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);
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);
2232 //-------------------------------------------------------------------------
2233 // Get the handle, if any.
2234 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2235 // Get the handle, and assert if not found.
2236 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2237 // Get the handle for a ref type.
2238 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2240 //-------------------------------------------------------------------------
2241 // Functions to display the trees
2244 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2246 void gtDispVN(GenTreePtr tree);
2247 void gtDispConst(GenTreePtr tree);
2248 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2249 void gtDispNodeName(GenTreePtr tree);
2250 void gtDispRegVal(GenTreePtr tree);
2262 void gtDispChild(GenTreePtr child,
2263 IndentStack* indentStack,
2265 __in_opt const char* msg = nullptr,
2266 bool topOnly = false);
2267 void gtDispTree(GenTreePtr tree,
2268 IndentStack* indentStack = nullptr,
2269 __in_opt const char* msg = nullptr,
2270 bool topOnly = false,
2271 bool isLIR = false);
2272 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2273 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2274 char* gtGetLclVarName(unsigned lclNum);
2275 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2276 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2277 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2278 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2279 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2280 void gtDispFieldSeq(FieldSeqNode* pfsn);
2282 void gtDispRange(LIR::ReadOnlyRange const& range);
2284 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2286 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2298 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2299 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2302 static fgWalkPreFn gtAssertColonCond;
2304 static fgWalkPreFn gtMarkColonCond;
2305 static fgWalkPreFn gtClearColonCond;
2307 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2308 bool gtHasCatchArg(GenTreePtr tree);
2309 bool gtHasUnmanagedCall(GenTreePtr tree);
2311 typedef ArrayStack<GenTree*> GenTreeStack;
2313 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2314 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2316 //=========================================================================
2317 // BasicBlock functions
2319 // This is a debug flag we will use to assert when creating block during codegen
2320 // as this interferes with procedure splitting. If you know what you're doing, set
2321 // it to true before creating the block. (DEBUG only)
2322 bool fgSafeBasicBlockCreation;
2325 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2328 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2329 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2333 XX The variables to be used by the code generator. XX
2335 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2340 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2341 // be placed in the stack frame and it's fields must be laid out sequentially.
2343 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2344 // a local variable that can be enregistered or placed in the stack frame.
2345 // The fields do not need to be laid out sequentially
2347 enum lvaPromotionType
2349 PROMOTION_TYPE_NONE, // The struct local is not promoted
2350 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2351 // and its field locals are independent of its parent struct local.
2352 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2353 // but its field locals depend on its parent struct local.
2356 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2357 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2359 /*****************************************************************************/
2361 enum FrameLayoutState
2364 INITIAL_FRAME_LAYOUT,
2365 PRE_REGALLOC_FRAME_LAYOUT,
2366 REGALLOC_FRAME_LAYOUT,
2367 TENTATIVE_FRAME_LAYOUT,
2372 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2373 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2374 bool lvaSortAgain; // true: We need to sort the lvaTable
2375 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2376 unsigned lvaCount; // total number of locals
2378 unsigned lvaRefCount; // total number of references to locals
2379 LclVarDsc* lvaTable; // variable descriptor table
2380 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2382 LclVarDsc** lvaRefSorted; // table sorted by refcount
2384 unsigned short lvaTrackedCount; // actual # of locals being tracked
2385 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2387 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2388 // Only for AMD64 System V cache the first caller stack homed argument.
2389 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2390 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2393 VARSET_TP lvaTrackedVars; // set of tracked variables
2395 #ifndef _TARGET_64BIT_
2396 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2398 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2400 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2401 // It that changes, this changes. VarSets from different epochs
2402 // cannot be meaningfully combined.
2404 unsigned GetCurLVEpoch()
2409 // reverse map of tracked number to var number
2410 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2412 #ifdef LEGACY_BACKEND
2413 // variable interference graph
2414 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2417 // variable preference graph
2418 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2422 // # of procs compiled a with double-aligned stack
2423 static unsigned s_lvaDoubleAlignedProcsCount;
2427 // Getters and setters for address-exposed and do-not-enregister local var properties.
2428 bool lvaVarAddrExposed(unsigned varNum);
2429 void lvaSetVarAddrExposed(unsigned varNum);
2430 bool lvaVarDoNotEnregister(unsigned varNum);
2432 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2433 enum DoNotEnregisterReason
2438 DNER_VMNeedsStackAddr,
2439 DNER_LiveInOutOfHandler,
2440 DNER_LiveAcrossUnmanagedCall,
2441 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2442 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2443 DNER_DepField, // It is a field of a dependently promoted struct
2444 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2445 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2446 #ifdef JIT32_GCENCODER
2451 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2453 unsigned lvaVarargsHandleArg;
2455 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2457 #endif // _TARGET_X86_
2459 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2460 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2461 #if FEATURE_FIXED_OUT_ARGS
2462 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2464 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2465 // that tracks whether the lock has been taken
2467 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2468 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2469 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2471 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2472 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2474 #if FEATURE_FIXED_OUT_ARGS
2475 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2476 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2477 #endif // FEATURE_FIXED_OUT_ARGS
2480 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2481 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2482 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2483 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2484 // this variable to be this scratch word whenever struct promotion occurs.
2485 unsigned lvaPromotedStructAssemblyScratchVar;
2486 #endif // _TARGET_ARM_
2489 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2490 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2493 unsigned lvaGenericsContextUseCount;
2495 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2496 // CORINFO_GENERICS_CTXT_FROM_THIS?
2497 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2499 //-------------------------------------------------------------------------
2500 // All these frame offsets are inter-related and must be kept in sync
2502 #if !FEATURE_EH_FUNCLETS
2503 // This is used for the callable handlers
2504 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2505 #endif // FEATURE_EH_FUNCLETS
2507 unsigned lvaCachedGenericContextArgOffs;
2508 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2511 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2513 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2515 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2516 // after the reg predict we will use a computed maxTmpSize
2517 // which is based upon the number of spill temps predicted by reg predict
2518 // All this is necessary because if we under-estimate the size of the spill
2519 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2521 // Pre codegen max spill temp size.
2522 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2524 //-------------------------------------------------------------------------
2526 unsigned lvaGetMaxSpillTempSize();
2528 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2529 #endif // _TARGET_ARM_
2530 void lvaAssignFrameOffsets(FrameLayoutState curState);
2531 void lvaFixVirtualFrameOffsets();
2533 #ifndef LEGACY_BACKEND
2534 void lvaUpdateArgsWithInitialReg();
2535 #endif // !LEGACY_BACKEND
2537 void lvaAssignVirtualFrameOffsetsToArgs();
2538 #ifdef UNIX_AMD64_ABI
2539 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2540 #else // !UNIX_AMD64_ABI
2541 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2542 #endif // !UNIX_AMD64_ABI
2543 void lvaAssignVirtualFrameOffsetsToLocals();
2544 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2545 #ifdef _TARGET_AMD64_
2546 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2547 bool lvaIsCalleeSavedIntRegCountEven();
2549 void lvaAlignFrame();
2550 void lvaAssignFrameOffsetsToPromotedStructs();
2551 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2554 void lvaDumpRegLocation(unsigned lclNum);
2555 void lvaDumpFrameLocation(unsigned lclNum);
2556 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2557 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2558 // layout state defined by lvaDoneFrameLayout
2561 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2562 // to avoid bugs from borderline cases.
2563 #define MAX_FrameSize 0x3FFFFFFF
2564 void lvaIncrementFrameSize(unsigned size);
2566 unsigned lvaFrameSize(FrameLayoutState curState);
2568 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2569 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2571 // Returns the caller-SP-relative offset for the local variable "varNum."
2572 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2574 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2575 int lvaGetSPRelativeOffset(unsigned varNum);
2577 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2578 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2580 //------------------------ For splitting types ----------------------------
2582 void lvaInitTypeRef();
2584 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2585 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2586 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2587 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2588 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2589 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2591 void lvaInitVarDsc(LclVarDsc* varDsc,
2593 CorInfoType corInfoType,
2594 CORINFO_CLASS_HANDLE typeHnd,
2595 CORINFO_ARG_LIST_HANDLE varList,
2596 CORINFO_SIG_INFO* varSig);
2598 static unsigned lvaTypeRefMask(var_types type);
2600 var_types lvaGetActualType(unsigned lclNum);
2601 var_types lvaGetRealType(unsigned lclNum);
2603 //-------------------------------------------------------------------------
2607 unsigned lvaLclSize(unsigned varNum);
2608 unsigned lvaLclExactSize(unsigned varNum);
2610 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2612 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2613 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2614 // the return result.
2615 bool lvaLclVarRefsAccum(
2616 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2618 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2619 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2620 // and (destructively) unions "trkedVars" into "*result".
