <op>= const" tree
5229 unsigned lpIterVar(); // iterator variable #
5230 int lpIterConst(); // the constant with which the iterator is incremented
5231 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5232 void VERIFY_lpIterTree();
5234 var_types lpIterOperType(); // For overflow instructions
5237 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5238 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5242 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5244 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5245 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5246 void VERIFY_lpTestTree();
5248 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5249 GenTreePtr lpIterator(); // the iterator node in the loop test
5250 GenTreePtr lpLimit(); // the limit node in the loop test
5252 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5253 // LPFLG_CONST_LIMIT
5254 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5256 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5257 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5258 // LPFLG_ARRLEN_LIMIT
5260 // Returns "true" iff "*this" contains the blk.
5261 bool lpContains(BasicBlock* blk)
5263 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5265 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5266 // to be equal, but requiring bottoms to be different.)
5267 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5269 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5272 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5273 // bottoms to be different.)
5274 bool lpContains(const LoopDsc& lp2)
5276 return lpContains(lp2.lpFirst, lp2.lpBottom);
5279 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5280 // (allowing firsts to be equal, but requiring bottoms to be different.)
5281 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5283 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5286 // Returns "true" iff "*this" is (properly) contained by "lp2"
5287 // (allowing firsts to be equal, but requiring bottoms to be different.)
5288 bool lpContainedBy(const LoopDsc& lp2)
5290 return lpContains(lp2.lpFirst, lp2.lpBottom);
5293 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5294 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5296 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5298 // Returns "true" iff "*this" is disjoint from "lp2".
5299 bool lpDisjoint(const LoopDsc& lp2)
5301 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5303 // Returns "true" iff the loop is well-formed (see code for defn).
5306 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5307 lpEntry->bbNum <= lpBottom->bbNum &&
5308 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5313 bool fgMightHaveLoop(); // returns true if there are any backedges
5314 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5317 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5318 unsigned char optLoopCount; // number of tracked loops
5321 unsigned optCallCount; // number of calls made in the method
5322 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5323 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5324 unsigned optLoopsCloned; // number of loops cloned in the current method.
5327 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5328 void optPrintLoopInfo(unsigned loopNum,
5330 BasicBlock* lpFirst,
5332 BasicBlock* lpEntry,
5333 BasicBlock* lpBottom,
5334 unsigned char lpExitCnt,
5336 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5337 void optPrintLoopInfo(unsigned lnum);
5338 void optPrintLoopRecording(unsigned lnum);
5340 void optCheckPreds();
5343 void optSetBlockWeights();
5345 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5347 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5349 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5351 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5352 unsigned optIsLoopIncrTree(GenTreePtr incr);
5353 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5354 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5355 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5356 bool optExtractInitTestIncr(BasicBlock* head,
5361 GenTreePtr* ppIncr);
5363 void optRecordLoop(BasicBlock* head,
5369 unsigned char exitCnt);
5371 void optFindNaturalLoops();
5373 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5374 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5375 bool optCanonicalizeLoopNest(unsigned char loopInd);
5377 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5378 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5379 bool optCanonicalizeLoop(unsigned char loopInd);
5381 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5382 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5383 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5384 bool optLoopContains(unsigned l1, unsigned l2);
5386 // Requires "loopInd" to be a valid index into the loop table.
5387 // Updates the loop table by changing loop "loopInd", whose head is required
5388 // to be "from", to be "to". Also performs this transformation for any
5389 // loop nested in "loopInd" that shares the same head as "loopInd".
5390 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5392 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5393 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5394 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5396 // Marks the containsCall information to "lnum" and any parent loops.
5397 void AddContainsCallAllContainingLoops(unsigned lnum);
5398 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5399 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5400 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5401 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5402 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5403 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5405 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5406 // of "from".) Copies the jump destination from "from" to "to".
5407 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5409 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5410 unsigned optLoopDepth(unsigned lnum)
5412 unsigned par = optLoopTable[lnum].lpParent;
5413 if (par == BasicBlock::NOT_IN_LOOP)
5419 return 1 + optLoopDepth(par);
5423 void fgOptWhileLoop(BasicBlock* block);
5425 bool optComputeLoopRep(int constInit,
5428 genTreeOps iterOper,
5430 genTreeOps testOper,
5433 unsigned* iterCount);
5434 #if FEATURE_STACK_FP_X87
5437 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5438 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5439 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5440 #endif // FEATURE_STACK_FP_X87
5443 static fgWalkPreFn optIsVarAssgCB;
5446 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5448 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5450 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5452 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5454 /**************************************************************************
5455 * Optimization conditions
5456 *************************************************************************/
5458 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5459 bool optPentium4(void);
5460 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5461 bool optAvoidIntMult(void);
5466 // The following is the upper limit on how many expressions we'll keep track
5467 // of for the CSE analysis.
5469 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5471 static const int MIN_CSE_COST = 2;
5473 // Keeps tracked cse indices
5474 BitVecTraits* cseTraits;
5477 /* Generic list of nodes - used by the CSE logic */
5485 typedef struct treeLst* treeLstPtr;
5489 treeStmtLst* tslNext;
5490 GenTreePtr tslTree; // tree node
5491 GenTreePtr tslStmt; // statement containing the tree
5492 BasicBlock* tslBlock; // block containing the statement
5495 typedef struct treeStmtLst* treeStmtLstPtr;
5497 // The following logic keeps track of expressions via a simple hash table.
5501 CSEdsc* csdNextInBucket; // used by the hash table
5503 unsigned csdHashValue; // the orginal hashkey
5505 unsigned csdIndex; // 1..optCSECandidateCount
5506 char csdLiveAcrossCall; // 0 or 1
5508 unsigned short csdDefCount; // definition count
5509 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5511 unsigned csdDefWtCnt; // weighted def count
5512 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5514 GenTreePtr csdTree; // treenode containing the 1st occurance
5515 GenTreePtr csdStmt; // stmt containing the 1st occurance
5516 BasicBlock* csdBlock; // block containing the 1st occurance
5518 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5519 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5521 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5522 // number, this will reflect it; otherwise, NoVN.
5525 static const size_t s_optCSEhashSize;
5526 CSEdsc** optCSEhash;
5529 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5531 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5532 // re-numbered with the bound to improve range check elimination
5534 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5535 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5539 CSEdsc* optCSEfindDsc(unsigned index);
5540 void optUnmarkCSE(GenTreePtr tree);
5542 // user defined callback data for the tree walk function optCSE_MaskHelper()
5543 struct optCSE_MaskData
5545 EXPSET_TP CSE_defMask;
5546 EXPSET_TP CSE_useMask;
5549 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5550 static fgWalkPreFn optCSE_MaskHelper;
5552 // This function walks all the node for an given tree
5553 // and return the mask of CSE definitions and uses for the tree
5555 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5557 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5558 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5559 bool optCSE_canSwap(GenTree* tree);
5561 static fgWalkPostFn optPropagateNonCSE;
5562 static fgWalkPreFn optHasNonCSEChild;
5564 static fgWalkPreFn optUnmarkCSEs;
5566 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5567 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5569 void optCleanupCSEs();
5572 void optEnsureClearCSEInfo();
5575 #endif // FEATURE_ANYCSE
5577 #if FEATURE_VALNUM_CSE
5578 /**************************************************************************
5579 * Value Number based CSEs
5580 *************************************************************************/
5583 void optOptimizeValnumCSEs();
5586 void optValnumCSE_Init();
5587 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5588 unsigned optValnumCSE_Locate();
5589 void optValnumCSE_InitDataFlow();
5590 void optValnumCSE_DataFlow();
5591 void optValnumCSE_Availablity();
5592 void optValnumCSE_Heuristic();
5593 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5595 #endif // FEATURE_VALNUM_CSE
5598 bool optDoCSE; // True when we have found a duplicate CSE tree
5599 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5600 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5601 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5602 unsigned optCSEstart; // The first local variable number that is a CSE
5603 unsigned optCSEcount; // The total count of CSE's introduced.
5604 unsigned optCSEweight; // The weight of the current block when we are
5605 // scanning for CSE expressions
5607 bool optIsCSEcandidate(GenTreePtr tree);
5609 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5611 bool lclNumIsTrueCSE(unsigned lclNum) const
5613 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5616 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5618 bool lclNumIsCSE(unsigned lclNum) const
5620 return lvaTable[lclNum].lvIsCSE;
5624 bool optConfigDisableCSE();
5625 bool optConfigDisableCSE2();
5627 void optOptimizeCSEs();
5629 #endif // FEATURE_ANYCSE
5637 unsigned ivaVar; // Variable we are interested in, or -1
5638 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5639 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5640 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5641 callInterf ivaMaskCall; // What kind of calls are there?
5644 static callInterf optCallInterf(GenTreeCall* call);
5647 // VN based copy propagation.
5648 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5649 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5650 LclNumToGenTreePtrStack;
5652 // Kill set to track variables with intervening definitions.
5653 VARSET_TP optCopyPropKillSet;
5655 // Copy propagation functions.
5656 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5657 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5658 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5659 bool optIsSsaLocal(GenTreePtr tree);
5660 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5661 void optVnCopyProp();
5663 /**************************************************************************
5664 * Early value propagation
5665 *************************************************************************/
5671 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5675 static unsigned GetHashCode(SSAName ssaNm)
5677 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5680 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5682 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5686 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5687 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5688 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5689 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5690 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5691 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5693 bool doesMethodHaveFatPointer()
5695 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5698 void setMethodHasFatPointer()
5700 optMethodFlags |= OMF_HAS_FATPOINTER;
5703 void clearMethodHasFatPointer()
5705 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5708 void addFatPointerCandidate(GenTreeCall* call)
5710 setMethodHasFatPointer();
5711 call->SetFatPointerCandidate();
5714 unsigned optMethodFlags;
5716 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5717 // No throughput diff was found with backward walk bound between 3-8.
5718 static const int optEarlyPropRecurBound = 5;
5720 enum class optPropKind
5728 bool gtIsVtableRef(GenTreePtr tree);
5729 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5730 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5731 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5732 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5733 bool optEarlyPropRewriteTree(GenTreePtr tree);
5734 bool optDoEarlyPropForBlock(BasicBlock* block);
5735 bool optDoEarlyPropForFunc();
5736 void optEarlyProp();
5737 void optFoldNullCheck(GenTreePtr tree);
5738 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5741 /**************************************************************************
5742 * Value/Assertion propagation
5743 *************************************************************************/
5745 // Data structures for assertion prop
5746 BitVecTraits* apTraits;
5749 enum optAssertionKind
5766 O1K_CONSTANT_LOOP_BND,
5787 optAssertionKind assertionKind;
5790 unsigned lclNum; // assigned to or property of this local var number
5798 struct AssertionDscOp1
5800 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5807 struct AssertionDscOp2
5809 optOp2Kind kind; // a const or copy assignment
5813 ssize_t iconVal; // integer
5814 unsigned iconFlags; // gtFlags
5816 struct Range // integer subrange
5830 bool IsCheckedBoundArithBound()
5832 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5834 bool IsCheckedBoundBound()
5836 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5838 bool IsConstantBound()
5840 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5841 op1.kind == O1K_CONSTANT_LOOP_BND);
5843 bool IsBoundsCheckNoThrow()
5845 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5848 bool IsCopyAssertion()
5850 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5853 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5855 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5856 a1->op2.kind == a2->op2.kind;
5859 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5861 if (kind == OAK_EQUAL)
5863 return kind2 == OAK_NOT_EQUAL;
5865 else if (kind == OAK_NOT_EQUAL)
5867 return kind2 == OAK_EQUAL;
5872 static ssize_t GetLowerBoundForIntegralType(var_types type)
5892 static ssize_t GetUpperBoundForIntegralType(var_types type)
5916 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5918 if (op1.kind != that->op1.kind)
5922 else if (op1.kind == O1K_ARR_BND)
5925 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5929 return ((vnBased && (op1.vn == that->op1.vn)) ||
5930 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5934 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5936 if (op2.kind != that->op2.kind)
5942 case O2K_IND_CNS_INT:
5944 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5946 case O2K_CONST_LONG:
5947 return (op2.lconVal == that->op2.lconVal);
5949 case O2K_CONST_DOUBLE:
5950 // exact match because of positive and negative zero.
5951 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5953 case O2K_LCLVAR_COPY:
5955 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5956 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5959 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5962 // we will return false
5966 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5972 bool Complementary(AssertionDsc* that, bool vnBased)
5974 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5975 HasSameOp2(that, vnBased);
5978 bool Equals(AssertionDsc* that, bool vnBased)
5980 if (assertionKind != that->assertionKind)
5984 else if (assertionKind == OAK_NO_THROW)
5986 assert(op2.kind == O2K_INVALID);
5987 return HasSameOp1(that, vnBased);
5991 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5997 static fgWalkPreFn optAddCopiesCallback;
5998 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5999 unsigned optAddCopyLclNum;
6000 GenTreePtr optAddCopyAsgnNode;
6002 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6003 bool optAssertionPropagated; // set to true if we modified the trees
6004 bool optAssertionPropagatedCurrentStmt;
6006 GenTreePtr optAssertionPropCurrentTree;
6008 AssertionIndex* optComplementaryAssertionMap;
6009 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6010 // using the value of a local var) for each local var
6011 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6012 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6013 AssertionIndex optMaxAssertionCount;
6016 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6017 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6018 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6019 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6020 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6021 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6023 AssertionIndex GetAssertionCount()
6025 return optAssertionCount;
6027 ASSERT_TP* bbJtrueAssertionOut;
6028 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6029 ValueNumToAssertsMap;
6030 ValueNumToAssertsMap* optValueNumToAsserts;
6032 // Assertion prop helpers.
6033 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6034 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6035 void optAssertionInit(bool isLocalProp);
6036 void optAssertionTraitsInit(AssertionIndex assertionCount);
6037 #if LOCAL_ASSERTION_PROP
6038 void optAssertionReset(AssertionIndex limit);
6039 void optAssertionRemove(AssertionIndex index);
6042 // Assertion prop data flow functions.
6043 void optAssertionPropMain();
6044 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6045 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6046 ASSERT_TP* optInitAssertionDataflowFlags();
6047 ASSERT_TP* optComputeAssertionGen();
6049 // Assertion Gen functions.
6050 void optAssertionGen(GenTreePtr tree);
6051 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6052 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6053 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6054 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6055 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6056 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6058 // Assertion creation functions.
6059 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6060 AssertionIndex optCreateAssertion(GenTreePtr op1,
6062 optAssertionKind assertionKind,
6063 AssertionDsc* assertion);
6064 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6066 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6067 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6068 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6070 void optPrintVnAssertionMapping();
6072 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6074 // Used for respective assertion propagations.
6075 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6076 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6077 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6078 bool optAssertionIsNonNull(GenTreePtr op,
6079 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6081 // Used for Relop propagation.
6082 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6083 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6084 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6086 // Assertion prop for lcl var functions.
6087 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6088 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6090 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6091 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6092 const GenTreePtr tree,
6093 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6094 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6096 // Assertion propagation functions.
6097 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6098 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6099 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6100 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6101 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6102 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6103 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6104 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6105 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6106 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6107 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6108 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6110 // Implied assertion functions.
6111 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6112 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6113 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6114 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6117 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6118 void optDebugCheckAssertion(AssertionDsc* assertion);
6119 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6121 void optAddCopies();
6122 #endif // ASSERTION_PROP
6124 /**************************************************************************
6126 *************************************************************************/
6129 struct LoopCloneVisitorInfo
6131 LoopCloneContext* context;
6134 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6135 : context(context), loopNum(loopNum), stmt(nullptr)
6140 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6141 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6142 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6143 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6144 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6145 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6146 void optObtainLoopCloningOpts(LoopCloneContext* context);
6147 bool optIsLoopClonable(unsigned loopInd);
6149 bool optCanCloneLoops();
6152 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6154 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6155 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6156 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6157 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6161 void optInsertLoopCloningStress(BasicBlock* head);
6163 #if COUNT_RANGECHECKS
6164 static unsigned optRangeChkRmv;
6165 static unsigned optRangeChkAll;
6174 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6179 RngChkDsc* rcdNextInBucket; // used by the hash table
6181 unsigned short rcdHashValue; // to make matching faster
6182 unsigned short rcdIndex; // 0..optRngChkCount-1
6184 GenTreePtr rcdTree; // the array index tree
6187 unsigned optRngChkCount;
6188 static const size_t optRngChkHashSize;
6190 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6191 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6193 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6196 bool optLoopsMarked;
6199 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6200 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6204 XX Does the register allocation and puts the remaining lclVars on the stack XX
6206 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6207 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6211 #ifndef LEGACY_BACKEND
6216 #else // LEGACY_BACKEND
6221 #endif // LEGACY_BACKEND
6223 #ifdef LEGACY_BACKEND
6225 void raAssignVars(); // register allocation
6226 #endif // LEGACY_BACKEND
6228 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6230 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6232 void raMarkStkVars();
6235 // Some things are used by both LSRA and regpredict allocators.
6237 FrameType rpFrameType;
6238 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6240 #ifdef LEGACY_BACKEND
6241 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6243 #endif // LEGACY_BACKEND
6245 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6247 #if FEATURE_FP_REGALLOC
6248 enum enumConfigRegisterFP
6250 CONFIG_REGISTER_FP_NONE = 0x0,
6251 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6252 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6253 CONFIG_REGISTER_FP_FULL = 0x3,
6255 enumConfigRegisterFP raConfigRegisterFP();
6256 #endif // FEATURE_FP_REGALLOC
6259 regMaskTP raConfigRestrictMaskFP();
6262 #ifndef LEGACY_BACKEND
6263 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6264 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6265 #else // LEGACY_BACKEND
6266 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6267 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6268 bool raNewBlocks; // True is we added killing blocks for FPU registers
6269 unsigned rpPasses; // Number of passes made by the register predicter
6270 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6271 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6272 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6273 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6274 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6275 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6276 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6277 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6278 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6279 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6280 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6281 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6283 bool rpRegAllocDone; // Set to true after we have completed register allocation
6285 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6287 void raSetupArgMasks(RegState* r);
6289 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6291 void raDumpVarIntf(); // Dump the variable to variable interference graph
6292 void raDumpRegIntf(); // Dump the variable to register interference graph
6294 void raAdjustVarIntf();
6296 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6298 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6300 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6301 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6303 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6305 static fgWalkPreFn rpMarkRegIntf;
6307 regMaskTP rpPredictAddressMode(
6308 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6310 void rpPredictRefAssign(unsigned lclNum);
6312 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6314 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6316 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6318 void rpPredictRegUse(); // Entry point
6320 unsigned raPredictTreeRegUse(GenTreePtr tree);
6321 unsigned raPredictListRegUse(GenTreePtr list);
6323 void raSetRegVarOrder(var_types regType,
6324 regNumber* customVarOrder,
6325 unsigned* customVarOrderSize,
6327 regMaskTP avoidReg);
6329 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6330 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6331 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6332 void raAddToStkPredict(unsigned val)
6334 unsigned newStkPredict = rpStkPredict + val;
6335 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6336 rpStkPredict = UINT_MAX - 1;
6338 rpStkPredict = newStkPredict;
6342 #if !FEATURE_FP_REGALLOC
6343 void raDispFPlifeInfo();
6347 regMaskTP genReturnRegForTree(GenTreePtr tree);
6348 #endif // LEGACY_BACKEND
6350 /* raIsVarargsStackArg is called by raMaskStkVars and by
6351 lvaSortByRefCount. It identifies the special case
6352 where a varargs function has a parameter passed on the
6353 stack, other than the special varargs handle. Such parameters
6354 require special treatment, because they cannot be tracked
6355 by the GC (their offsets in the stack are not known
6359 bool raIsVarargsStackArg(unsigned lclNum)
6363 LclVarDsc* varDsc = &lvaTable[lclNum];
6365 assert(varDsc->lvIsParam);
6367 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6369 #else // _TARGET_X86_
6373 #endif // _TARGET_X86_
6376 #ifdef LEGACY_BACKEND
6377 // Records the current prediction, if it's better than any previous recorded prediction.
6378 void rpRecordPrediction();
6379 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6380 void rpUseRecordedPredictionIfBetter();
6382 // Data members used in the methods above.
6383 unsigned rpBestRecordedStkPredict;
6384 struct VarRegPrediction
6386 bool m_isEnregistered;
6387 regNumberSmall m_regNum;
6388 regNumberSmall m_otherReg;
6390 VarRegPrediction* rpBestRecordedPrediction;
6391 #endif // LEGACY_BACKEND
6394 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6395 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6399 XX Get to the class and method info from the Execution Engine given XX
6400 XX tokens for the class and method XX
6402 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6403 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6407 /* These are the different addressing modes used to access a local var.
6408 * The JIT has to report the location of the locals back to the EE
6409 * for debugging purposes.
6415 VLT_REG_BYREF, // this type is currently only used for value types on X64
6418 VLT_STK_BYREF, // this type is currently only used for value types on X64
6432 siVarLocType vlType;
6435 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6437 // VLT_REG_BYREF -- the specified register contains the address of the variable
6445 // VLT_STK -- Any 32 bit value which is on the stack
6446 // eg. [ESP+0x20], or [EBP-0x28]
6447 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6448 // eg. mov EAX, [ESP+0x20]; [EAX]
6452 regNumber vlsBaseReg;
6453 NATIVE_OFFSET vlsOffset;
6456 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6465 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6466 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6474 regNumber vlrssBaseReg;
6475 NATIVE_OFFSET vlrssOffset;
6479 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6480 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6486 regNumber vlsrsBaseReg;
6487 NATIVE_OFFSET vlsrsOffset;
6493 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6494 // eg 2 DWords at [ESP+0x10]
6498 regNumber vls2BaseReg;
6499 NATIVE_OFFSET vls2Offset;
6502 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6503 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6510 // VLT_FIXED_VA -- fixed argument of a varargs function.
6511 // The argument location depends on the size of the variable
6512 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6513 // location of the first arg. This argument can then be accessed
6514 // relative to the position of the first arg
6518 unsigned vlfvOffset;
6525 void* rpValue; // pointer to the in-process
6526 // location of the value.
6532 bool vlIsInReg(regNumber reg);
6533 bool vlIsOnStk(regNumber reg, signed offset);
6536 /*************************************************************************/
6541 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6542 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6543 CORINFO_CALLINFO_FLAGS flags,
6544 CORINFO_CALL_INFO* pResult);
6545 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6547 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6548 CORINFO_ACCESS_FLAGS flags,
6549 CORINFO_FIELD_INFO* pResult);
6553 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6555 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6557 bool IsSuperPMIException(unsigned code)
6559 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6561 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6562 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6563 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6564 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6565 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6566 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6567 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6568 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6572 case EXCEPTIONCODE_DebugBreakorAV:
6573 case EXCEPTIONCODE_MC:
6574 case EXCEPTIONCODE_LWM:
6575 case EXCEPTIONCODE_SASM:
6576 case EXCEPTIONCODE_SSYM:
6577 case EXCEPTIONCODE_CALLUTILS:
6578 case EXCEPTIONCODE_TYPEUTILS:
6579 case EXCEPTIONCODE_ASSERT:
6586 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6587 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6589 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6590 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6593 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6594 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6595 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6597 // VOM info, method sigs
6599 void eeGetSig(unsigned sigTok,
6600 CORINFO_MODULE_HANDLE scope,
6601 CORINFO_CONTEXT_HANDLE context,
6602 CORINFO_SIG_INFO* retSig);
6604 void eeGetCallSiteSig(unsigned sigTok,
6605 CORINFO_MODULE_HANDLE scope,
6606 CORINFO_CONTEXT_HANDLE context,
6607 CORINFO_SIG_INFO* retSig);
6609 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6611 // Method entry-points, instrs
6613 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6615 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6617 CORINFO_EE_INFO eeInfo;
6618 bool eeInfoInitialized;
6620 CORINFO_EE_INFO* eeGetEEInfo();
6622 // Gets the offset of a SDArray's first element
6623 unsigned eeGetArrayDataOffset(var_types type);
6624 // Gets the offset of a MDArray's first element
6625 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6627 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6629 // Returns the page size for the target machine as reported by the EE.
6630 inline size_t eeGetPageSize()
6632 return eeGetEEInfo()->osPageSize;
6635 // Returns the frame size at which we will generate a loop to probe the stack.
6636 inline size_t getVeryLargeFrameSize()
6639 // The looping probe code is 40 bytes, whereas the straight-line probing for
6640 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6641 // or greater, to generate smaller code.
6642 return 2 * eeGetPageSize();
6644 return 3 * eeGetPageSize();
6648 //------------------------------------------------------------------------
6649 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6651 // It represents Abi and target specific registers for the parameter.