2621 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2623 ALLVARSET_VALARG_TP allVars,
2624 VARSET_VALARG_TP trkdVars);
2626 bool lvaHaveManyLocals() const;
2628 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2629 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2630 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2633 void lvaSortByRefCount();
2634 void lvaDumpRefCounts();
2636 void lvaMarkLocalVars(BasicBlock* block);
2638 void lvaMarkLocalVars(); // Local variable ref-counting
2640 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2642 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2644 void lvaIncRefCnts(GenTreePtr tree);
2645 void lvaDecRefCnts(GenTreePtr tree);
2647 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2648 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2649 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2652 struct lvaStressLclFldArgs
2654 Compiler* m_pCompiler;
2658 static fgWalkPreFn lvaStressLclFldCB;
2659 void lvaStressLclFld();
2661 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2662 void lvaDispVarSet(VARSET_VALARG_TP set);
2667 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2669 int lvaFrameAddress(int varNum, bool* pFPbased);
2672 bool lvaIsParameter(unsigned varNum);
2673 bool lvaIsRegArgument(unsigned varNum);
2674 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2675 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2676 // that writes to arg0
2678 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2679 // (this is an overload of lvIsTemp because there are no temp parameters).
2680 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2681 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2682 bool lvaIsImplicitByRefLocal(unsigned varNum)
2684 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2685 LclVarDsc* varDsc = &(lvaTable[varNum]);
2686 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2688 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2691 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2695 // Returns true if this local var is a multireg struct
2696 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2698 // If the local is a TYP_STRUCT, get/set a class handle describing it
2699 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2700 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2702 // If the local is TYP_REF, set or update the associated class information.
2703 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2704 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2705 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2706 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2708 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2710 // Info about struct fields
2711 struct lvaStructFieldInfo
2713 CORINFO_FIELD_HANDLE fldHnd;
2714 unsigned char fldOffset;
2715 unsigned char fldOrdinal;
2718 CORINFO_CLASS_HANDLE fldTypeHnd;
2721 // Info about struct to be promoted.
2722 struct lvaStructPromotionInfo
2724 CORINFO_CLASS_HANDLE typeHnd;
2726 bool requiresScratchVar;
2729 unsigned char fieldCnt;
2730 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2732 lvaStructPromotionInfo()
2733 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2738 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2739 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2740 lvaStructPromotionInfo* StructPromotionInfo,
2742 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2743 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2744 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2745 #if !defined(_TARGET_64BIT_)
2746 void lvaPromoteLongVars();
2747 #endif // !defined(_TARGET_64BIT_)
2748 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2749 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2750 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2751 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2752 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2753 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2754 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2756 #if defined(FEATURE_SIMD)
2757 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2759 assert(varDsc->lvType == TYP_SIMD12);
2760 assert(varDsc->lvExactSize == 12);
2762 #if defined(_TARGET_64BIT_)
2763 assert(varDsc->lvSize() == 16);
2765 #else // !defined(_TARGET_64BIT_)
2767 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2768 // already does this calculation. However, we also need to prevent mapping types if the var is a
2769 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2770 // However, we don't know this until late, so we may have already pretended the field is bigger
2772 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2781 #endif // !defined(_TARGET_64BIT_)
2783 #endif // defined(FEATURE_SIMD)
2785 BYTE* lvaGetGcLayout(unsigned varNum);
2786 bool lvaTypeIsGC(unsigned varNum);
2787 unsigned lvaGSSecurityCookie; // LclVar number
2788 bool lvaTempsHaveLargerOffsetThanVars();
2790 unsigned lvaSecurityObject; // variable representing the security object on the stack
2791 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2793 #if FEATURE_EH_FUNCLETS
2794 unsigned lvaPSPSym; // variable representing the PSPSym
2797 InlineInfo* impInlineInfo;
2798 InlineStrategy* m_inlineStrategy;
2800 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2801 Compiler* impInlineRoot();
2803 #if defined(DEBUG) || defined(INLINE_DATA)
2804 unsigned __int64 getInlineCycleCount()
2806 return m_compCycles;
2808 #endif // defined(DEBUG) || defined(INLINE_DATA)
2810 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2811 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2813 //=========================================================================
2815 //=========================================================================
2818 //---------------- Local variable ref-counting ----------------------------
2821 BasicBlock* lvaMarkRefsCurBlock;
2822 GenTreePtr lvaMarkRefsCurStmt;
2824 BasicBlock::weight_t lvaMarkRefsWeight;
2826 void lvaMarkLclRefs(GenTreePtr tree);
2828 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2829 void SetVolatileHint(LclVarDsc* varDsc);
2831 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2832 PerSsaArray lvMemoryPerSsaData;
2833 unsigned lvMemoryNumSsaNames;
2836 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2837 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2838 // not an SSA variable).
2839 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2841 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2842 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2844 assert(ssaNum < lvMemoryNumSsaNames);
2845 return &lvMemoryPerSsaData.GetRef(ssaNum);
2849 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2850 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2854 XX Imports the given method and converts it to semantic trees XX
2856 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2857 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2863 void impImport(BasicBlock* method);
2865 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2866 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2867 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2868 CORINFO_CLASS_HANDLE impGetStringClass();
2869 CORINFO_CLASS_HANDLE impGetObjectClass();
2871 //=========================================================================
2873 //=========================================================================
2876 //-------------------- Stack manipulation ---------------------------------
2878 unsigned impStkSize; // Size of the full stack
2880 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2882 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2884 struct SavedStack // used to save/restore stack contents.
2886 unsigned ssDepth; // number of values on stack
2887 StackEntry* ssTrees; // saved tree values
2890 bool impIsPrimitive(CorInfoType type);
2891 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2893 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2895 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2896 void impPushNullObjRefOnStack();
2897 StackEntry impPopStack();
2898 StackEntry& impStackTop(unsigned n = 0);
2899 unsigned impStackHeight();
2901 void impSaveStackState(SavedStack* savePtr, bool copy);
2902 void impRestoreStackState(SavedStack* savePtr);
2904 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2905 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2906 CORINFO_CALL_INFO* pCallInfo);
2908 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2910 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2912 bool impCanPInvokeInline();
2913 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2914 void impCheckForPInvokeCall(
2915 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2916 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2917 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2919 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2920 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2921 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2923 var_types impImportCall(OPCODE opcode,
2924 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2925 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2927 GenTreePtr newobjThis,
2929 CORINFO_CALL_INFO* callInfo,
2930 IL_OFFSET rawILOffset);
2932 void impDevirtualizeCall(GenTreeCall* call,
2934 CORINFO_METHOD_HANDLE* method,
2935 unsigned* methodFlags,
2936 CORINFO_CONTEXT_HANDLE* contextHandle,
2937 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2939 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2941 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2943 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2946 var_types impImportJitTestLabelMark(int numArgs);
2949 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2951 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2953 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2954 CORINFO_ACCESS_FLAGS access,
2955 CORINFO_FIELD_INFO* pFieldInfo,
2958 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2960 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2962 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2964 void impImportLeave(BasicBlock* block);
2965 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2966 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2967 CORINFO_CLASS_HANDLE clsHnd,
2968 CORINFO_METHOD_HANDLE method,
2969 CORINFO_SIG_INFO* sig,
2973 CorInfoIntrinsics* pIntrinsicID);
2974 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2975 CORINFO_SIG_INFO* sig,
2978 CorInfoIntrinsics intrinsicID);
2979 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2981 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2983 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2984 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2985 CORINFO_THIS_TRANSFORM transform);
2987 //----------------- Manipulating the trees and stmts ----------------------
2989 GenTreePtr impTreeList; // Trees for the BB being imported
2990 GenTreePtr impTreeLast; // The last tree for the current BB
2994 CHECK_SPILL_ALL = -1,
2995 CHECK_SPILL_NONE = -2
2999 void impBeginTreeList();
3000 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
3001 void impEndTreeList(BasicBlock* block);
3002 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3003 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3004 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3005 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3006 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3007 void impAssignTempGen(unsigned tmp,
3010 GenTreePtr* pAfterStmt = nullptr,
3011 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3012 BasicBlock* block = nullptr);
3013 void impAssignTempGen(unsigned tmpNum,
3015 CORINFO_CLASS_HANDLE structHnd,
3017 GenTreePtr* pAfterStmt = nullptr,
3018 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3019 BasicBlock* block = nullptr);
3020 GenTreePtr impCloneExpr(GenTreePtr tree,
3022 CORINFO_CLASS_HANDLE structHnd,
3024 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3025 GenTreePtr impAssignStruct(GenTreePtr dest,
3027 CORINFO_CLASS_HANDLE structHnd,
3029 GenTreePtr* pAfterStmt = nullptr,
3030 BasicBlock* block = nullptr);
3031 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3033 CORINFO_CLASS_HANDLE structHnd,
3035 GenTreePtr* pAfterStmt = nullptr,
3036 BasicBlock* block = nullptr);
3038 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3039 CORINFO_CLASS_HANDLE structHnd,
3043 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3044 BYTE* gcLayout = nullptr,
3045 unsigned* numGCVars = nullptr,
3046 var_types* simdBaseType = nullptr);
3048 GenTreePtr impNormStructVal(GenTreePtr structVal,
3049 CORINFO_CLASS_HANDLE structHnd,
3051 bool forceNormalization = false);
3053 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3054 BOOL* pRuntimeLookup = nullptr,
3055 BOOL mustRestoreHandle = FALSE,
3056 BOOL importParent = FALSE);
3058 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3059 BOOL* pRuntimeLookup = nullptr,
3060 BOOL mustRestoreHandle = FALSE)
3062 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3065 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3066 CORINFO_LOOKUP* pLookup,
3068 void* compileTimeHandle);
3070 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3072 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3073 CORINFO_LOOKUP* pLookup,
3074 void* compileTimeHandle);
3076 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3078 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3079 CorInfoHelpFunc helper,
3081 GenTreeArgList* arg = nullptr,
3082 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3084 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3086 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3089 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3090 CORINFO_CLASS_HANDLE typeClass,
3094 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3095 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3096 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3097 static bool IsMathIntrinsic(GenTreePtr tree);
3100 //----------------- Importing the method ----------------------------------
3102 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3105 unsigned impCurOpcOffs;
3106 const char* impCurOpcName;
3107 bool impNestedStackSpill;
3109 // For displaying instrs with generated native code (-n:B)
3110 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3111 void impNoteLastILoffs();
3114 /* IL offset of the stmt currently being imported. It gets set to
3115 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3116 updated at IL offsets for which we have to report mapping info.