6653 class VirtualStubParamInfo
6656 VirtualStubParamInfo(bool isCoreRTABI)
6658 #if defined(_TARGET_X86_)
6661 #elif defined(_TARGET_AMD64_)
6672 #elif defined(_TARGET_ARM_)
6683 #elif defined(_TARGET_ARM64_)
6687 #error Unsupported or unset target architecture
6690 #ifdef LEGACY_BACKEND
6691 #if defined(_TARGET_X86_)
6692 predict = PREDICT_REG_EAX;
6693 #elif defined(_TARGET_ARM_)
6694 predict = PREDICT_REG_R4;
6696 #error Unsupported or unset target architecture
6698 #endif // LEGACY_BACKEND
6701 regNumber GetReg() const
6706 _regMask_enum GetRegMask() const
6711 #ifdef LEGACY_BACKEND
6712 rpPredictReg GetPredict() const
6720 _regMask_enum regMask;
6722 #ifdef LEGACY_BACKEND
6723 rpPredictReg predict;
6727 VirtualStubParamInfo* virtualStubParamInfo;
6729 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6731 return eeGetEEInfo()->targetAbi == abi;
6734 inline bool generateCFIUnwindCodes()
6736 #ifdef UNIX_AMD64_ABI
6737 return IsTargetAbi(CORINFO_CORERT_ABI);
6745 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6747 // Debugging support - Line number info
6749 void eeGetStmtOffsets();
6751 unsigned eeBoundariesCount;
6753 struct boundariesDsc
6755 UNATIVE_OFFSET nativeIP;
6757 unsigned sourceReason;
6758 } * eeBoundaries; // Boundaries to report to EE
6759 void eeSetLIcount(unsigned count);
6760 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6764 static void eeDispILOffs(IL_OFFSET offs);
6765 static void eeDispLineInfo(const boundariesDsc* line);
6766 void eeDispLineInfos();
6769 // Debugging support - Local var info
6773 unsigned eeVarsCount;
6775 struct VarResultInfo
6777 UNATIVE_OFFSET startOffset;
6778 UNATIVE_OFFSET endOffset;
6782 void eeSetLVcount(unsigned count);
6783 void eeSetLVinfo(unsigned which,
6784 UNATIVE_OFFSET startOffs,
6785 UNATIVE_OFFSET length,
6790 const siVarLoc& loc);
6794 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6795 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6798 // ICorJitInfo wrappers
6800 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6802 void eeAllocUnwindInfo(BYTE* pHotCode,
6808 CorJitFuncKind funcKind);
6810 void eeSetEHcount(unsigned cEH);
6812 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6814 WORD eeGetRelocTypeHint(void* target);
6816 // ICorStaticInfo wrapper functions
6818 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6820 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6822 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6825 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6826 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6827 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6828 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6830 template <typename ParamType>
6831 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6833 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6836 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6838 // Utility functions
6840 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6843 const wchar_t* eeGetCPString(size_t stringHandle);
6846 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6848 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6849 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6851 static fgWalkPreFn CountSharedStaticHelper;
6852 static bool IsSharedStaticHelper(GenTreePtr tree);
6853 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6854 static bool IsGcSafePoint(GenTreePtr tree);
6856 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6857 // returns true/false if 'field' is a Jit Data offset
6858 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6859 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6860 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6862 /*****************************************************************************/
6867 enum TEMP_USAGE_TYPE
6873 static var_types tmpNormalizeType(var_types type);
6874 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6875 void tmpRlsTemp(TempDsc* temp);
6876 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6879 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6880 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6884 bool tmpAllFree() const;
6887 #ifndef LEGACY_BACKEND
6888 void tmpPreAllocateTemps(var_types type, unsigned count);
6889 #endif // !LEGACY_BACKEND
6892 #ifdef LEGACY_BACKEND
6893 unsigned tmpIntSpillMax; // number of int-sized spill temps
6894 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6895 #endif // LEGACY_BACKEND
6897 unsigned tmpCount; // Number of temps
6898 unsigned tmpSize; // Size of all the temps
6901 // Used by RegSet::rsSpillChk()
6902 unsigned tmpGetCount; // Temps which haven't been released yet
6905 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6907 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6908 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6911 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6912 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6916 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6917 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6921 CodeGenInterface* codeGen;
6923 // The following holds information about instr offsets in terms of generated code.
6927 IPmappingDsc* ipmdNext; // next line# record
6928 IL_OFFSETX ipmdILoffsx; // the instr offset
6929 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6930 bool ipmdIsLabel; // Can this code be a branch label?
6933 // Record the instr offset mapping to the generated code
6935 IPmappingDsc* genIPmappingList;
6936 IPmappingDsc* genIPmappingLast;
6938 // Managed RetVal - A side hash table meant to record the mapping from a
6939 // GT_CALL node to its IL offset. This info is used to emit sequence points
6940 // that can be used by debugger to determine the native offset at which the
6941 // managed RetVal will be available.
6943 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6944 // favor of a side table for two reasons: 1) We need IL offset for only those
6945 // GT_CALL nodes (created during importation) that correspond to an IL call and
6946 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6947 // structure and IL offset is needed only when generating debuggable code. Therefore
6948 // it is desirable to avoid memory size penalty in retail scenarios.
6949 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6950 CallSiteILOffsetTable;
6951 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6953 unsigned genReturnLocal; // Local number for the return value when applicable.
6954 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6956 // The following properties are part of CodeGenContext. Getters are provided here for
6957 // convenience and backward compatibility, but the properties can only be set by invoking
6958 // the setter on CodeGenContext directly.
6960 __declspec(property(get = getEmitter)) emitter* genEmitter;
6961 emitter* getEmitter()
6963 return codeGen->getEmitter();
6966 const bool isFramePointerUsed()
6968 return codeGen->isFramePointerUsed();
6971 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6972 bool getInterruptible()
6974 return codeGen->genInterruptible;
6976 void setInterruptible(bool value)
6978 codeGen->setInterruptible(value);
6982 const bool genDoubleAlign()
6984 return codeGen->doDoubleAlign();
6986 DWORD getCanDoubleAlign();
6987 bool shouldDoubleAlign(unsigned refCntStk,
6989 unsigned refCntWtdReg,
6990 unsigned refCntStkParam,
6991 unsigned refCntWtdStkDbl);
6992 #endif // DOUBLE_ALIGN
6994 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6995 bool getFullPtrRegMap()
6997 return codeGen->genFullPtrRegMap;
6999 void setFullPtrRegMap(bool value)
7001 codeGen->setFullPtrRegMap(value);
7004 // Things that MAY belong either in CodeGen or CodeGenContext
7006 #if FEATURE_EH_FUNCLETS
7007 FuncInfoDsc* compFuncInfos;
7008 unsigned short compCurrFuncIdx;
7009 unsigned short compFuncInfoCount;
7011 unsigned short compFuncCount()
7013 assert(fgFuncletsCreated);
7014 return compFuncInfoCount;
7017 #else // !FEATURE_EH_FUNCLETS
7019 // This is a no-op when there are no funclets!
7020 void genUpdateCurrentFunclet(BasicBlock* block)
7025 FuncInfoDsc compFuncInfoRoot;
7027 static const unsigned compCurrFuncIdx = 0;
7029 unsigned short compFuncCount()
7034 #endif // !FEATURE_EH_FUNCLETS
7036 FuncInfoDsc* funCurrentFunc();
7037 void funSetCurrentFunc(unsigned funcIdx);
7038 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7039 unsigned int funGetFuncIdx(BasicBlock* block);
7043 VARSET_TP compCurLife; // current live variables
7044 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7046 template <bool ForCodeGen>
7047 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7049 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7051 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7054 template <bool ForCodeGen>
7055 void compUpdateLife(GenTreePtr tree);
7057 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7058 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7059 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7060 template <bool ForCodeGen>
7061 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7063 template <bool ForCodeGen>
7064 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7066 // Gets a register mask that represent the kill set for a helper call since
7067 // not all JIT Helper calls follow the standard ABI on the target architecture.
7068 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7070 // Gets a register mask that represent the kill set for a NoGC helper call.
7071 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7074 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7075 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7076 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7077 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7078 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7079 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7080 #endif // _TARGET_ARM_
7082 // 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
7084 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7086 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7087 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7088 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7089 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7090 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7091 // for the tracked var indices of the field vars, as in a live var set).
7092 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7094 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7096 if (m_promotedStructDeathVars == nullptr)
7098 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7100 return m_promotedStructDeathVars;
7104 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7105 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7109 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7110 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7113 #if !defined(__GNUC__)
7114 #pragma region Unwind information
7119 // Infrastructure functions: start/stop/reserve/emit.
7122 void unwindBegProlog();
7123 void unwindEndProlog();
7124 void unwindBegEpilog();
7125 void unwindEndEpilog();
7126 void unwindReserve();
7127 void unwindEmit(void* pHotCode, void* pColdCode);
7130 // Specific unwind information functions: called by code generation to indicate a particular
7131 // prolog or epilog unwindable instruction has been generated.
7134 void unwindPush(regNumber reg);
7135 void unwindAllocStack(unsigned size);
7136 void unwindSetFrameReg(regNumber reg, unsigned offset);
7137 void unwindSaveReg(regNumber reg, unsigned offset);
7139 #if defined(_TARGET_ARM_)
7140 void unwindPushMaskInt(regMaskTP mask);
7141 void unwindPushMaskFloat(regMaskTP mask);
7142 void unwindPopMaskInt(regMaskTP mask);
7143 void unwindPopMaskFloat(regMaskTP mask);
7144 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7145 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7146 // called via unwindPadding().
7147 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7148 // instruction and the current location.
7149 #endif // _TARGET_ARM_
7151 #if defined(_TARGET_ARM64_)
7153 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7154 // instruction and the current location.
7155 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7156 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7157 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7158 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7159 void unwindSaveNext(); // unwind code: save_next
7160 void unwindReturn(regNumber reg); // ret lr
7161 #endif // defined(_TARGET_ARM64_)
7164 // Private "helper" functions for the unwind implementation.
7168 #if FEATURE_EH_FUNCLETS
7169 void unwindGetFuncLocations(FuncInfoDsc* func,
7170 bool getHotSectionData,
7171 /* OUT */ emitLocation** ppStartLoc,
7172 /* OUT */ emitLocation** ppEndLoc);
7173 #endif // FEATURE_EH_FUNCLETS
7175 void unwindReserveFunc(FuncInfoDsc* func);
7176 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7178 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7180 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7181 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7183 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7185 #if defined(_TARGET_AMD64_)
7187 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7189 void unwindBegPrologWindows();
7190 void unwindPushWindows(regNumber reg);
7191 void unwindAllocStackWindows(unsigned size);
7192 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7193 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7195 #ifdef UNIX_AMD64_ABI
7196 void unwindBegPrologCFI();
7197 void unwindPushCFI(regNumber reg);
7198 void unwindAllocStackCFI(unsigned size);
7199 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7200 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7201 int mapRegNumToDwarfReg(regNumber reg);
7202 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7203 #endif // UNIX_AMD64_ABI
7204 #elif defined(_TARGET_ARM_)
7206 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7207 void unwindPushPopMaskFloat(regMaskTP mask);
7208 void unwindSplit(FuncInfoDsc* func);
7210 #endif // _TARGET_ARM_
7212 #if !defined(__GNUC__)
7213 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7217 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7218 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7222 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7223 XX that contains the distinguished, well-known SIMD type definitions). XX
7225 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7226 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7229 // Get highest available instruction set for floating point codegen
7230 InstructionSet getFloatingPointInstructionSet()
7232 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7235 return InstructionSet_AVX;
7240 return InstructionSet_SSE3_4;
7244 assert(canUseSSE2());
7245 return InstructionSet_SSE2;
7247 assert(!"getFPInstructionSet() is not implemented for target arch");
7249 return InstructionSet_NONE;
7253 // Get highest available instruction set for SIMD codegen
7254 InstructionSet getSIMDInstructionSet()
7256 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7257 return getFloatingPointInstructionSet();
7259 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7261 return InstructionSet_NONE;
7267 // Should we support SIMD intrinsics?
7270 // Have we identified any SIMD types?
7271 // This is currently used by struct promotion to avoid getting type information for a struct
7272 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7274 bool _usesSIMDTypes;
7275 bool usesSIMDTypes()
7277 return _usesSIMDTypes;
7279 void setUsesSIMDTypes(bool value)
7281 _usesSIMDTypes = value;
7284 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7285 // that require indexed access to the individual fields of the vector, which is not well supported
7286 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7287 unsigned lvaSIMDInitTempVarNum;
7290 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7291 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7292 CORINFO_CLASS_HANDLE SIMDIntHandle;
7293 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7294 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7295 CORINFO_CLASS_HANDLE SIMDShortHandle;
7296 CORINFO_CLASS_HANDLE SIMDByteHandle;
7297 CORINFO_CLASS_HANDLE SIMDLongHandle;
7298 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7299 CORINFO_CLASS_HANDLE SIMDULongHandle;
7300 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7301 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7302 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7303 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7305 // Get the handle for a SIMD type.