3117 It also includes flag bits, so use jitGetILoffs()
3118 to get the actual IL offset value.
3121 IL_OFFSETX impCurStmtOffs;
3122 void impCurStmtOffsSet(IL_OFFSET offs);
3124 void impNoteBranchOffs();
3126 unsigned impInitBlockLineInfo();
3128 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3129 bool impIsThis(GenTreePtr obj);
3130 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3131 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3132 bool impIsAnySTLOC(OPCODE opcode)
3134 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3135 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3138 GenTreeArgList* impPopList(unsigned count,
3140 CORINFO_SIG_INFO* sig,
3141 GenTreeArgList* prefixTree = nullptr);
3143 GenTreeArgList* impPopRevList(unsigned count,
3145 CORINFO_SIG_INFO* sig,
3146 unsigned skipReverseCount = 0);
3149 * Get current IL offset with stack-empty info incoporated
3151 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3153 //---------------- Spilling the importer stack ----------------------------
3155 // The maximum number of bytes of IL processed without clean stack state.
3156 // It allows to limit the maximum tree size and depth.
3157 static const unsigned MAX_TREE_SIZE = 200;
3158 bool impCanSpillNow(OPCODE prevOpcode);
3164 SavedStack pdSavedStack;
3165 ThisInitState pdThisPtrInit;
3168 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3169 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3171 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3172 ExpandArray<BYTE> impPendingBlockMembers;
3174 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3175 // Operates on the map in the top-level ancestor.
3176 BYTE impGetPendingBlockMember(BasicBlock* blk)
3178 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3181 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3182 // Operates on the map in the top-level ancestor.
3183 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3185 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3188 bool impCanReimport;
3190 bool impSpillStackEntry(unsigned level,
3194 bool bAssertOnRecursion,
3199 void impSpillStackEnsure(bool spillLeaves = false);
3200 void impEvalSideEffects();
3201 void impSpillSpecialSideEff();
3202 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3203 void impSpillValueClasses();
3204 void impSpillEvalStack();
3205 static fgWalkPreFn impFindValueClasses;
3206 void impSpillLclRefs(ssize_t lclNum);
3208 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3210 void impImportBlockCode(BasicBlock* block);
3212 void impReimportMarkBlock(BasicBlock* block);
3213 void impReimportMarkSuccessors(BasicBlock* block);
3215 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3217 void impImportBlockPending(BasicBlock* block);
3219 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3220 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3221 // for the block, but instead, just re-uses the block's existing EntryState.
3222 void impReimportBlockPending(BasicBlock* block);
3224 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3226 void impImportBlock(BasicBlock* block);
3228 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3229 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3230 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3231 // the variables that will be used -- and for all the predecessors of those successors, and the
3232 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3233 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3234 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3235 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3236 // of local variable numbers, so we represent them with the base local variable number), returns that.
3237 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3238 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3239 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3240 // on which kind of member of the clique the block is).
3241 unsigned impGetSpillTmpBase(BasicBlock* block);
3243 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3244 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3245 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3246 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3247 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3248 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3249 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3250 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3251 // successors receive a native int. Similarly float and double are unified to double.
3252 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3253 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3254 // predecessors, so they insert an upcast if needed).
3255 void impReimportSpillClique(BasicBlock* block);
3257 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3258 // block, and represent the predecessor and successor members of the clique currently being computed.
3259 // *** Access to these will need to be locked in a parallel compiler.
3260 ExpandArray<BYTE> impSpillCliquePredMembers;
3261 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3269 // Abstract class for receiving a callback while walking a spill clique
3270 class SpillCliqueWalker
3273 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3276 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3277 class SetSpillTempsBase : public SpillCliqueWalker
3282 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3285 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3288 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3289 class ReimportSpillClique : public SpillCliqueWalker
3294 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3297 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3300 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3301 // predecessor or successor within the spill clique
3302 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3304 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3305 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3306 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3307 void impRetypeEntryStateTemps(BasicBlock* blk);
3309 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3310 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3312 void impPushVar(GenTree* op, typeInfo tiRetVal);
3313 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3314 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3316 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3318 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3319 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3320 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3323 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3326 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3327 struct BlockListNode
3330 BlockListNode* m_next;
3331 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3334 void* operator new(size_t sz, Compiler* comp);
3336 BlockListNode* impBlockListNodeFreeList;
3338 BlockListNode* AllocBlockListNode();
3339 void FreeBlockListNode(BlockListNode* node);
3341 bool impIsValueType(typeInfo* pTypeInfo);
3342 var_types mangleVarArgsType(var_types type);
3345 regNumber getCallArgIntRegister(regNumber floatReg);
3346 regNumber getCallArgFloatRegister(regNumber intReg);
3347 #endif // FEATURE_VARARG
3350 static unsigned jitTotalMethodCompiled;
3354 static LONG jitNestingLevel;
3357 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3359 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3361 // STATIC inlining decision based on the IL code.
3362 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3363 CORINFO_METHOD_INFO* methInfo,
3365 InlineResult* inlineResult);
3367 void impCheckCanInline(GenTreePtr call,
3368 CORINFO_METHOD_HANDLE fncHandle,
3370 CORINFO_CONTEXT_HANDLE exactContextHnd,
3371 InlineCandidateInfo** ppInlineCandidateInfo,
3372 InlineResult* inlineResult);
3374 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3375 GenTreePtr curArgVal,
3377 InlineResult* inlineResult);
3379 void impInlineInitVars(InlineInfo* pInlineInfo);
3381 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3383 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3385 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3387 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3388 GenTreePtr variableBeingDereferenced,
3389 InlArgInfo* inlArgInfo);
3391 void impMarkInlineCandidate(GenTreePtr call,
3392 CORINFO_CONTEXT_HANDLE exactContextHnd,
3393 bool exactContextNeedsRuntimeLookup,
3394 CORINFO_CALL_INFO* callInfo);
3396 bool impTailCallRetTypeCompatible(var_types callerRetType,
3397 CORINFO_CLASS_HANDLE callerRetTypeClass,
3398 var_types calleeRetType,
3399 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3401 bool impIsTailCallILPattern(bool tailPrefixed,
3403 const BYTE* codeAddrOfNextOpcode,
3404 const BYTE* codeEnd,
3406 bool* IsCallPopRet = nullptr);
3408 bool impIsImplicitTailCallCandidate(
3409 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3411 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3414 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3419 XX Info about the basic-blocks, their contents and the flow analysis XX
3421 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3422 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3426 BasicBlock* fgFirstBB; // Beginning of the basic block list
3427 BasicBlock* fgLastBB; // End of the basic block list
3428 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3429 #if FEATURE_EH_FUNCLETS
3430 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3432 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3434 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3435 unsigned fgEdgeCount; // # of control flow edges between the BBs
3436 unsigned fgBBcount; // # of BBs in the method
3438 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3440 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3441 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3442 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3443 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3445 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3446 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3447 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3448 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3449 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3450 // index). The arrays are of size fgBBNumMax + 1.
3451 unsigned* fgDomTreePreOrder;
3452 unsigned* fgDomTreePostOrder;
3454 bool fgBBVarSetsInited;
3456 // Allocate array like T* a = new T[fgBBNumMax + 1];
3457 // Using helper so we don't keep forgetting +1.
3458 template <typename T>
3459 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3461 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3464 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3465 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3466 // cannot be meaningfully combined. Note that new blocks can be created with higher
3467 // block numbers without changing the basic block epoch. These blocks *cannot*
3468 // participate in a block set until the blocks are all renumbered, causing the epoch
3469 // to change. This is useful if continuing to use previous block sets is valuable.
3470 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3471 unsigned fgCurBBEpoch;
3473 unsigned GetCurBasicBlockEpoch()
3475 return fgCurBBEpoch;
3478 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3479 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3480 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3481 unsigned fgCurBBEpochSize;
3483 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3484 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3485 unsigned fgBBSetCountInSizeTUnits;
3487 void NewBasicBlockEpoch()
3489 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3491 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3493 fgCurBBEpochSize = fgBBNumMax + 1;
3494 fgBBSetCountInSizeTUnits =
3495 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3498 // All BlockSet objects are now invalid!
3499 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3500 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3504 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3505 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3506 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3507 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3509 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3510 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3511 // array of size_t bitsets), then print that out.