7306 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7308 if (simdBaseType == TYP_FLOAT)
7313 return SIMDVector2Handle;
7315 return SIMDVector3Handle;
7317 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7319 return SIMDVector4Handle;
7328 assert(simdType == getSIMDVectorType());
7329 switch (simdBaseType)
7332 return SIMDFloatHandle;
7334 return SIMDDoubleHandle;
7336 return SIMDIntHandle;
7338 return SIMDUShortHandle;
7340 return SIMDUShortHandle;
7342 return SIMDUByteHandle;
7344 return SIMDShortHandle;
7346 return SIMDByteHandle;
7348 return SIMDLongHandle;
7350 return SIMDUIntHandle;
7352 return SIMDULongHandle;
7354 assert(!"Didn't find a class handle for simdType");
7356 return NO_CLASS_HANDLE;
7360 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7361 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7362 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7364 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7365 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7366 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7367 bool isSIMDTypeLocal(GenTree* tree)
7369 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7372 // Returns true if the type of the tree is a byref of TYP_SIMD
7373 bool isAddrOfSIMDType(GenTree* tree)
7375 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7377 switch (tree->OperGet())
7380 return varTypeIsSIMD(tree->gtGetOp1());
7382 case GT_LCL_VAR_ADDR:
7383 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7386 return isSIMDTypeLocal(tree);
7393 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7395 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7396 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7397 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7400 // Returns base type of a TYP_SIMD local.
7401 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7402 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7404 if (isSIMDTypeLocal(tree))
7406 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7412 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7414 return info.compCompHnd->isInSIMDModule(clsHnd);
7417 bool isSIMDClass(typeInfo* pTypeInfo)
7419 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7422 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7423 // if it is not a SIMD type or is an unsupported base type.
7424 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7426 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7428 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7431 // Get SIMD Intrinsic info given the method handle.
7432 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7433 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7434 CORINFO_METHOD_HANDLE methodHnd,
7435 CORINFO_SIG_INFO* sig,
7438 var_types* baseType,
7439 unsigned* sizeBytes);
7441 // Pops and returns GenTree node from importers type stack.
7442 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7443 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7445 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7446 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7448 // Creates a GT_SIMD tree for Select operation
7449 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7451 unsigned simdVectorSize,
7456 // Creates a GT_SIMD tree for Min/Max operation
7457 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7458 CORINFO_CLASS_HANDLE typeHnd,
7460 unsigned simdVectorSize,
7464 // Transforms operands and returns the SIMD intrinsic to be applied on
7465 // transformed operands to obtain given relop result.
7466 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7467 CORINFO_CLASS_HANDLE typeHnd,
7468 unsigned simdVectorSize,
7469 var_types* baseType,
7473 // Creates a GT_SIMD tree for Abs intrinsic.
7474 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7476 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7477 // Transforms operands and returns the SIMD intrinsic to be applied on
7478 // transformed operands to obtain == comparison result.
7479 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7480 unsigned simdVectorSize,
7484 // Transforms operands and returns the SIMD intrinsic to be applied on
7485 // transformed operands to obtain > comparison result.
7486 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7487 unsigned simdVectorSize,
7491 // Transforms operands and returns the SIMD intrinsic to be applied on
7492 // transformed operands to obtain >= comparison result.
7493 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7494 unsigned simdVectorSize,
7498 // Transforms operands and returns the SIMD intrinsic to be applied on
7499 // transformed operands to obtain >= comparison result in case of int32
7500 // and small int base type vectors.
7501 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7502 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7503 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7505 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7506 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7507 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7508 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7509 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7511 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7512 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7513 GenTreePtr newobjThis,
7514 CORINFO_CLASS_HANDLE clsHnd,
7515 CORINFO_METHOD_HANDLE method,
7516 CORINFO_SIG_INFO* sig,
7519 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7521 // Whether SIMD vector occupies part of SIMD register.
7522 // SSE2: vector2f/3f are considered sub register SIMD types.
7523 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7524 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7526 unsigned sizeBytes = 0;
7527 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7528 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7531 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7533 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7536 // Get the type for the hardware SIMD vector.
7537 // This is the maximum SIMD type supported for this target.
7538 var_types getSIMDVectorType()
7540 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7547 assert(canUseSSE2());
7551 assert(!"getSIMDVectorType() unimplemented on target arch");
7556 // Get the size of the SIMD type in bytes
7557 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7559 unsigned sizeBytes = 0;
7560 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7564 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7565 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7567 // Get the the number of elements of basetype of SIMD vector given by its type handle
7568 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7570 // Get preferred alignment of SIMD type.
7571 int getSIMDTypeAlignment(var_types simdType);
7573 // Get the number of bytes in a SIMD Vector for the current compilation.
7574 unsigned getSIMDVectorRegisterByteLength()
7576 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7579 return YMM_REGSIZE_BYTES;
7583 assert(canUseSSE2());
7584 return XMM_REGSIZE_BYTES;
7587 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7592 // The minimum and maximum possible number of bytes in a SIMD vector.
7593 unsigned int maxSIMDStructBytes()
7595 return getSIMDVectorRegisterByteLength();
7597 unsigned int minSIMDStructBytes()
7599 return emitTypeSize(TYP_SIMD8);
7602 #ifdef FEATURE_AVX_SUPPORT
7603 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7604 static const unsigned maxPossibleSIMDStructBytes = 32;
7605 #else // !FEATURE_AVX_SUPPORT
7606 static const unsigned maxPossibleSIMDStructBytes = 16;
7607 #endif // !FEATURE_AVX_SUPPORT
7609 // Returns the codegen type for a given SIMD size.
7610 var_types getSIMDTypeForSize(unsigned size)
7612 var_types simdType = TYP_UNDEF;
7615 simdType = TYP_SIMD8;
7617 else if (size == 12)
7619 simdType = TYP_SIMD12;
7621 else if (size == 16)
7623 simdType = TYP_SIMD16;
7625 #ifdef FEATURE_AVX_SUPPORT
7626 else if (size == 32)
7628 simdType = TYP_SIMD32;
7630 #endif // FEATURE_AVX_SUPPORT
7633 noway_assert(!"Unexpected size for SIMD type");
7638 unsigned getSIMDInitTempVarNum()
7640 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7642 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7643 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7645 return lvaSIMDInitTempVarNum;
7648 #endif // FEATURE_SIMD
7651 //------------------------------------------------------------------------
7652 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7654 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7655 // candidate for enregistration.
7657 unsigned largestEnregisterableStructSize()
7660 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7661 if (vectorRegSize > TARGET_POINTER_SIZE)
7663 return vectorRegSize;
7666 #endif // FEATURE_SIMD
7668 return TARGET_POINTER_SIZE;
7673 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7674 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7675 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7677 // Is this var is of type simd struct?
7678 bool lclVarIsSIMDType(unsigned varNum)
7680 LclVarDsc* varDsc = lvaTable + varNum;
7681 return varDsc->lvIsSIMDType();
7684 // Is this Local node a SIMD local?
7685 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7687 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7690 // Returns true if the TYP_SIMD locals on stack are aligned at their
7691 // preferred byte boundary specified by getSIMDTypeAlignment().
7693 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7694 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7695 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7696 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7697 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7698 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7699 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7702 bool isSIMDTypeLocalAligned(unsigned varNum)
7704 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7705 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7708 int off = lvaFrameAddress(varNum, &ebpBased);
7709 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7710 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7711 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7714 #endif // FEATURE_SIMD
7719 // Whether SSE2 is available
7720 bool canUseSSE2() const
7722 #ifdef _TARGET_XARCH_
7723 return opts.compCanUseSSE2;
7729 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7730 bool CanUseSSE3_4() const
7732 #ifdef _TARGET_XARCH_
7733 return opts.compCanUseSSE3_4;
7739 bool canUseAVX() const
7741 #ifdef FEATURE_AVX_SUPPORT
7742 return opts.compCanUseAVX;
7749 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7750 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7754 XX Generic info about the compilation and the method being compiled. XX
7755 XX It is responsible for driving the other phases. XX
7756 XX It is also responsible for all the memory management. XX
7758 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7759 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7763 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7765 InlineResult* compInlineResult; // The result of importing the inlinee method.
7767 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7768 bool compJmpOpUsed; // Does the method do a JMP
7769 bool compLongUsed; // Does the method use TYP_LONG
7770 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7771 bool compTailCallUsed; // Does the method do a tailcall
7772 bool compLocallocUsed; // Does the method use localloc.
7773 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7774 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7775 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7777 // NOTE: These values are only reliable after
7778 // the importing is completely finished.
7780 #ifdef LEGACY_BACKEND
7781 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7782 // we can iterate over these efficiently.
7785 #if CPU_USES_BLOCK_MOVE
7786 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7790 // State information - which phases have completed?
7791 // These are kept together for easy discoverability
7793 bool bRangeAllowStress;
7794 bool compCodeGenDone;
7795 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7796 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7797 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7798 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7801 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7802 bool fgLocalVarLivenessChanged;
7804 bool compStackProbePrologDone;
7806 #ifndef LEGACY_BACKEND
7808 #endif // !LEGACY_BACKEND
7809 bool compRationalIRForm;
7811 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7813 bool compGeneratingProlog;
7814 bool compGeneratingEpilog;
7815 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7816 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7817 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7818 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7819 bool getNeedsGSSecurityCookie() const
7821 return compNeedsGSSecurityCookie;
7823 void setNeedsGSSecurityCookie()
7825 compNeedsGSSecurityCookie = true;
7828 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7829 // frame layout calculations, this is the level we are currently
7832 //---------------------------- JITing options -----------------------------
7845 JitFlags* jitFlags; // all flags passed from the EE
7846 unsigned compFlags; // method attributes
7848 codeOptimize compCodeOpt; // what type of code optimizations
7852 #ifdef _TARGET_XARCH_
7853 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7854 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7856 #ifdef FEATURE_AVX_SUPPORT
7857 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7858 #endif // FEATURE_AVX_SUPPORT
7859 #endif // _TARGET_XARCH_
7861 // optimize maximally and/or favor speed over size?
7863 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7864 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7865 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7866 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7867 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7869 // Maximun number of locals before turning off the inlining
7870 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7873 unsigned instrCount;
7874 unsigned lvRefCount;
7875 bool compMinOptsIsSet;
7877 bool compMinOptsIsUsed;
7879 inline bool MinOpts()
7881 assert(compMinOptsIsSet);
7882 compMinOptsIsUsed = true;
7885 inline bool IsMinOptsSet()
7887 return compMinOptsIsSet;
7890 inline bool MinOpts()
7894 inline bool IsMinOptsSet()
7896 return compMinOptsIsSet;
7899 inline void SetMinOpts(bool val)
7901 assert(!compMinOptsIsUsed);
7902 assert(!compMinOptsIsSet || (compMinOpts == val));
7904 compMinOptsIsSet = true;
7907 // true if the CLFLG_* for an optimization is set.
7908 inline bool OptEnabled(unsigned optFlag)
7910 return !!(compFlags & optFlag);
7913 #ifdef FEATURE_READYTORUN_COMPILER
7914 inline bool IsReadyToRun()
7916 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7919 inline bool IsReadyToRun()
7925 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7926 // PInvoke transitions inline (e.g. when targeting CoreRT).
7927 inline bool ShouldUsePInvokeHelpers()
7929 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7932 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7934 inline bool IsReversePInvoke()
7936 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7939 // true if we must generate code compatible with JIT32 quirks
7940 inline bool IsJit32Compat()
7942 #if defined(_TARGET_X86_)
7943 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7949 // true if we must generate code compatible with Jit64 quirks
7950 inline bool IsJit64Compat()
7952 #if defined(_TARGET_AMD64_)
7953 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7954 #elif !defined(FEATURE_CORECLR)
7961 bool compScopeInfo; // Generate the LocalVar info ?
7962 bool compDbgCode; // Generate debugger-friendly code?
7963 bool compDbgInfo; // Gather debugging info?
7966 #ifdef PROFILING_SUPPORTED
7967 bool compNoPInvokeInlineCB;
7969 static const bool compNoPInvokeInlineCB;
7973 bool compGcChecks; // Check arguments and return values to ensure they are sane
7974 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7975 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7979 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7980 // to be allocated on the stack.
7981 // It will be set to true in the following cases:
7982 // 1. When the method being compiled has a declarative security
7983 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7984 // This is also the case when we inject a prolog and epilog in the method.
7986 // 2. When the method being compiled has imperative security (i.e. the method
7987 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7989 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7991 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7992 // which gets reported as a GC root to stackwalker.
7993 // (See also ICodeManager::GetAddrOfSecurityObject.)
7998 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7999 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
8003 #ifdef UNIX_AMD64_ABI
8004 // This flag is indicating if there is a need to align the frame.
8005 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8006 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8007 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8008 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8009 // there are calls and making sure the frame alignment logic is executed.
8010 bool compNeedToAlignFrame;
8011 #endif // UNIX_AMD64_ABI
8013 bool compProcedureSplitting; // Separate cold code from hot code
8015 bool genFPorder; // Preserve FP order (operations are non-commutative)
8016 bool genFPopt; // Can we do frame-pointer-omission optimization?