3512 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3519 void EnsureBasicBlockEpoch()
3521 if (fgCurBBEpochSize != fgBBNumMax + 1)
3523 NewBasicBlockEpoch();
3527 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3528 void fgEnsureFirstBBisScratch();
3529 bool fgFirstBBisScratch();
3530 bool fgBBisScratch(BasicBlock* block);
3532 void fgExtendEHRegionBefore(BasicBlock* block);
3533 void fgExtendEHRegionAfter(BasicBlock* block);
3535 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3537 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3539 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3542 BasicBlock* nearBlk,
3543 bool putInFilter = false,
3544 bool runRarely = false,
3545 bool insertAtEnd = false);
3547 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3549 bool runRarely = false,
3550 bool insertAtEnd = false);
3552 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3554 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3555 BasicBlock* afterBlk,
3556 unsigned xcptnIndex,
3557 bool putInTryRegion);
3559 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3560 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3561 void fgUnlinkBlock(BasicBlock* block);
3563 unsigned fgMeasureIR();
3565 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3566 bool fgMultipleNots;
3569 bool fgModified; // True if the flow graph has been modified recently
3570 bool fgComputePredsDone; // Have we computed the bbPreds list
3571 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3572 bool fgDomsComputed; // Have we computed the dominator sets?
3573 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3575 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3576 bool fgHasPostfix; // any postfix ++/-- found?
3577 unsigned fgIncrCount; // number of increment nodes found
3579 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3583 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3584 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3587 bool fgRemoveRestOfBlock; // true if we know that we will throw
3588 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3590 // There are two modes for ordering of the trees.
3591 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3592 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3593 // by traversing the tree according to the order of the operands.
3594 // - In FGOrderLinear, the dominant ordering is the linear order.
3601 FlowGraphOrder fgOrder;
3603 // The following are boolean flags that keep track of the state of internal data structures
3605 bool fgStmtListThreaded;
3606 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3607 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3608 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3609 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3610 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3611 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3612 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3613 // This is derived from the profile data
3614 // or is BB_UNITY_WEIGHT when we don't have profile data
3616 #if FEATURE_EH_FUNCLETS
3617 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3618 #endif // FEATURE_EH_FUNCLETS
3620 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3621 // since fgMorphTree can be called from several places
3622 bool fgExpandInline; // indicates that we are creating tree for the inliner
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 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3659 BasicBlock* handler,
3660 BlockToBlockMap& continuationMap);
3662 GenTreePtr fgGetCritSectOfStaticMethod();
3664 #if FEATURE_EH_FUNCLETS
3666 void fgAddSyncMethodEnterExit();
3668 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3670 void fgConvertSyncReturnToLeave(BasicBlock* block);
3672 #endif // FEATURE_EH_FUNCLETS
3674 void fgAddReversePInvokeEnterExit();
3676 bool fgMoreThanOneReturnBlock();
3678 // The number of separate return points in the method.
3679 unsigned fgReturnCount;
3681 void fgAddInternal();
3683 bool fgFoldConditional(BasicBlock* block);
3685 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3686 void fgMorphBlocks();
3688 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3690 void fgCheckArgCnt();
3691 void fgSetOptions();
3694 static fgWalkPreFn fgAssertNoQmark;
3695 void fgPreExpandQmarkChecks(GenTreePtr expr);
3696 void fgPostExpandQmarkChecks();
3697 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3700 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3702 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3703 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3704 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3705 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3706 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3708 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3709 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3710 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3711 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3713 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3714 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3715 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3716 void fgExpandQmarkNodes();
3720 // Do "simple lowering." This functionality is (conceptually) part of "general"
3721 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3722 void fgSimpleLowering();
3724 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3726 GenTreePtr fgInitThisClass();
3728 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3730 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3732 inline bool backendRequiresLocalVarLifetimes()
3734 #if defined(LEGACY_BACKEND)
3737 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3741 void fgLocalVarLiveness();
3743 void fgLocalVarLivenessInit();
3745 #ifdef LEGACY_BACKEND
3746 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3748 void fgPerNodeLocalVarLiveness(GenTree* node);
3750 void fgPerBlockLocalVarLiveness();
3752 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3754 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3756 // This is used in the liveness computation, as a temporary. When we use the
3757 // arbitrary-length VarSet representation, it is better not to allocate a new one
3759 VARSET_TP fgMarkIntfUnionVS;
3761 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3763 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3765 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3767 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3769 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3771 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3773 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3775 void fgComputeLife(VARSET_TP& life,
3776 GenTreePtr startNode,
3778 VARSET_VALARG_TP volatileVars,
3779 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3781 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3783 bool fgRemoveDeadStore(GenTree** pTree,
3785 VARSET_VALARG_TP life,
3787 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3789 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3791 // For updating liveset during traversal AFTER fgComputeLife has completed
3792 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3793 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3795 // Returns the set of live variables after endTree,
3796 // assuming that liveSet is the set of live variables BEFORE tree.
3797 // Requires that fgComputeLife has completed, and that tree is in the same
3798 // statement as endTree, and that it comes before endTree in execution order
3800 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3802 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3803 while (tree != nullptr && tree != endTree->gtNext)
3805 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3806 tree = tree->gtNext;
3808 assert(tree == endTree->gtNext);
3812 void fgInterBlockLocalVarLiveness();
3814 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3815 // "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
3816 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3817 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3818 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3819 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3820 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3822 if (m_opAsgnVarDefSsaNums == nullptr)
3824 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3826 return m_opAsgnVarDefSsaNums;
3829 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3830 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3831 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3833 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3835 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3836 // Except: assumes that lcl is a def, and if it is
3837 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3838 // rather than the "use" SSA number recorded in the tree "lcl".
3839 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3841 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3842 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3843 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3844 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3845 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3847 // (byref addrS1 = &s1,
3848 // *(addrS1 * offsetof(f0)) = s2f0,
3850 // *(addrS1 * offsetof(fn)) = s2fn)
3852 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3853 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3854 // give it SSA names and value numbers?
3856 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3857 // end with an instance of the structure below, whose fields are described in the declaration.
3858 struct IndirectAssignmentAnnotation
3860 unsigned m_lclNum; // The local num that is being indirectly assigned.
3861 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3862 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3863 // be the singleton field sequence "g". The individual assignments would
3864 // further append the fields of "s.g" to that.
3865 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3866 // structure has a single field).
3867 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3868 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3871 IndirectAssignmentAnnotation(unsigned lclNum,
3872 FieldSeqNode* fldSeq,
3874 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3875 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3876 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3880 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3881 NodeToIndirAssignMap;
3882 NodeToIndirAssignMap* m_indirAssignMap;
3883 NodeToIndirAssignMap* GetIndirAssignMap()
3885 if (m_indirAssignMap == nullptr)
3887 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3888 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3889 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3891 return m_indirAssignMap;
3894 // Performs SSA conversion.
3897 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3898 void fgResetForSsa();
3900 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3902 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3903 inline bool fgExcludeFromSsa(unsigned lclNum);
3905 // The value numbers for this compilation.
3906 ValueNumStore* vnStore;
3909 ValueNumStore* GetValueNumStore()
3914 // Do value numbering (assign a value number to each
3916 void fgValueNumber();
3918 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3919 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3920 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3921 // match the element type of the array or fldSeq. When this type doesn't match
3922 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3924 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3927 FieldSeqNode* fldSeq,
3931 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3932 // has been parsed to yield the other input arguments. If evaluation of the address
3933 // can raise exceptions, those should be captured in the exception set "excVN."
3934 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3935 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3936 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3937 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3938 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3940 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3941 CORINFO_CLASS_HANDLE elemTypeEq,
3945 FieldSeqNode* fldSeq);
3947 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3948 // by evaluating the array index expression "tree". Returns the value number resulting from
3949 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3950 // "GT_IND" that does the dereference, and it is given the returned value number.
3951 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3953 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3954 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3956 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3958 // Utility functions for fgValueNumber.
3960 // Perform value-numbering for the trees in "blk".
3961 void fgValueNumberBlock(BasicBlock* blk);
3963 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3964 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3965 // assumed for the memoryKind at the start "entryBlk".
3966 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3968 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3969 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3970 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3972 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3974 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3976 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3977 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3978 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3980 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3981 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3983 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3984 // value in that SSA #.
3985 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3987 // The input 'tree' is a leaf node that is a constant
3988 // Assign the proper value number to the tree
3989 void fgValueNumberTreeConst(GenTreePtr tree);
3991 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3992 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3994 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3996 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3998 // Does value-numbering for a block assignment.
3999 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4001 // Does value-numbering for a cast tree.
4002 void fgValueNumberCastTree(GenTreePtr tree);
4004 // Does value-numbering for an intrinsic tree.
4005 void fgValueNumberIntrinsic(GenTreePtr tree);
4007 // Does value-numbering for a call. We interpret some helper calls.
4008 void fgValueNumberCall(GenTreeCall* call);
4010 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4011 void fgUpdateArgListVNs(GenTreeArgList* args);
4013 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4014 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4016 // Requires "helpCall" to be a helper call. Assigns it a value number;
4017 // we understand the semantics of some of the calls. Returns "true" if
4018 // the call may modify the heap (we assume arbitrary memory side effects if so).
4019 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4021 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4022 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4024 // These are the current value number for the memory implicit variables while
4025 // doing value numbering. These are the value numbers under the "liberal" interpretation
4026 // of memory values; the "conservative" interpretation needs no VN, since every access of
4027 // memory yields an unknown value.