8017 bool altJit; // True if we are an altjit and are compiling this method
8020 bool optRepeat; // Repeat optimizer phases k times
8024 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8025 bool dspCode; // Display native code generated
8026 bool dspEHTable; // Display the EH table reported to the VM
8027 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8028 bool dspEmit; // Display emitter output
8029 bool dspLines; // Display source-code lines intermixed with native code output
8030 bool dmpHex; // Display raw bytes in hex of native code output
8031 bool varNames; // Display variables names in native code output
8032 bool disAsm; // Display native code as it is generated
8033 bool disAsmSpilled; // Display native code when any register spilling occurs
8034 bool disDiffable; // Makes the Disassembly code 'diff-able'
8035 bool disAsm2; // Display native code after it is generated using external disassembler
8036 bool dspOrder; // Display names of each of the methods that we ngen/jit
8037 bool dspUnwind; // Display the unwind info output
8038 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8039 bool compLongAddress; // Force using large pseudo instructions for long address
8040 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8041 bool dspGCtbls; // Display the GC tables
8045 bool doLateDisasm; // Run the late disassembler
8046 #endif // LATE_DISASM
8048 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8049 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8050 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8051 static const bool dspGCtbls = true;
8054 // We need stack probes to guarantee that we won't trigger a stack overflow
8055 // when calling unmanaged code until they get a chance to set up a frame, because
8056 // the EE will have no idea where it is.
8058 // We will only be doing this currently for hosted environments. Unfortunately
8059 // we need to take care of stubs, so potentially, we will have to do the probes
8060 // for any call. We have a plan for not needing for stubs though
8061 bool compNeedStackProbes;
8063 #ifdef PROFILING_SUPPORTED
8064 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8065 // This option helps make the JIT behave as if it is running under a profiler.
8066 bool compJitELTHookEnabled;
8067 #endif // PROFILING_SUPPORTED
8069 #if FEATURE_TAILCALL_OPT
8070 // Whether opportunistic or implicit tail call optimization is enabled.
8071 bool compTailCallOpt;
8072 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8073 bool compTailCallLoopOpt;
8077 static const bool compUseSoftFP = true;
8078 #else // !ARM_SOFTFP
8079 static const bool compUseSoftFP = false;
8082 GCPollType compGCPollType;
8086 static bool s_pAltJitExcludeAssembliesListInitialized;
8087 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8092 template <typename T>
8095 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8098 template <typename T>
8101 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8104 static int dspTreeID(GenTree* tree)
8106 return tree->gtTreeID;
8108 static void printTreeID(GenTree* tree)
8110 if (tree == nullptr)
8116 printf("[%06d]", dspTreeID(tree));
8123 #define STRESS_MODES \
8127 /* "Variations" stress areas which we try to mix up with each other. */ \
8128 /* These should not be exhaustively used as they might */ \
8129 /* hide/trivialize other areas */ \
8132 STRESS_MODE(DBL_ALN) \
8133 STRESS_MODE(LCL_FLDS) \
8134 STRESS_MODE(UNROLL_LOOPS) \
8135 STRESS_MODE(MAKE_CSE) \
8136 STRESS_MODE(LEGACY_INLINE) \
8137 STRESS_MODE(CLONE_EXPR) \
8138 STRESS_MODE(USE_FCOMI) \
8139 STRESS_MODE(USE_CMOV) \
8141 STRESS_MODE(BB_PROFILE) \
8142 STRESS_MODE(OPT_BOOLS_GC) \
8143 STRESS_MODE(REMORPH_TREES) \
8144 STRESS_MODE(64RSLT_MUL) \
8145 STRESS_MODE(DO_WHILE_LOOPS) \
8146 STRESS_MODE(MIN_OPTS) \
8147 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8148 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8149 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8150 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8151 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8152 STRESS_MODE(NULL_OBJECT_CHECK) \
8153 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8154 STRESS_MODE(RANDOM_INLINE) \
8155 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8156 STRESS_MODE(GENERIC_VARN) \
8158 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8160 STRESS_MODE(COUNT_VARN) \
8162 /* "Check" stress areas that can be exhaustively used if we */ \
8163 /* dont care about performance at all */ \
8165 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8166 STRESS_MODE(CHK_FLOW_UPDATE) \
8167 STRESS_MODE(EMITTER) \
8168 STRESS_MODE(CHK_REIMPORT) \
8169 STRESS_MODE(FLATFP) \
8170 STRESS_MODE(GENERIC_CHECK) \
8175 #define STRESS_MODE(mode) STRESS_##mode,
8182 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8183 BYTE compActiveStressModes[STRESS_COUNT];
8186 #define MAX_STRESS_WEIGHT 100
8188 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8192 bool compInlineStress()
8194 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8197 bool compRandomInlineStress()
8199 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8204 bool compTailCallStress()
8207 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8213 codeOptimize compCodeOpt()
8216 // Switching between size & speed has measurable throughput impact
8217 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8218 // DEBUG, but should generate identical code between CHK & RET builds,
8219 // so that's not acceptable.
8220 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8221 // Investigate the cause of the throughput regression.
8223 return opts.compCodeOpt;
8225 return BLENDED_CODE;
8229 //--------------------- Info about the procedure --------------------------
8233 COMP_HANDLE compCompHnd;
8234 CORINFO_MODULE_HANDLE compScopeHnd;
8235 CORINFO_CLASS_HANDLE compClassHnd;
8236 CORINFO_METHOD_HANDLE compMethodHnd;
8237 CORINFO_METHOD_INFO* compMethodInfo;
8239 BOOL hasCircularClassConstraints;
8240 BOOL hasCircularMethodConstraints;
8242 #if defined(DEBUG) || defined(LATE_DISASM)
8243 const char* compMethodName;
8244 const char* compClassName;
8245 const char* compFullName;
8246 #endif // defined(DEBUG) || defined(LATE_DISASM)
8248 #if defined(DEBUG) || defined(INLINE_DATA)
8249 // Method hash is logcally const, but computed
8251 mutable unsigned compMethodHashPrivate;
8252 unsigned compMethodHash() const;
8253 #endif // defined(DEBUG) || defined(INLINE_DATA)
8255 #ifdef PSEUDORANDOM_NOP_INSERTION
8256 // things for pseudorandom nop insertion
8257 unsigned compChecksum;
8261 // The following holds the FLG_xxxx flags for the method we're compiling.
8264 // The following holds the class attributes for the method we're compiling.
8265 unsigned compClassAttr;
8267 const BYTE* compCode;
8268 IL_OFFSET compILCodeSize; // The IL code size
8269 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8270 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8271 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8272 // (2) the code is hot/cold split, and we issued less code than we expected
8273 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8275 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8276 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8277 bool compIsContextful : 1; // contextful method
8278 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8279 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8280 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8281 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8282 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8284 var_types compRetType; // Return type of the method as declared in IL
8285 var_types compRetNativeType; // Normalized return type as per target arch ABI
8286 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8287 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8288 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8289 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8290 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8291 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8292 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8293 unsigned compMaxStack;
8294 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8295 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8297 unsigned compCallUnmanaged; // count of unmanaged calls
8298 unsigned compLvFrameListRoot; // lclNum for the Frame root
8299 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8300 // You should generally use compHndBBtabCount instead: it is the
8301 // current number of EH clauses (after additions like synchronized
8302 // methods and funclets, and removals like unreachable code deletion).
8304 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8305 // and the VM expects that, or the JIT is a "self-host" compiler
8306 // (e.g., x86 hosted targeting x86) and the VM expects that.
8308 /* The following holds IL scope information about local variables.
8311 unsigned compVarScopesCount;
8312 VarScopeDsc* compVarScopes;
8314 /* The following holds information about instr offsets for
8315 * which we need to report IP-mappings
8318 IL_OFFSET* compStmtOffsets; // sorted
8319 unsigned compStmtOffsetsCount;
8320 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8322 #define CPU_X86 0x0100 // The generic X86 CPU
8323 #define CPU_X86_PENTIUM_4 0x0110
8325 #define CPU_X64 0x0200 // The generic x64 CPU
8326 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8327 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8329 #define CPU_ARM 0x0300 // The generic ARM CPU
8331 unsigned genCPU; // What CPU are we running on
8334 // Returns true if the method being compiled returns a non-void and non-struct value.
8335 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8336 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8337 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8338 // Methods returning such structs are considered to return non-struct return value and
8339 // this method returns true in that case.
8340 bool compMethodReturnsNativeScalarType()
8342 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8345 // Returns true if the method being compiled returns RetBuf addr as its return value
8346 bool compMethodReturnsRetBufAddr()
8348 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8349 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8351 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8352 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8353 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8354 // methods with hidden RetBufArg.
8356 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8357 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8358 // returning the address of RetBuf.
8360 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8361 // to be returned in RAX.
8362 CLANG_FORMAT_COMMENT_ANCHOR;
8364 #ifdef _TARGET_AMD64_
8365 return (info.compRetBuffArg != BAD_VAR_NUM);
8366 #else // !_TARGET_AMD64_
8367 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8368 #endif // !_TARGET_AMD64_
8371 // Returns true if the method returns a value in more than one return register
8372 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8373 // TODO-ARM64: Does this apply for ARM64 too?
8374 bool compMethodReturnsMultiRegRetType()
8376 #if FEATURE_MULTIREG_RET
8377 #if defined(_TARGET_X86_)
8378 // On x86 only 64-bit longs are returned in multiple registers
8379 return varTypeIsLong(info.compRetNativeType);
8380 #else // targets: X64-UNIX, ARM64 or ARM32
8381 // On all other targets that support multireg return values:
8382 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8383 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8384 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8385 #endif // TARGET_XXX
8387 #else // not FEATURE_MULTIREG_RET
8389 // For this architecture there are no multireg returns
8392 #endif // FEATURE_MULTIREG_RET
8395 #if FEATURE_MULTIREG_ARGS
8396 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8397 // return the gcPtr layout for the pointers sized fields
8398 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8399 #endif // FEATURE_MULTIREG_ARGS
8401 // Returns true if the method being compiled returns a value
8402 bool compMethodHasRetVal()
8404 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8405 compMethodReturnsMultiRegRetType();
8410 void compDispLocalVars();
8414 //-------------------------- Global Compiler Data ------------------------------------
8417 static unsigned s_compMethodsCount; // to produce unique label names
8418 unsigned compGenTreeID;
8419 unsigned compBasicBlockID;
8422 BasicBlock* compCurBB; // the current basic block in process
8423 GenTreePtr compCurStmt; // the current statement in process
8425 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8428 // The following is used to create the 'method JIT info' block.
8429 size_t compInfoBlkSize;
8430 BYTE* compInfoBlkAddr;
8432 EHblkDsc* compHndBBtab; // array of EH data
8433 unsigned compHndBBtabCount; // element count of used elements in EH data array
8434 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8436 #if defined(_TARGET_X86_)
8438 //-------------------------------------------------------------------------
8439 // Tracking of region covered by the monitor in synchronized methods
8440 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8441 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8443 #endif // !_TARGET_X86_
8445 Phases previousCompletedPhase; // the most recently completed phase
8447 //-------------------------------------------------------------------------
8448 // The following keeps track of how many bytes of local frame space we've
8449 // grabbed so far in the current function, and how many argument bytes we
8450 // need to pop when we return.
8453 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8455 // Count of callee-saved regs we pushed in the prolog.
8456 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8457 // In case of Amd64 this doesn't include float regs saved on stack.
8458 unsigned compCalleeRegsPushed;
8460 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8461 // Mask of callee saved float regs on stack.
8462 regMaskTP compCalleeFPRegsSavedMask;
8464 #ifdef _TARGET_AMD64_
8465 // Quirk for VS debug-launch scenario to work:
8466 // Bytes of padding between save-reg area and locals.
8467 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8468 unsigned compVSQuirkStackPaddingNeeded;
8469 bool compQuirkForPPPflag;
8472 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8474 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8475 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8476 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8478 //-------------------------------------------------------------------------
8480 static void compStartup(); // One-time initialization
8481 static void compShutdown(); // One-time finalization
8483 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8486 static void compDisplayStaticSizes(FILE* fout);
8488 //------------ Some utility functions --------------
8490 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8491 void** ppIndirection); /* OUT */
8493 // Several JIT/EE interface functions return a CorInfoType, and also return a
8494 // class handle as an out parameter if the type is a value class. Returns the
8495 // size of the type these describe.
8496 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8499 // Components used by the compiler may write unit test suites, and
8500 // have them run within this method. They will be run only once per process, and only
8501 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8502 // These should fail by asserting.
8503 void compDoComponentUnitTestsOnce();
8506 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8507 CORINFO_MODULE_HANDLE classPtr,
8508 COMP_HANDLE compHnd,
8509 CORINFO_METHOD_INFO* methodInfo,
8510 void** methodCodePtr,
8511 ULONG* methodCodeSize,
8512 JitFlags* compileFlags);
8513 void compCompileFinish();
8514 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8515 COMP_HANDLE compHnd,
8516 CORINFO_METHOD_INFO* methodInfo,
8517 void** methodCodePtr,
8518 ULONG* methodCodeSize,
8519 JitFlags* compileFlags,
8520 CorInfoInstantiationVerification instVerInfo);
8522 ArenaAllocator* compGetAllocator();
8524 #if MEASURE_MEM_ALLOC
8526 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8530 unsigned allocCnt; // # of allocs
8531 UINT64 allocSz; // total size of those alloc.
8532 UINT64 allocSzMax; // Maximum single allocation.
8533 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8534 UINT64 nraTotalSizeAlloc;
8535 UINT64 nraTotalSizeUsed;
8537 static const char* s_CompMemKindNames[]; // Names of the kinds.
8539 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8541 for (int i = 0; i < CMK_Count; i++)
8543 allocSzByKind[i] = 0;
8546 MemStats(const MemStats& ms)
8547 : allocCnt(ms.allocCnt)
8548 , allocSz(ms.allocSz)
8549 , allocSzMax(ms.allocSzMax)
8550 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8551 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8553 for (int i = 0; i < CMK_Count; i++)
8555 allocSzByKind[i] = ms.allocSzByKind[i];
8559 // Until we have ubiquitous constructors.
8562 this->MemStats::MemStats();
8565 void AddAlloc(size_t sz, CompMemKind cmk)
8569 if (sz > allocSzMax)
8573 allocSzByKind[cmk] += sz;
8576 void Print(FILE* f); // Print these stats to f.
8577 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8579 MemStats genMemStats;
8581 struct AggregateMemStats : public MemStats
8585 AggregateMemStats() : MemStats(), nMethods(0)
8589 void Add(const MemStats& ms)
8592 allocCnt += ms.allocCnt;
8593 allocSz += ms.allocSz;
8594 allocSzMax = max(allocSzMax, ms.allocSzMax);
8595 for (int i = 0; i < CMK_Count; i++)
8597 allocSzByKind[i] += ms.allocSzByKind[i];
8599 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8600 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8603 void Print(FILE* f); // Print these stats to jitstdout.
8606 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8607 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8608 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8610 #endif // MEASURE_MEM_ALLOC
8612 #if LOOP_HOIST_STATS
8613 unsigned m_loopsConsidered;
8614 bool m_curLoopHasHoistedExpression;
8615 unsigned m_loopsWithHoistedExpressions;
8616 unsigned m_totalHoistedExpressions;
8618 void AddLoopHoistStats();
8619 void PrintPerMethodLoopHoistStats();
8621 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8622 static unsigned s_loopsConsidered;
8623 static unsigned s_loopsWithHoistedExpressions;
8624 static unsigned s_totalHoistedExpressions;
8626 static void PrintAggregateLoopHoistStats(FILE* f);
8627 #endif // LOOP_HOIST_STATS
8629 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8630 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8631 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8632 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8633 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8634 void compFreeMem(void*);
8636 bool compIsForImportOnly();
8637 bool compIsForInlining();
8638 bool compDonotInline();
8641 const char* compLocalVarName(unsigned varNum, unsigned offs);
8642 VarName compVarName(regNumber reg, bool isFloatReg = false);
8643 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8644 const char* compRegPairName(regPairNo regPair);
8645 const char* compRegNameForSize(regNumber reg, size_t size);
8646 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8647 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8648 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8651 //-------------------------------------------------------------------------
8653 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8655 struct VarScopeMapInfo
8657 VarScopeListNode* head;
8658 VarScopeListNode* tail;
8659 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8661 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8668 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8669 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8671 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8672 VarNumToScopeDscMap;
8674 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8675 VarNumToScopeDscMap* compVarScopeMap;
8677 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8679 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8681 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8683 void compInitVarScopeMap();
8685 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8686 // enter scope, sorted by instr offset
8687 unsigned compNextEnterScope;
8689 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8690 // go out of scope, sorted by instr offset
8691 unsigned compNextExitScope;
8693 void compInitScopeLists();
8695 void compResetScopeLists();
8697 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8699 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8701 void compProcessScopesUntil(unsigned offset,
8703 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8704 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8707 void compDispScopeLists();
8710 bool compIsProfilerHookNeeded();
8712 //-------------------------------------------------------------------------
8713 /* Statistical Data Gathering */
8715 void compJitStats(); // call this function and enable
8716 // various ifdef's below for statistical data
8719 void compCallArgStats();
8720 static void compDispCallArgStats(FILE* fout);
8723 //-------------------------------------------------------------------------
8730 ArenaAllocator* compAllocator;
8733 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8734 // suitable for use by utilcode collection types.
8735 IAllocator* compAsIAllocator;
8737 #if MEASURE_MEM_ALLOC
8738 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8739 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8740 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8742 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8744 #endif // MEASURE_MEM_ALLOC
8746 void compFunctionTraceStart();
8747 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8750 size_t compMaxUncheckedOffsetForNullObject;
8752 void compInitOptions(JitFlags* compileFlags);
8754 void compSetProcessor();
8755 void compInitDebuggingInfo();
8756 void compSetOptimizationLevel();
8757 #ifdef _TARGET_ARMARCH_
8758 bool compRsvdRegCheck(FrameLayoutState curState);
8760 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8762 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8763 void ResetOptAnnotations();
8765 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8766 void RecomputeLoopInfo();
8768 #ifdef PROFILING_SUPPORTED
8769 // Data required for generating profiler Enter/Leave/TailCall hooks
8771 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8772 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8773 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8776 #ifdef _TARGET_AMD64_
8777 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8780 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8781 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8783 IAllocator* getAllocator()
8785 return compAsIAllocator;
8788 #if MEASURE_MEM_ALLOC
8789 IAllocator* getAllocatorBitset()
8791 return compAsIAllocatorBitset;
8793 IAllocator* getAllocatorGC()
8795 return compAsIAllocatorGC;
8797 IAllocator* getAllocatorLoopHoist()
8799 return compAsIAllocatorLoopHoist;
8801 #else // !MEASURE_MEM_ALLOC
8802 IAllocator* getAllocatorBitset()
8804 return compAsIAllocator;
8806 IAllocator* getAllocatorGC()
8808 return compAsIAllocator;
8810 IAllocator* getAllocatorLoopHoist()
8812 return compAsIAllocator;
8814 #endif // !MEASURE_MEM_ALLOC
8817 IAllocator* getAllocatorDebugOnly()
8819 #if MEASURE_MEM_ALLOC
8820 return compAsIAllocatorDebugOnly;
8821 #else // !MEASURE_MEM_ALLOC
8822 return compAsIAllocator;
8823 #endif // !MEASURE_MEM_ALLOC
8828 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8829 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8833 XX Checks for type compatibility and merges types XX
8835 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8836 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8840 // Set to TRUE if verification cannot be skipped for this method
8841 // If we detect unverifiable code, we will lazily check
8842 // canSkipMethodVerification() to see if verification is REALLY needed.
8843 BOOL tiVerificationNeeded;
8845 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8846 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8847 BOOL tiIsVerifiableCode;
8849 // Set to TRUE if runtime callout is needed for this method
8850 BOOL tiRuntimeCalloutNeeded;
8852 // Set to TRUE if security prolog/epilog callout is needed for this method
8853 // Note: This flag is different than compNeedSecurityCheck.
8854 // compNeedSecurityCheck means whether or not a security object needs
8855 // to be allocated on the stack, which is currently true for EnC as well.
8856 // tiSecurityCalloutNeeded means whether or not security callouts need
8857 // to be inserted in the jitted code.
8858 BOOL tiSecurityCalloutNeeded;
8860 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8861 // This support is necessary to suport attributes that are not described in
8862 // for example, signatures. For example, the permanent home byref (byref that
8863 // points to the gc heap), isn't a property of method signatures, therefore,
8864 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8865 // but when deciding if we need to reimport a block, we need to take these
8867 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8869 // Returns TRUE if child is equal to or a subtype of parent.
8870 // normalisedForStack indicates that both types are normalised for the stack
8871 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8873 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8874 // *pDest is modified to represent the merged type. Sets "*changed" to true
8875 // if this changes "*pDest".
8876 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8878 // Set pDest from the primitive value type.
8879 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8881 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8884 // <BUGNUM> VSW 471305
8885 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8886 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8887 // We use a "short" as we need to push/pop this scope.
8889 short compRegSetCheckLevel;
8893 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8894 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8896 XX IL verification stuff XX
8899 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8900 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8904 // The following is used to track liveness of local variables, initialization
8905 // of valueclass constructors, and type safe use of IL instructions.
8907 // dynamic state info needed for verification
8908 EntryState verCurrentState;
8910 // this ptr of object type .ctors are considered intited only after
8911 // the base class ctor is called, or an alternate ctor is called.
8912 // An uninited this ptr can be used to access fields, but cannot
8913 // be used to call a member function.
8914 BOOL verTrackObjCtorInitState;
8916 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8918 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8919 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8920 void verInitCurrentState();
8921 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8923 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8924 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8925 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8927 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8928 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8929 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8930 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8931 typeInfo verMakeTypeInfo(CorInfoType ciType,
8932 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8933 BOOL verIsSDArray(typeInfo ti);
8934 typeInfo verGetArrayElemType(typeInfo ti);
8936 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8937 BOOL verNeedsVerification();
8938 BOOL verIsByRefLike(const typeInfo& ti);
8939 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8941 // generic type variables range over types that satisfy IsBoxable
8942 BOOL verIsBoxable(const typeInfo& ti);
8944 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8945 DEBUGARG(unsigned line));
8946 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8947 DEBUGARG(unsigned line));
8948 bool verCheckTailCallConstraint(OPCODE opcode,
8949 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8950 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8951 // on a type parameter?
8952 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8953 // return false to the caller.
8954 // If false, it will throw.
8956 bool verIsBoxedValueType(typeInfo ti);
8958 void verVerifyCall(OPCODE opcode,
8959 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8960 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8962 bool readonlyCall, // is this a "readonly." call?
8963 const BYTE* delegateCreateStart,
8964 const BYTE* codeAddr,
8965 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8967 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8969 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8970 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8971 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8972 const CORINFO_FIELD_INFO& fieldInfo,
8973 const typeInfo* tiThis,
8975 BOOL allowPlainStructAsThis = FALSE);
8976 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8977 void verVerifyThisPtrInitialised();
8978 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8980 // Register allocator
8981 void raInitStackFP();
8982 void raEnregisterVarsPrePassStackFP();
8983 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8984 void raEnregisterVarsPostPassStackFP();
8985 void raGenerateFPRefCounts();
8986 void raEnregisterVarsStackFP();
8987 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8989 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8990 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8992 // returns true if enregistering v1 would save more mem accesses than v2
8993 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8996 void raDumpHeightsStackFP();
8997 void raDumpVariableRegIntfFloat();
9000 #if FEATURE_STACK_FP_X87
9002 // Currently, we use FP transition blocks in only 2 situations:
9004 // -conditional jump on longs where FP stack differs with target: it's not strictly
9005 // necessary, but its low frequency and the code would get complicated if we try to
9006 // inline the FP stack adjustment, as we have a lot of special casing going on to try
9007 // minimize the way we generate the jump code.
9008 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
9009 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
9011 // However, transition blocks have 2 problems
9013 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9014 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9015 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9016 // in the right place without preordering them), this causes us to have to generate the transition
9017 // blocks in the cold area if we want procedure splitting.
9020 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9021 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9022 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9023 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9024 // a big change in the exception.
9026 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9027 // optimizations. For these 2 cases:
9029 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9030 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9031 // a switch statement.
9033 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9034 // current procedure splitting and exception code have.
9035 bool compMayHaveTransitionBlocks;
9037 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9039 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9041 unsigned raCntStkStackFP;
9042 unsigned raCntWtdStkDblStackFP;
9043 unsigned raCntStkParamDblStackFP;
9045 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9046 // TODO: Do we want to put this in LclVarDsc?
9047 unsigned raPayloadStackFP[lclMAX_TRACKED];
9048 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9050 // Useful for debugging
9051 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9053 #endif // FEATURE_STACK_FP_X87
9056 // One line log function. Default level is 0. Increasing it gives you
9057 // more log information
9059 // levels are currently unused: #define JITDUMP(level,...) ();
9060 void JitLogEE(unsigned level, const char* fmt, ...);
9062 bool compDebugBreak;
9064 bool compJitHaltMethod();
9069 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9070 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9072 XX GS Security checks for unsafe buffers XX
9074 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9075 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9078 struct ShadowParamVarInfo
9080 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9081 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9083 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9085 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9086 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9087 // slots and update all trees to refer to shadow slots is done immediately after
9088 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9089 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9090 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9091 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9092 // creating a shadow slot even though this routine returns true.