4028 ValueNum fgCurMemoryVN[MemoryKindCount];
4030 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4031 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4032 // is 1, and the rest is an encoding of "elemTyp".
4033 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4035 if (elemStructType != nullptr)
4037 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4038 varTypeIsIntegral(elemTyp));
4039 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4040 return elemStructType;
4044 elemTyp = varTypeUnsignedToSigned(elemTyp);
4045 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4048 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4049 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4050 // the struct type of the element).
4051 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4053 size_t clsHndVal = size_t(clsHnd);
4054 if (clsHndVal & 0x1)
4056 return var_types(clsHndVal >> 1);
4064 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4065 var_types getJitGCType(BYTE gcType);
4067 enum structPassingKind
4069 SPK_Unknown, // Invalid value, never returned
4070 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4071 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4072 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4073 // parameters registers are used, then the stack will be used)
4074 // for X86 passed on the stack, for ARM32 passed in registers
4075 // or the stack or split between registers and the stack.
4076 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4078 }; // The struct is passed/returned by reference to a copy/buffer.
4080 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4081 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4082 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4083 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4085 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4087 // Get the type that is used to pass values of the given struct type.
4088 // If you have already retrieved the struct size then pass it as the optional third argument
4090 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4091 structPassingKind* wbPassStruct,
4092 unsigned structSize = 0);
4094 // Get the type that is used to return values of the given struct type.
4095 // If you have already retrieved the struct size then pass it as the optional third argument
4097 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4098 structPassingKind* wbPassStruct = nullptr,
4099 unsigned structSize = 0);
4102 // Print a representation of "vnp" or "vn" on standard output.
4103 // If "level" is non-zero, we also print out a partial expansion of the value.
4104 void vnpPrint(ValueNumPair vnp, unsigned level);
4105 void vnPrint(ValueNum vn, unsigned level);
4108 // Dominator computation member functions
4109 // Not exposed outside Compiler
4111 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4113 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4115 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4116 // flow graph. We first assume the fields bbIDom on each
4117 // basic block are invalid. This computation is needed later
4118 // by fgBuildDomTree to build the dominance tree structure.
4119 // Based on: A Simple, Fast Dominance Algorithm
4120 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4122 void fgCompDominatedByExceptionalEntryBlocks();
4124 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4125 // Note: this is relatively slow compared to calling fgDominate(),
4126 // especially if dealing with a single block versus block check.
4128 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4130 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4132 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4134 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4136 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4138 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4139 // processed in topological sort, this function takes care of that.
4141 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4143 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4144 // Returns this as a set.
4146 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4147 // root nodes. Returns this as a set.
4150 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4153 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4154 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4157 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4158 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4159 // && postOrder(A) >= postOrder(B) making the computation O(1).
4160 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4162 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4164 void fgUpdateChangedFlowGraph();
4167 // Compute the predecessors of the blocks in the control flow graph.
4168 void fgComputePreds();
4170 // Remove all predecessor information.
4171 void fgRemovePreds();
4173 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4174 // before the full predecessors lists are computed.
4175 void fgComputeCheapPreds();
4178 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4180 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4190 // Initialize the per-block variable sets (used for liveness analysis).
4191 void fgInitBlockVarSets();
4193 // true if we've gone through and created GC Poll calls.
4194 bool fgGCPollsCreated;
4195 void fgMarkGCPollBlocks();
4196 void fgCreateGCPolls();
4197 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4199 // Requires that "block" is a block that returns from
4200 // a finally. Returns the number of successors (jump targets of
4201 // of blocks in the covered "try" that did a "LEAVE".)
4202 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4204 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4205 // a finally. Returns its "i"th successor (jump targets of
4206 // of blocks in the covered "try" that did a "LEAVE".)
4207 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4208 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4211 // Factor out common portions of the impls of the methods above.
4212 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4215 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4216 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4217 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4218 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4219 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4220 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4221 // we leave the entry associated with the block, but it will no longer be accessed.)
4222 struct SwitchUniqueSuccSet
4224 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4225 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4228 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4229 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4230 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4231 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4234 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4235 BlockToSwitchDescMap;
4238 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4239 // iteration over only the distinct successors.
4240 BlockToSwitchDescMap* m_switchDescMap;
4243 BlockToSwitchDescMap* GetSwitchDescMap()
4245 if (m_switchDescMap == nullptr)
4247 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4249 return m_switchDescMap;
4252 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4253 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4254 // we don't accidentally look up and return the wrong switch data.
4255 void InvalidateUniqueSwitchSuccMap()
4257 m_switchDescMap = nullptr;
4260 // Requires "switchBlock" to be a block that ends in a switch. Returns
4261 // the corresponding SwitchUniqueSuccSet.
4262 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4264 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4265 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4266 // remove it from "this", and ensure that "to" is a member.
4267 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4269 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4270 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4272 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4274 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4276 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4278 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4280 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4282 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4284 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4286 void fgRemoveBlockAsPred(BasicBlock* block);
4288 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4290 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4292 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4294 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4296 flowList* fgAddRefPred(BasicBlock* block,
4297 BasicBlock* blockPred,
4298 flowList* oldEdge = nullptr,
4299 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4302 void fgFindBasicBlocks();
4304 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4306 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4308 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4309 bool putInTryRegion,
4310 BasicBlock* startBlk,
4312 BasicBlock* nearBlk,
4313 BasicBlock* jumpBlk,
4316 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4318 void fgRemoveEmptyBlocks();
4320 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4322 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4324 void fgCreateLoopPreHeader(unsigned lnum);
4326 void fgUnreachableBlock(BasicBlock* block);
4328 void fgRemoveConditionalJump(BasicBlock* block);
4330 BasicBlock* fgLastBBInMainFunction();
4332 BasicBlock* fgEndBBAfterMainFunction();
4334 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4336 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4338 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4340 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4342 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4344 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4346 bool fgRenumberBlocks();
4348 bool fgExpandRarelyRunBlocks();
4350 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4352 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4354 enum FG_RELOCATE_TYPE
4356 FG_RELOCATE_TRY, // relocate the 'try' region
4357 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4359 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4361 #if FEATURE_EH_FUNCLETS
4362 #if defined(_TARGET_ARM_)
4363 void fgClearFinallyTargetBit(BasicBlock* block);
4364 #endif // defined(_TARGET_ARM_)
4365 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4366 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4367 void fgInsertFuncletPrologBlock(BasicBlock* block);
4368 void fgCreateFuncletPrologBlocks();
4369 void fgCreateFunclets();
4370 #else // !FEATURE_EH_FUNCLETS
4371 bool fgRelocateEHRegions();
4372 #endif // !FEATURE_EH_FUNCLETS
4374 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4376 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4378 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4380 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4382 bool fgOptimizeEmptyBlock(BasicBlock* block);
4384 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4386 bool fgOptimizeBranch(BasicBlock* bJump);
4388 bool fgOptimizeSwitchBranches(BasicBlock* block);
4390 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4392 bool fgOptimizeSwitchJumps();
4394 void fgPrintEdgeWeights();
4396 void fgComputeEdgeWeights();
4398 void fgReorderBlocks();
4400 void fgDetermineFirstColdBlock();
4402 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4404 bool fgUpdateFlowGraph(bool doTailDup = false);
4406 void fgFindOperOrder();
4408 // method that returns if you should split here
4409 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4411 void fgSetBlockOrder();
4413 void fgRemoveReturnBlock(BasicBlock* block);
4415 /* Helper code that has been factored out */
4416 inline void fgConvertBBToThrowBB(BasicBlock* block);
4418 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4419 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4420 GenTreePtr fgMakeTmpArgNode(
4421 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4423 // The following check for loops that don't execute calls
4424 bool fgLoopCallMarked;
4426 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4427 void fgLoopCallMark();
4429 void fgMarkLoopHead(BasicBlock* block);
4431 unsigned fgGetCodeEstimate(BasicBlock* block);
4434 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4435 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4436 bool fgDumpFlowGraph(Phases phase);
4438 #endif // DUMP_FLOWGRAPHS
4443 void fgDispBBLiveness(BasicBlock* block);
4444 void fgDispBBLiveness();
4445 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4446 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4447 void fgDispBasicBlocks(bool dumpTrees = false);
4448 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4449 void fgDumpBlock(BasicBlock* block);
4450 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4452 static fgWalkPreFn fgStress64RsltMulCB;
4453 void fgStress64RsltMul();
4454 void fgDebugCheckUpdate();
4455 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4456 void fgDebugCheckBlockLinks();
4457 void fgDebugCheckLinks(bool morphTrees = false);
4458 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4459 void fgDebugCheckFlags(GenTreePtr tree);
4460 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4461 void fgDebugCheckTryFinallyExits();
4464 #ifdef LEGACY_BACKEND
4465 static void fgOrderBlockOps(GenTreePtr tree,
4469 GenTreePtr* opsPtr, // OUT
4470 regMaskTP* regsPtr); // OUT
4471 #endif // LEGACY_BACKEND
4473 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4474 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4476 inline bool fgIsInlining()
4478 return fgExpandInline;
4481 void fgTraverseRPO();
4483 //--------------------- Walking the trees in the IR -----------------------
4488 fgWalkPreFn* wtprVisitorFn;
4489 fgWalkPostFn* wtpoVisitorFn;
4490 void* pCallbackData; // user-provided data
4491 bool wtprLclsOnly; // whether to only visit lclvar nodes
4492 GenTreePtr parent; // parent of current node, provided to callback
4493 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4495 bool printModified; // callback can use this
4499 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4500 fgWalkPreFn* visitor,
4501 void* pCallBackData = nullptr,
4502 bool lclVarsOnly = false,
4503 bool computeStack = false);
4505 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4506 fgWalkPreFn* preVisitor,
4507 fgWalkPostFn* postVisitor,
4508 void* pCallBackData = nullptr);
4510 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4514 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4515 fgWalkPostFn* visitor,
4516 void* pCallBackData = nullptr,
4517 bool computeStack = false);
4519 // An fgWalkPreFn that looks for expressions that have inline throws in
4520 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4521 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4522 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4523 // properly propagated to parent trees). It returns WALK_CONTINUE
4525 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4526 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4527 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4529 /**************************************************************************
4531 *************************************************************************/
4534 friend class SsaBuilder;
4535 friend struct ValueNumberState;
4537 //--------------------- Detect the basic blocks ---------------------------
4539 BasicBlock** fgBBs; // Table of pointers to the BBs
4541 void fgInitBBLookup();
4542 BasicBlock* fgLookupBB(unsigned addr);
4544 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4546 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4548 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4550 void fgLinkBasicBlocks();
4552 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4554 void fgCheckBasicBlockControlFlow();
4556 void fgControlFlowPermitted(BasicBlock* blkSrc,
4557 BasicBlock* blkDest,
4558 BOOL IsLeave = false /* is the src a leave block */);
4560 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4562 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4564 void fgAdjustForAddressExposedOrWrittenThis();
4566 bool fgProfileData_ILSizeMismatch;
4567 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4568 ULONG fgProfileBufferCount;
4569 ULONG fgNumProfileRuns;
4571 unsigned fgStressBBProf()
4574 unsigned result = JitConfig.JitStressBBProf();
4577 if (compStressCompile(STRESS_BB_PROFILE, 15))
4588 bool fgHaveProfileData();
4589 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4590 void fgInstrumentMethod();
4593 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4594 // or if we have some fake profile data for the stress mode
4595 bool fgIsUsingProfileWeights()
4597 return (fgHaveProfileData() || fgStressBBProf());
4600 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4601 // or BB_UNITY_WEIGHT when we aren't using profile data.