9094 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9095 // required. There are two cases under which a reg arg could potentially be used from its
9097 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9098 // b) LSRA spills it
9100 // Possible solution to address case (a)
9101 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9102 // in this routine. Note that live out of exception handler is something we may not be
9103 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9104 // Therefore, for methods with exception handling and need GS cookie check we might have
9105 // to take conservative approach.
9107 // Possible solution to address case (b)
9108 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9109 // create a new spill temp if the method needs GS cookie check.
9110 return varDsc->lvIsParam;
9111 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9112 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9119 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9124 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9125 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9126 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9128 void gsGSChecksInitCookie(); // Grabs cookie variable
9129 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9130 bool gsFindVulnerableParams(); // Shadow param analysis code
9131 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9133 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9134 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9136 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9137 // This can be overwritten by setting complus_JITInlineSize env variable.
9139 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9142 #ifdef FEATURE_JIT_METHOD_PERF
9143 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9144 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9146 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9147 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9149 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9151 #if MEASURE_CLRAPI_CALLS
9152 // Thin wrappers that call into JitTimer (if present).
9153 inline void CLRApiCallEnter(unsigned apix);
9154 inline void CLRApiCallLeave(unsigned apix);
9157 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9158 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9163 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9164 // These variables are associated with maintaining SQM data about compile time.
9165 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9166 // in the current compilation.
9167 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9168 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9169 // the inlining phase in the current compilation.
9170 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9172 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9173 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9174 // type-loading and class initialization).
9175 void RecordStateAtEndOfInlining();
9176 // Assumes being called at the end of compilation. Update the SQM state.
9177 void RecordStateAtEndOfCompilation();
9179 #ifdef FEATURE_CLRSQM
9180 // Does anything SQM related necessary at process shutdown time.
9181 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9182 #endif // FEATURE_CLRSQM
9185 #if FUNC_INFO_LOGGING
9186 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9187 // filename to write it to.
9188 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9189 #endif // FUNC_INFO_LOGGING
9191 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9193 // Is the compilation in a full trust context?
9194 bool compIsFullTrust();
9197 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9198 #endif // MEASURE_NOWAY
9200 #ifndef FEATURE_TRACELOGGING
9201 // Should we actually fire the noway assert body and the exception handler?
9202 bool compShouldThrowOnNoway();
9203 #else // FEATURE_TRACELOGGING
9204 // Should we actually fire the noway assert body and the exception handler?
9205 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9207 // Telemetry instance to use per method compilation.
9208 JitTelemetry compJitTelemetry;
9210 // Get common parameters that have to be logged with most telemetry data.
9211 void compGetTelemetryDefaults(const char** assemblyName,
9212 const char** scopeName,
9213 const char** methodName,
9214 unsigned* methodHash);
9215 #endif // !FEATURE_TRACELOGGING
9219 NodeToTestDataMap* m_nodeTestData;
9221 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9222 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9223 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9224 // Current kept in this.
9226 NodeToTestDataMap* GetNodeTestData()
9228 Compiler* compRoot = impInlineRoot();
9229 if (compRoot->m_nodeTestData == nullptr)
9231 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9233 return compRoot->m_nodeTestData;
9236 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9238 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9239 // currently occur in the AST graph.
9240 NodeToIntMap* FindReachableNodesInNodeTestData();
9242 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9243 // test data, associate that data with "to".
9244 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9246 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9247 // have annotations, attach similar annotations to the corresponding nodes in "to".
9248 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9250 // These are the methods that test that the various conditions implied by the
9251 // test attributes are satisfied.
9252 void JitTestCheckSSA(); // SSA builder tests.
9253 void JitTestCheckVN(); // Value numbering tests.
9256 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9258 FieldSeqStore* m_fieldSeqStore;
9260 FieldSeqStore* GetFieldSeqStore()
9262 Compiler* compRoot = impInlineRoot();
9263 if (compRoot->m_fieldSeqStore == nullptr)
9265 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9266 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9267 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9269 return compRoot->m_fieldSeqStore;
9272 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9274 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9275 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9276 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9277 // attach the field sequence directly to the address node.
9278 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9280 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9282 // Don't need to worry about inlining here
9283 if (m_zeroOffsetFieldMap == nullptr)
9285 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9287 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9288 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9290 return m_zeroOffsetFieldMap;
9293 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9294 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9295 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9296 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9297 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9298 // record the the field sequence using the ZeroOffsetFieldMap described above.
9300 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9301 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9302 // CoreRT. Such case is handled same as the default case.
9303 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9305 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9307 NodeToArrayInfoMap* m_arrayInfoMap;
9309 NodeToArrayInfoMap* GetArrayInfoMap()
9311 Compiler* compRoot = impInlineRoot();
9312 if (compRoot->m_arrayInfoMap == nullptr)
9314 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9315 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9316 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9318 return compRoot->m_arrayInfoMap;
9321 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9323 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9324 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9325 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9326 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9328 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9330 // Use the same map for GCHeap and ByrefExposed when their states match.
9331 memoryKind = ByrefExposed;
9334 assert(memoryKind < MemoryKindCount);
9335 Compiler* compRoot = impInlineRoot();
9336 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9338 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9339 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9340 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9342 return compRoot->m_memorySsaMap[memoryKind];
9345 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9346 CORINFO_CLASS_HANDLE m_refAnyClass;
9347 CORINFO_FIELD_HANDLE GetRefanyDataField()
9349 if (m_refAnyClass == nullptr)
9351 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9353 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9355 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9357 if (m_refAnyClass == nullptr)
9359 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9361 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9365 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9367 #if ALLVARSET_COUNTOPS
9368 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9371 static HelperCallProperties s_helperCallProperties;
9373 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9374 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9375 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9377 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9380 unsigned __int8* offset0,
9381 unsigned __int8* offset1);
9382 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9383 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9385 void fgMorphMultiregStructArgs(GenTreeCall* call);
9386 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9388 }; // end of class Compiler
9390 // Inline methods of CompAllocator.
9391 void* CompAllocator::Alloc(size_t sz)
9393 #if MEASURE_MEM_ALLOC
9394 return m_comp->compGetMem(sz, m_cmk);
9396 return m_comp->compGetMem(sz);
9400 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9402 #if MEASURE_MEM_ALLOC
9403 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9405 return m_comp->compGetMemArray(elems, elemSize);
9409 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9410 inline LclVarDsc::LclVarDsc(Compiler* comp)
9411 : // Initialize the ArgRegs to REG_STK.
9412 // The morph will do the right thing to change
9413 // to the right register if passed in register.
9416 #if FEATURE_MULTIREG_ARGS
9417 _lvOtherArgReg(REG_STK)
9419 #endif // FEATURE_MULTIREG_ARGS
9421 lvRefBlks(BlockSetOps::UninitVal())
9423 #endif // ASSERTION_PROP
9424 lvPerSsaData(comp->getAllocator())
9428 //---------------------------------------------------------------------------------------------------------------------
9429 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9431 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9432 // shown in parentheses):
9434 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9435 // of a misnomer, as the first entry will always be the current node.
9437 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9438 // argument before visiting the node's operands.
9440 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9441 // argument after visiting the node's operands.
9443 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9444 // `DoPreOrder` must be true if this option is true.
9446 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9447 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9448 // visited before the first).
9450 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9452 // A simple pre-order visitor might look something like the following:
9454 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9459 // DoPreOrder = true
9462 // unsigned m_count;
9464 // CountingVisitor(Compiler* compiler)
9465 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9469 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9475 // This visitor would then be used like so:
9477 // CountingVisitor countingVisitor(compiler);
9478 // countingVisitor.WalkTree(root);
9480 template <typename TVisitor>
9481 class GenTreeVisitor
9484 typedef Compiler::fgWalkResult fgWalkResult;
9488 ComputeStack = false,
9490 DoPostOrder = false,
9491 DoLclVarsOnly = false,
9492 UseExecutionOrder = false,
9495 Compiler* m_compiler;
9496 ArrayStack<GenTree*> m_ancestors;
9498 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9500 assert(compiler != nullptr);
9502 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9503 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9506 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9508 return fgWalkResult::WALK_CONTINUE;
9511 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9513 return fgWalkResult::WALK_CONTINUE;
9517 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9519 assert(use != nullptr);
9521 GenTree* node = *use;
9523 if (TVisitor::ComputeStack)
9525 m_ancestors.Push(node);
9528 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9529 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9531 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9532 if (result == fgWalkResult::WALK_ABORT)
9538 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9544 switch (node->OperGet())
9549 case GT_LCL_VAR_ADDR:
9550 case GT_LCL_FLD_ADDR:
9551 if (TVisitor::DoLclVarsOnly)
9553 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9554 if (result == fgWalkResult::WALK_ABORT)
9570 case GT_MEMORYBARRIER:
9575 case GT_START_NONGC:
9577 #if !FEATURE_EH_FUNCLETS
9579 #endif // !FEATURE_EH_FUNCLETS
9581 #ifndef LEGACY_BACKEND
9583 #endif // LEGACY_BACKEND
9586 case GT_CLS_VAR_ADDR:
9590 case GT_PINVOKE_PROLOG:
9591 case GT_PINVOKE_EPILOG:
9595 // Lclvar unary operators
9596 case GT_STORE_LCL_VAR:
9597 case GT_STORE_LCL_FLD:
9598 if (TVisitor::DoLclVarsOnly)
9600 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9601 if (result == fgWalkResult::WALK_ABORT)
9608 // Standard unary operators
9636 GenTreeUnOp* const unOp = node->AsUnOp();
9637 if (unOp->gtOp1 != nullptr)
9639 result = WalkTree(&unOp->gtOp1, unOp);
9640 if (result == fgWalkResult::WALK_ABORT)
9651 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9653 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9654 if (result == fgWalkResult::WALK_ABORT)
9658 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9659 if (result == fgWalkResult::WALK_ABORT)
9663 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9664 if (result == fgWalkResult::WALK_ABORT)
9671 case GT_ARR_BOUNDS_CHECK:
9674 #endif // FEATURE_SIMD
9676 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9678 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9679 if (result == fgWalkResult::WALK_ABORT)
9683 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9684 if (result == fgWalkResult::WALK_ABORT)
9693 GenTreeField* const field = node->AsField();
9695 if (field->gtFldObj != nullptr)
9697 result = WalkTree(&field->gtFldObj, field);
9698 if (result == fgWalkResult::WALK_ABORT)
9708 GenTreeArrElem* const arrElem = node->AsArrElem();
9710 result = WalkTree(&arrElem->gtArrObj, arrElem);
9711 if (result == fgWalkResult::WALK_ABORT)
9716 const unsigned rank = arrElem->gtArrRank;
9717 for (unsigned dim = 0; dim < rank; dim++)
9719 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9720 if (result == fgWalkResult::WALK_ABORT)
9730 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9732 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9733 if (result == fgWalkResult::WALK_ABORT)
9737 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9738 if (result == fgWalkResult::WALK_ABORT)
9742 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9743 if (result == fgWalkResult::WALK_ABORT)
9752 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9754 GenTree** op1Use = &dynBlock->gtOp1;
9755 GenTree** op2Use = &dynBlock->gtDynamicSize;
9757 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9759 std::swap(op1Use, op2Use);
9762 result = WalkTree(op1Use, dynBlock);
9763 if (result == fgWalkResult::WALK_ABORT)
9767 result = WalkTree(op2Use, dynBlock);
9768 if (result == fgWalkResult::WALK_ABORT)
9775 case GT_STORE_DYN_BLK:
9777 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9779 GenTree** op1Use = &dynBlock->gtOp1;
9780 GenTree** op2Use = &dynBlock->gtOp2;
9781 GenTree** op3Use = &dynBlock->gtDynamicSize;
9783 if (TVisitor::UseExecutionOrder)
9785 if (dynBlock->IsReverseOp())
9787 std::swap(op1Use, op2Use);
9789 if (dynBlock->gtEvalSizeFirst)
9791 std::swap(op3Use, op2Use);
9792 std::swap(op2Use, op1Use);
9796 result = WalkTree(op1Use, dynBlock);
9797 if (result == fgWalkResult::WALK_ABORT)
9801 result = WalkTree(op2Use, dynBlock);
9802 if (result == fgWalkResult::WALK_ABORT)
9806 result = WalkTree(op3Use, dynBlock);
9807 if (result == fgWalkResult::WALK_ABORT)
9816 GenTreeCall* const call = node->AsCall();
9818 if (call->gtCallObjp != nullptr)
9820 result = WalkTree(&call->gtCallObjp, call);
9821 if (result == fgWalkResult::WALK_ABORT)
9827 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9829 result = WalkTree(args->pCurrent(), call);
9830 if (result == fgWalkResult::WALK_ABORT)
9836 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9838 result = WalkTree(args->pCurrent(), call);
9839 if (result == fgWalkResult::WALK_ABORT)
9845 if (call->gtCallType == CT_INDIRECT)
9847 if (call->gtCallCookie != nullptr)
9849 result = WalkTree(&call->gtCallCookie, call);
9850 if (result == fgWalkResult::WALK_ABORT)
9856 result = WalkTree(&call->gtCallAddr, call);
9857 if (result == fgWalkResult::WALK_ABORT)
9863 if (call->gtControlExpr != nullptr)
9865 result = WalkTree(&call->gtControlExpr, call);
9866 if (result == fgWalkResult::WALK_ABORT)
9878 assert(node->OperIsBinary());
9880 GenTreeOp* const op = node->AsOp();
9882 GenTree** op1Use = &op->gtOp1;
9883 GenTree** op2Use = &op->gtOp2;
9885 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9887 std::swap(op1Use, op2Use);
9890 if (*op1Use != nullptr)
9892 result = WalkTree(op1Use, op);
9893 if (result == fgWalkResult::WALK_ABORT)
9899 if (*op2Use != nullptr)
9901 result = WalkTree(op2Use, op);
9902 if (result == fgWalkResult::WALK_ABORT)
9912 // Finally, visit the current node
9913 if (TVisitor::DoPostOrder)
9915 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9918 if (TVisitor::ComputeStack)
9927 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9928 class GenericTreeWalker final
9929 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
9934 ComputeStack = computeStack,
9935 DoPreOrder = doPreOrder,
9936 DoPostOrder = doPostOrder,
9937 DoLclVarsOnly = doLclVarsOnly,
9938 UseExecutionOrder = useExecutionOrder,
9942 Compiler::fgWalkData* m_walkData;
9945 GenericTreeWalker(Compiler::fgWalkData* walkData)
9946 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
9948 , m_walkData(walkData)
9950 assert(walkData != nullptr);
9954 walkData->parentStack = &this->m_ancestors;
9958 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9960 m_walkData->parent = user;
9961 return m_walkData->wtprVisitorFn(use, m_walkData);
9964 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9966 m_walkData->parent = user;
9967 return m_walkData->wtpoVisitorFn(use, m_walkData);
9971 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
9977 DoLclVarsOnly = true
9980 IncLclVarRefCountsVisitor(Compiler* compiler);
9981 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9983 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9986 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
9992 DoLclVarsOnly = true
9995 DecLclVarRefCountsVisitor(Compiler* compiler);
9996 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9998 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10002 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10003 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10005 XX Miscellaneous Compiler stuff XX
10007 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10008 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10011 // Values used to mark the types a stack slot is used for
10013 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10014 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10015 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10016 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10017 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10018 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10019 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10020 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10022 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10024 /*****************************************************************************
10026 * Variables to keep track of total code amounts.
10031 extern size_t grossVMsize;
10032 extern size_t grossNCsize;
10033 extern size_t totalNCsize;
10035 extern unsigned genMethodICnt;
10036 extern unsigned genMethodNCnt;
10037 extern size_t gcHeaderISize;
10038 extern size_t gcPtrMapISize;
10039 extern size_t gcHeaderNSize;
10040 extern size_t gcPtrMapNSize;
10042 #endif // DISPLAY_SIZES
10044 /*****************************************************************************
10046 * Variables to keep track of basic block counts (more data on 1 BB methods)
10049 #if COUNT_BASIC_BLOCKS
10050 extern Histogram bbCntTable;
10051 extern Histogram bbOneBBSizeTable;
10054 /*****************************************************************************
10056 * Used by optFindNaturalLoops to gather statistical information such as
10057 * - total number of natural loops
10058 * - number of loops with 1, 2, ... exit conditions
10059 * - number of loops that have an iterator (for like)
10060 * - number of loops that have a constant iterator
10065 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10066 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10067 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10068 extern unsigned totalLoopCount; // counts the total number of natural loops
10069 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10070 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10071 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10072 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10074 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10075 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10076 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10077 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10078 extern Histogram loopCountTable; // Histogram of loop counts
10079 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10081 #endif // COUNT_LOOPS
10083 /*****************************************************************************
10084 * variables to keep track of how many iterations we go in a dataflow pass
10089 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10090 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10092 #endif // DATAFLOW_ITER
10094 #if MEASURE_BLOCK_SIZE
10095 extern size_t genFlowNodeSize;
10096 extern size_t genFlowNodeCnt;
10097 #endif // MEASURE_BLOCK_SIZE
10099 #if MEASURE_NODE_SIZE
10100 struct NodeSizeStats
10104 genTreeNodeCnt = 0;
10105 genTreeNodeSize = 0;
10106 genTreeNodeActualSize = 0;
10109 // Count of tree nodes allocated.
10110 unsigned __int64 genTreeNodeCnt;
10112 // The size we allocate.
10113 unsigned __int64 genTreeNodeSize;
10115 // The actual size of the node. Note that the actual size will likely be smaller
10116 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10117 // a smaller node to a larger one. TODO-Cleanup: add stats on
10118 // SetOper()/ChangeOper() usage to quantify this.
10119 unsigned __int64 genTreeNodeActualSize;
10121 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10122 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10123 extern Histogram genTreeNcntHist;
10124 extern Histogram genTreeNsizHist;
10125 #endif // MEASURE_NODE_SIZE
10127 /*****************************************************************************
10128 * Count fatal errors (including noway_asserts).
10132 extern unsigned fatal_badCode;
10133 extern unsigned fatal_noWay;
10134 extern unsigned fatal_NOMEM;
10135 extern unsigned fatal_noWayAssertBody;
10137 extern unsigned fatal_noWayAssertBodyArgs;
10139 extern unsigned fatal_NYI;
10140 #endif // MEASURE_FATAL
10142 /*****************************************************************************
10146 #ifdef _TARGET_XARCH_
10148 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10149 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10150 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10152 const instruction INS_AND = INS_and;
10153 const instruction INS_OR = INS_or;
10154 const instruction INS_XOR = INS_xor;
10155 const instruction INS_NEG = INS_neg;
10156 const instruction INS_TEST = INS_test;
10157 const instruction INS_MUL = INS_imul;
10158 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10159 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10160 const instruction INS_BREAKPOINT = INS_int3;
10161 const instruction INS_ADDC = INS_adc;
10162 const instruction INS_SUBC = INS_sbb;
10163 const instruction INS_NOT = INS_not;
10167 #ifdef _TARGET_ARM_
10169 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10170 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10171 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10173 const instruction INS_AND = INS_and;
10174 const instruction INS_OR = INS_orr;
10175 const instruction INS_XOR = INS_eor;
10176 const instruction INS_NEG = INS_rsb;
10177 const instruction INS_TEST = INS_tst;
10178 const instruction INS_MUL = INS_mul;
10179 const instruction INS_MULADD = INS_mla;
10180 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10181 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10182 const instruction INS_BREAKPOINT = INS_bkpt;
10183 const instruction INS_ADDC = INS_adc;
10184 const instruction INS_SUBC = INS_sbc;
10185 const instruction INS_NOT = INS_mvn;
10187 const instruction INS_ABS = INS_vabs;
10188 const instruction INS_ROUND = INS_invalid;
10189 const instruction INS_SQRT = INS_vsqrt;
10193 #ifdef _TARGET_ARM64_
10195 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10196 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10197 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10199 const instruction INS_AND = INS_and;
10200 const instruction INS_OR = INS_orr;
10201 const instruction INS_XOR = INS_eor;
10202 const instruction INS_NEG = INS_neg;
10203 const instruction INS_TEST = INS_tst;
10204 const instruction INS_MUL = INS_mul;
10205 const instruction INS_MULADD = INS_madd;
10206 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10207 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10208 const instruction INS_BREAKPOINT = INS_bkpt;
10209 const instruction INS_ADDC = INS_adc;
10210 const instruction INS_SUBC = INS_sbc;
10211 const instruction INS_NOT = INS_mvn;
10213 const instruction INS_ABS = INS_fabs;
10214 const instruction INS_ROUND = INS_frintn;
10215 const instruction INS_SQRT = INS_fsqrt;
10219 /*****************************************************************************/
10221 extern const BYTE genTypeSizes[];
10222 extern const BYTE genTypeAlignments[];
10223 extern const BYTE genTypeStSzs[];
10224 extern const BYTE genActualTypes[];
10226 /*****************************************************************************/
10228 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10229 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10231 #ifdef _TARGET_ARM_
10232 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10233 #elif defined(_TARGET_ARM64_)
10234 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10237 /*****************************************************************************/
10239 #define REG_CORRUPT regNumber(REG_NA + 1)
10240 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10241 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10243 /*****************************************************************************/
10245 extern BasicBlock dummyBB;
10247 /*****************************************************************************/
10248 /*****************************************************************************/
10250 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10251 // nodes of a statement in execution order.
10252 // __stmt: a GT_STMT type GenTree*
10253 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10255 #define foreach_treenode_execution_order(__node, __stmt) \
10256 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10258 // foreach_block: An iterator over all blocks in the function.
10259 // __compiler: the Compiler* object
10260 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10262 #define foreach_block(__compiler, __block) \
10263 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10265 /*****************************************************************************/
10266 /*****************************************************************************/
10270 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10272 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10273 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10275 XX Debugging helpers XX
10277 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10278 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10281 /*****************************************************************************/
10282 /* The following functions are intended to be called from the debugger, to dump
10283 * various data structures. The can be used in the debugger Watch or Quick Watch
10284 * windows. They are designed to be short to type and take as few arguments as
10285 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10286 * See the function definition comment for more details.
10289 void cBlock(Compiler* comp, BasicBlock* block);
10290 void cBlocks(Compiler* comp);
10291 void cBlocksV(Compiler* comp);
10292 void cTree(Compiler* comp, GenTree* tree);
10293 void cTrees(Compiler* comp);
10294 void cEH(Compiler* comp);
10295 void cVar(Compiler* comp, unsigned lclNum);
10296 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10297 void cVars(Compiler* comp);
10298 void cVarsFinal(Compiler* comp);
10299 void cBlockPreds(Compiler* comp, BasicBlock* block);
10300 void cReach(Compiler* comp);
10301 void cDoms(Compiler* comp);
10302 void cLiveness(Compiler* comp);
10303 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10305 void cFuncIR(Compiler* comp);
10306 void cBlockIR(Compiler* comp, BasicBlock* block);
10307 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10308 void cTreeIR(Compiler* comp, GenTree* tree);
10309 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10310 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10311 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10312 int cOperandIR(Compiler* comp, GenTree* operand);
10313 int cLeafIR(Compiler* comp, GenTree* tree);
10314 int cIndirIR(Compiler* comp, GenTree* tree);
10315 int cListIR(Compiler* comp, GenTree* list);
10316 int cSsaNumIR(Compiler* comp, GenTree* tree);
10317 int cValNumIR(Compiler* comp, GenTree* tree);
10318 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10320 void dBlock(BasicBlock* block);
10323 void dTree(GenTree* tree);
10326 void dVar(unsigned lclNum);
10327 void dVarDsc(LclVarDsc* varDsc);
10330 void dBlockPreds(BasicBlock* block);
10334 void dCVarSet(VARSET_VALARG_TP vars);
10336 void dVarSet(VARSET_VALARG_TP vars);
10337 void dRegMask(regMaskTP mask);
10340 void dBlockIR(BasicBlock* block);
10341 void dTreeIR(GenTree* tree);
10342 void dLoopIR(Compiler::LoopDsc* loop);
10343 void dLoopNumIR(unsigned loopNum);
10344 int dTabStopIR(int curr, int tabstop);
10345 int dTreeTypeIR(GenTree* tree);
10346 int dTreeKindsIR(GenTree* tree);
10347 int dTreeFlagsIR(GenTree* tree);
10348 int dOperandIR(GenTree* operand);
10349 int dLeafIR(GenTree* tree);
10350 int dIndirIR(GenTree* tree);
10351 int dListIR(GenTree* list);
10352 int dSsaNumIR(GenTree* tree);
10353 int dValNumIR(GenTree* tree);
10354 int dDependsIR(GenTree* comma);
10357 GenTree* dFindTree(GenTree* tree, unsigned id);
10358 GenTree* dFindTree(unsigned id);
10359 GenTreeStmt* dFindStmt(unsigned id);
10360 BasicBlock* dFindBlock(unsigned bbNum);
10364 #include "compiler.hpp" // All the shared inline functions
10366 /*****************************************************************************/
10367 #endif //_COMPILER_H_
10368 /*****************************************************************************/