4602 unsigned fgProfileRunsCount()
4604 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4607 //-------- Insert a statement at the start or end of a basic block --------
4611 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4615 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4617 public: // Used by linear scan register allocation
4618 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4621 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4622 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4624 public: // Used by linear scan register allocation
4625 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4628 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4630 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4632 // Create a new temporary variable to hold the result of *ppTree,
4633 // and transform the graph accordingly.
4634 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4635 GenTree* fgMakeMultiUse(GenTree** ppTree);
4638 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4639 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4640 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4642 //-------- Determine the order in which the trees will be evaluated -------
4644 unsigned fgTreeSeqNum;
4645 GenTree* fgTreeSeqLst;
4646 GenTree* fgTreeSeqBeg;
4648 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4649 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4650 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4651 void fgSetStmtSeq(GenTree* tree);
4652 void fgSetBlockOrder(BasicBlock* block);
4654 //------------------------- Morphing --------------------------------------
4656 unsigned fgPtrArgCntCur;
4657 unsigned fgPtrArgCntMax;
4658 hashBv* fgOutgoingArgTemps;
4659 hashBv* fgCurrentlyInUseArgTemps;
4661 bool compCanEncodePtrArgCntMax();
4663 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4666 void fgMoveOpsLeft(GenTreePtr tree);
4669 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4671 bool fgIsThrow(GenTreePtr tree);
4673 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4674 bool fgIsBlockCold(BasicBlock* block);
4676 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4678 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4680 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4682 bool fgMorphRelopToQmark(GenTreePtr tree);
4684 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4685 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4686 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4687 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4688 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4689 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4690 // small; hence the other fields of MorphAddrContext.
4691 enum MorphAddrContextKind
4696 struct MorphAddrContext
4698 MorphAddrContextKind m_kind;
4699 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4700 // top-level indirection and here have been constants.
4701 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4702 // In that case, is the sum of those constant offsets.
4704 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4709 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4710 static MorphAddrContext s_CopyBlockMAC;
4713 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4714 var_types* baseTypeOut,
4716 unsigned* simdSizeOut,
4717 bool ignoreUsedInSIMDIntrinsic = false);
4718 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4719 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4720 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4721 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4723 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4724 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4725 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4727 #endif // FEATURE_SIMD
4728 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4729 GenTreePtr fgMorphCast(GenTreePtr tree);
4730 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4731 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4733 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4736 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4737 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4739 void fgFixupStructReturn(GenTreePtr call);
4740 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4741 bool fgAddrCouldBeNull(GenTreePtr addr);
4742 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4743 bool fgCanFastTailCall(GenTreeCall* call);
4744 void fgMorphTailCall(GenTreeCall* call);
4745 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4746 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4747 fgArgTabEntryPtr argTabEntry,
4749 IL_OFFSETX callILOffset,
4750 GenTreePtr tmpAssignmentInsertionPoint,
4751 GenTreePtr paramAssignmentInsertionPoint);
4752 static int fgEstimateCallStackSize(GenTreeCall* call);
4753 GenTreePtr fgMorphCall(GenTreeCall* call);
4754 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4755 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4757 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4758 static fgWalkPreFn fgFindNonInlineCandidate;
4760 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4761 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4762 CORINFO_RESOLVED_TOKEN* ldftnToken);
4763 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4764 void fgAssignSetVarDef(GenTreePtr tree);
4765 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4766 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4767 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4768 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4769 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4770 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4771 void fgMorphUnsafeBlk(GenTreeObj* obj);
4772 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4773 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4774 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4775 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4776 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4777 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4778 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4780 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4781 GenTreePtr fgMorphConst(GenTreePtr tree);
4784 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4787 #if LOCAL_ASSERTION_PROP
4788 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4789 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4791 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4793 GenTreeStmt* fgMorphStmt;
4795 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4796 // used when morphing big offset.
4798 //----------------------- Liveness analysis -------------------------------
4800 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4801 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4803 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4804 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4805 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4807 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4809 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4811 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4812 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4814 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4815 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4817 void fgExtendDbgScopes();
4818 void fgExtendDbgLifetimes();
4821 void fgDispDebugScopes();
4824 //-------------------------------------------------------------------------
4826 // The following keeps track of any code we've added for things like array
4827 // range checking or explicit calls to enable GC, and so on.
4832 AddCodeDsc* acdNext;
4833 BasicBlock* acdDstBlk; // block to which we jump
4835 SpecialCodeKind acdKind; // what kind of a special block is this?
4836 unsigned short acdStkLvl;
4840 static unsigned acdHelper(SpecialCodeKind codeKind);
4842 AddCodeDsc* fgAddCodeList;
4844 bool fgRngChkThrowAdded;
4845 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4847 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4849 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4852 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4855 bool fgIsCodeAdded();
4857 bool fgIsThrowHlpBlk(BasicBlock* block);
4858 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4860 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4862 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4863 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4864 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4865 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4866 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4868 #if FEATURE_MULTIREG_RET
4869 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4870 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4871 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4872 #endif // FEATURE_MULTIREG_RET
4874 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4877 static fgWalkPreFn fgDebugCheckInlineCandidates;
4879 void CheckNoFatPointerCandidatesLeft();
4880 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4883 void fgPromoteStructs();
4884 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4885 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4887 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4888 void fgMarkImplicitByRefArgs();
4890 // Change implicit byrefs' types from struct to pointer, and for any that were
4891 // promoted, create new promoted struct temps.
4892 void fgRetypeImplicitByRefArgs();
4894 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4895 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4896 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4898 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4899 void fgMarkDemotedImplicitByRefArgs();
4901 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4902 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4903 void fgMarkAddressExposedLocals();
4904 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4906 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4908 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4910 // The given local variable, required to be a struct variable, is being assigned via
4911 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4912 // the variable is not enregistered, and is therefore not promoted independently.
4913 void fgLclFldAssign(unsigned lclNum);
4915 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4916 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4917 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4918 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4921 bool fgPrintInlinedMethods;
4924 bool fgIsBigOffset(size_t offset);
4926 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4927 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4928 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4929 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4930 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4933 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4934 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4938 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4939 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4946 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4949 void optRemoveRangeCheck(
4950 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4951 bool optIsRangeCheckRemovable(GenTreePtr tree);
4954 static fgWalkPreFn optValidRangeCheckIndex;
4955 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4958 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4960 /**************************************************************************
4962 *************************************************************************/
4965 // Do hoisting for all loops.
4966 void optHoistLoopCode();
4968 // To represent sets of VN's that have already been hoisted in outer loops.
4969 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4970 typedef VNToBoolMap VNSet;
4972 struct LoopHoistContext
4975 // The set of variables hoisted in the current loop (or nullptr if there are none).
4976 VNSet* m_pHoistedInCurLoop;
4979 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4980 VNSet m_hoistedInParentLoops;
4981 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4982 // Previous decisions on loop-invariance of value numbers in the current loop.
4983 VNToBoolMap m_curLoopVnInvariantCache;
4985 VNSet* GetHoistedInCurLoop(Compiler* comp)
4987 if (m_pHoistedInCurLoop == nullptr)
4989 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4991 return m_pHoistedInCurLoop;
4994 VNSet* ExtractHoistedInCurLoop()
4996 VNSet* res = m_pHoistedInCurLoop;
4997 m_pHoistedInCurLoop = nullptr;
5001 LoopHoistContext(Compiler* comp)
5002 : m_pHoistedInCurLoop(nullptr)
5003 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5004 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5009 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5010 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5011 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5012 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5014 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5015 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5016 // "m_hoistedInParentLoops".
5018 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5020 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5021 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5022 // expressions to "hoistInLoop".
5023 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5025 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5026 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5028 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5029 // that are invariant in loop "lnum" (an index into the optLoopTable)
5030 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5031 // expressions to "hoistInLoop".
5032 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5033 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5034 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5035 bool optHoistLoopExprsForTree(GenTreePtr tree,
5037 LoopHoistContext* hoistCtxt,
5038 bool* firstBlockAndBeforeSideEffect,
5040 bool* pCctorDependent);
5042 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5043 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5045 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5046 // Constants and init values are always loop invariant.
5047 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5048 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5050 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5051 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5052 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5053 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5054 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5056 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5057 // in the loop table.
5058 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5060 // Records the set of "side effects" of all loops: fields (object instance and static)
5061 // written to, and SZ-array element type equivalence classes updated.
5062 void optComputeLoopSideEffects();
5065 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5066 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5067 // static) written to, and SZ-array element type equivalence classes updated.
5068 void optComputeLoopNestSideEffects(unsigned lnum);
5070 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5071 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5073 // Hoist the expression "expr" out of loop "lnum".
5074 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5077 void optOptimizeBools();
5080 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5082 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5085 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5087 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5088 // the loop into a "do-while" loop
5089 // Also finds all natural loops and records them in the loop table
5091 // Optionally clone loops in the loop table.
5092 void optCloneLoops();
5094 // Clone loop "loopInd" in the loop table.
5095 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5097 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5098 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5099 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5101 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5103 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5106 // This enumeration describes what is killed by a call.
5110 CALLINT_NONE, // no interference (most helpers)
5111 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5112 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5113 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5114 CALLINT_ALL, // kills everything (normal method call)
5118 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5119 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5120 // in bbNext order; we use comparisons on the bbNum to decide order.)
5121 // The blocks that define the body are
5122 // first <= top <= entry <= bottom .
5123 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5124 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5125 // Compiler::optFindNaturalLoops().
5128 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5129 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5130 // loop, but not the outer loop.)
5131 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5133 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5134 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5135 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5137 callInterf lpAsgCall; // "callInterf" for calls in the loop
5138 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5139 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5141 unsigned short lpFlags; // Mask of the LPFLG_* constants
5143 unsigned char lpExitCnt; // number of exits from the loop
5145 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5146 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5147 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5148 // (Actually, an "immediately" nested loop --
5149 // no other child of this loop is a parent of lpChild.)
5150 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5151 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5152 // by following "lpChild" then "lpSibling" links.
5154 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5155 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5157 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5158 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5159 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5161 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5162 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5164 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5165 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5166 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5167 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5169 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5170 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5171 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5173 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5174 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5175 // type are assigned to.
5177 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5178 // memory side effects. If this is set, the fields below
5179 // may not be accurate (since they become irrelevant.)
5180 bool lpContainsCall; // True if executing the loop body *may* execute a call
5182 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5183 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5185 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5187 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5188 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5190 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5192 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5193 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5195 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5196 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5198 JitSimplerHashBehavior>
5200 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5201 // instance fields modified
5204 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5205 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5207 JitSimplerHashBehavior>
5209 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5210 // arrays of that type are modified
5213 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5214 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5216 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5217 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5218 // (shifted left, with a low-order bit set to distinguish.)
5219 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5220 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5222 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5224 GenTreePtr lpIterTree; // The "i <op>= const" tree
5225 unsigned lpIterVar(); // iterator variable #
5226 int lpIterConst(); // the constant with which the iterator is incremented
5227 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5228 void VERIFY_lpIterTree();
5230 var_types lpIterOperType(); // For overflow instructions
5233 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5234 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5238 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5240 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5241 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5242 void VERIFY_lpTestTree();
5244 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5245 GenTreePtr lpIterator(); // the iterator node in the loop test
5246 GenTreePtr lpLimit(); // the limit node in the loop test
5248 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5249 // LPFLG_CONST_LIMIT
5250 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5252 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5253 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5254 // LPFLG_ARRLEN_LIMIT
5256 // Returns "true" iff "*this" contains the blk.
5257 bool lpContains(BasicBlock* blk)
5259 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5261 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5262 // to be equal, but requiring bottoms to be different.)
5263 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5265 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5268 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5269 // bottoms to be different.)
5270 bool lpContains(const LoopDsc& lp2)
5272 return lpContains(lp2.lpFirst, lp2.lpBottom);
5275 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5276 // (allowing firsts to be equal, but requiring bottoms to be different.)
5277 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5279 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5282 // Returns "true" iff "*this" is (properly) contained by "lp2"
5283 // (allowing firsts to be equal, but requiring bottoms to be different.)
5284 bool lpContainedBy(const LoopDsc& lp2)
5286 return lpContains(lp2.lpFirst, lp2.lpBottom);
5289 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5290 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5292 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5294 // Returns "true" iff "*this" is disjoint from "lp2".
5295 bool lpDisjoint(const LoopDsc& lp2)
5297 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5299 // Returns "true" iff the loop is well-formed (see code for defn).
5302 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5303 lpEntry->bbNum <= lpBottom->bbNum &&
5304 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5309 bool fgMightHaveLoop(); // returns true if there are any backedges
5310 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5313 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5314 unsigned char optLoopCount; // number of tracked loops
5317 unsigned optCallCount; // number of calls made in the method
5318 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5319 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5320 unsigned optLoopsCloned; // number of loops cloned in the current method.
5323 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5324 void optPrintLoopInfo(unsigned loopNum,
5326 BasicBlock* lpFirst,
5328 BasicBlock* lpEntry,
5329 BasicBlock* lpBottom,
5330 unsigned char lpExitCnt,
5332 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5333 void optPrintLoopInfo(unsigned lnum);
5334 void optPrintLoopRecording(unsigned lnum);
5336 void optCheckPreds();
5339 void optSetBlockWeights();
5341 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5343 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5345 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5347 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5348 unsigned optIsLoopIncrTree(GenTreePtr incr);
5349 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5350 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5351 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5352 bool optExtractInitTestIncr(BasicBlock* head,
5357 GenTreePtr* ppIncr);
5359 void optRecordLoop(BasicBlock* head,
5365 unsigned char exitCnt);
5367 void optFindNaturalLoops();
5369 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5370 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5371 bool optCanonicalizeLoopNest(unsigned char loopInd);
5373 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5374 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5375 bool optCanonicalizeLoop(unsigned char loopInd);
5377 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5378 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5379 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5380 bool optLoopContains(unsigned l1, unsigned l2);
5382 // Requires "loopInd" to be a valid index into the loop table.
5383 // Updates the loop table by changing loop "loopInd", whose head is required
5384 // to be "from", to be "to". Also performs this transformation for any
5385 // loop nested in "loopInd" that shares the same head as "loopInd".
5386 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5388 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5389 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5390 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5392 // Marks the containsCall information to "lnum" and any parent loops.
5393 void AddContainsCallAllContainingLoops(unsigned lnum);
5394 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5395 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5396 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5397 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5398 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5399 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5401 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5402 // of "from".) Copies the jump destination from "from" to "to".
5403 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5405 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5406 unsigned optLoopDepth(unsigned lnum)
5408 unsigned par = optLoopTable[lnum].lpParent;
5409 if (par == BasicBlock::NOT_IN_LOOP)
5415 return 1 + optLoopDepth(par);
5419 void fgOptWhileLoop(BasicBlock* block);
5421 bool optComputeLoopRep(int constInit,
5424 genTreeOps iterOper,
5426 genTreeOps testOper,
5429 unsigned* iterCount);
5430 #if FEATURE_STACK_FP_X87
5433 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5434 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5435 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5436 #endif // FEATURE_STACK_FP_X87
5439 static fgWalkPreFn optIsVarAssgCB;
5442 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5444 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5446 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5448 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5450 /**************************************************************************
5451 * Optimization conditions
5452 *************************************************************************/
5454 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5455 bool optPentium4(void);
5456 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5457 bool optAvoidIntMult(void);
5462 // The following is the upper limit on how many expressions we'll keep track
5463 // of for the CSE analysis.
5465 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5467 static const int MIN_CSE_COST = 2;
5469 // Keeps tracked cse indices
5470 BitVecTraits* cseTraits;
5473 /* Generic list of nodes - used by the CSE logic */
5481 typedef struct treeLst* treeLstPtr;
5485 treeStmtLst* tslNext;
5486 GenTreePtr tslTree; // tree node
5487 GenTreePtr tslStmt; // statement containing the tree
5488 BasicBlock* tslBlock; // block containing the statement
5491 typedef struct treeStmtLst* treeStmtLstPtr;
5493 // The following logic keeps track of expressions via a simple hash table.
5497 CSEdsc* csdNextInBucket; // used by the hash table
5499 unsigned csdHashValue; // the orginal hashkey
5501 unsigned csdIndex; // 1..optCSECandidateCount
5502 char csdLiveAcrossCall; // 0 or 1
5504 unsigned short csdDefCount; // definition count
5505 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5507 unsigned csdDefWtCnt; // weighted def count
5508 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5510 GenTreePtr csdTree; // treenode containing the 1st occurance
5511 GenTreePtr csdStmt; // stmt containing the 1st occurance
5512 BasicBlock* csdBlock; // block containing the 1st occurance
5514 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5515 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5517 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5518 // number, this will reflect it; otherwise, NoVN.
5521 static const size_t s_optCSEhashSize;
5522 CSEdsc** optCSEhash;
5525 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5527 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5528 // re-numbered with the bound to improve range check elimination
5530 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5531 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5535 CSEdsc* optCSEfindDsc(unsigned index);
5536 void optUnmarkCSE(GenTreePtr tree);
5538 // user defined callback data for the tree walk function optCSE_MaskHelper()
5539 struct optCSE_MaskData
5541 EXPSET_TP CSE_defMask;
5542 EXPSET_TP CSE_useMask;
5545 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5546 static fgWalkPreFn optCSE_MaskHelper;
5548 // This function walks all the node for an given tree
5549 // and return the mask of CSE definitions and uses for the tree
5551 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5553 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5554 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5555 bool optCSE_canSwap(GenTree* tree);
5557 static fgWalkPostFn optPropagateNonCSE;
5558 static fgWalkPreFn optHasNonCSEChild;
5560 static fgWalkPreFn optUnmarkCSEs;
5562 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5563 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5565 void optCleanupCSEs();
5568 void optEnsureClearCSEInfo();
5571 #endif // FEATURE_ANYCSE
5573 #if FEATURE_VALNUM_CSE
5574 /**************************************************************************
5575 * Value Number based CSEs
5576 *************************************************************************/
5579 void optOptimizeValnumCSEs();
5582 void optValnumCSE_Init();
5583 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5584 unsigned optValnumCSE_Locate();
5585 void optValnumCSE_InitDataFlow();
5586 void optValnumCSE_DataFlow();
5587 void optValnumCSE_Availablity();
5588 void optValnumCSE_Heuristic();
5589 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5591 #endif // FEATURE_VALNUM_CSE
5594 bool optDoCSE; // True when we have found a duplicate CSE tree
5595 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5596 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5597 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5598 unsigned optCSEstart; // The first local variable number that is a CSE
5599 unsigned optCSEcount; // The total count of CSE's introduced.
5600 unsigned optCSEweight; // The weight of the current block when we are
5601 // scanning for CSE expressions
5603 bool optIsCSEcandidate(GenTreePtr tree);
5605 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5607 bool lclNumIsTrueCSE(unsigned lclNum) const
5609 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5612 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5614 bool lclNumIsCSE(unsigned lclNum) const
5616 return lvaTable[lclNum].lvIsCSE;
5620 bool optConfigDisableCSE();
5621 bool optConfigDisableCSE2();
5623 void optOptimizeCSEs();
5625 #endif // FEATURE_ANYCSE
5633 unsigned ivaVar; // Variable we are interested in, or -1
5634 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5635 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5636 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5637 callInterf ivaMaskCall; // What kind of calls are there?
5640 static callInterf optCallInterf(GenTreeCall* call);
5643 // VN based copy propagation.
5644 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5645 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5646 LclNumToGenTreePtrStack;
5648 // Kill set to track variables with intervening definitions.
5649 VARSET_TP optCopyPropKillSet;
5651 // Copy propagation functions.
5652 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5653 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5654 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5655 bool optIsSsaLocal(GenTreePtr tree);
5656 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5657 void optVnCopyProp();
5659 /**************************************************************************
5660 * Early value propagation
5661 *************************************************************************/
5667 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5671 static unsigned GetHashCode(SSAName ssaNm)
5673 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5676 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5678 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5682 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5683 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5684 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5685 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5686 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5687 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5689 bool doesMethodHaveFatPointer()
5691 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5694 void setMethodHasFatPointer()
5696 optMethodFlags |= OMF_HAS_FATPOINTER;
5699 void clearMethodHasFatPointer()
5701 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5704 void addFatPointerCandidate(GenTreeCall* call)
5706 setMethodHasFatPointer();
5707 call->SetFatPointerCandidate();
5710 unsigned optMethodFlags;
5712 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5713 // No throughput diff was found with backward walk bound between 3-8.
5714 static const int optEarlyPropRecurBound = 5;
5716 enum class optPropKind
5724 bool gtIsVtableRef(GenTreePtr tree);
5725 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5726 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5727 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5728 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5729 bool optEarlyPropRewriteTree(GenTreePtr tree);
5730 bool optDoEarlyPropForBlock(BasicBlock* block);
5731 bool optDoEarlyPropForFunc();
5732 void optEarlyProp();
5733 void optFoldNullCheck(GenTreePtr tree);
5734 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5737 /**************************************************************************
5738 * Value/Assertion propagation
5739 *************************************************************************/
5741 // Data structures for assertion prop
5742 BitVecTraits* apTraits;
5745 enum optAssertionKind
5762 O1K_CONSTANT_LOOP_BND,
5783 optAssertionKind assertionKind;
5786 unsigned lclNum; // assigned to or property of this local var number
5794 struct AssertionDscOp1
5796 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5803 struct AssertionDscOp2
5805 optOp2Kind kind; // a const or copy assignment
5809 ssize_t iconVal; // integer
5810 unsigned iconFlags; // gtFlags
5812 struct Range // integer subrange
5826 bool IsCheckedBoundArithBound()
5828 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5830 bool IsCheckedBoundBound()
5832 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5834 bool IsConstantBound()
5836 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5837 op1.kind == O1K_CONSTANT_LOOP_BND);
5839 bool IsBoundsCheckNoThrow()
5841 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5844 bool IsCopyAssertion()
5846 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5849 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5851 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5852 a1->op2.kind == a2->op2.kind;
5855 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5857 if (kind == OAK_EQUAL)
5859 return kind2 == OAK_NOT_EQUAL;
5861 else if (kind == OAK_NOT_EQUAL)
5863 return kind2 == OAK_EQUAL;
5868 static ssize_t GetLowerBoundForIntegralType(var_types type)
5888 static ssize_t GetUpperBoundForIntegralType(var_types type)
5912 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5914 if (op1.kind != that->op1.kind)
5918 else if (op1.kind == O1K_ARR_BND)
5921 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5925 return ((vnBased && (op1.vn == that->op1.vn)) ||
5926 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5930 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5932 if (op2.kind != that->op2.kind)
5938 case O2K_IND_CNS_INT:
5940 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5942 case O2K_CONST_LONG:
5943 return (op2.lconVal == that->op2.lconVal);
5945 case O2K_CONST_DOUBLE:
5946 // exact match because of positive and negative zero.
5947 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5949 case O2K_LCLVAR_COPY:
5951 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5952 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5955 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5958 // we will return false
5962 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5968 bool Complementary(AssertionDsc* that, bool vnBased)
5970 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5971 HasSameOp2(that, vnBased);
5974 bool Equals(AssertionDsc* that, bool vnBased)
5976 if (assertionKind != that->assertionKind)
5980 else if (assertionKind == OAK_NO_THROW)
5982 assert(op2.kind == O2K_INVALID);
5983 return HasSameOp1(that, vnBased);
5987 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5993 static fgWalkPreFn optAddCopiesCallback;
5994 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5995 unsigned optAddCopyLclNum;
5996 GenTreePtr optAddCopyAsgnNode;
5998 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5999 bool optAssertionPropagated; // set to true if we modified the trees
6000 bool optAssertionPropagatedCurrentStmt;
6002 GenTreePtr optAssertionPropCurrentTree;
6004 AssertionIndex* optComplementaryAssertionMap;
6005 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6006 // using the value of a local var) for each local var
6007 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6008 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6009 AssertionIndex optMaxAssertionCount;
6012 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6013 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6014 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6015 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6016 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6017 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6019 AssertionIndex GetAssertionCount()
6021 return optAssertionCount;
6023 ASSERT_TP* bbJtrueAssertionOut;
6024 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6025 ValueNumToAssertsMap;
6026 ValueNumToAssertsMap* optValueNumToAsserts;
6028 // Assertion prop helpers.
6029 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6030 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6031 void optAssertionInit(bool isLocalProp);
6032 void optAssertionTraitsInit(AssertionIndex assertionCount);
6033 #if LOCAL_ASSERTION_PROP
6034 void optAssertionReset(AssertionIndex limit);
6035 void optAssertionRemove(AssertionIndex index);
6038 // Assertion prop data flow functions.
6039 void optAssertionPropMain();
6040 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6041 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6042 ASSERT_TP* optInitAssertionDataflowFlags();
6043 ASSERT_TP* optComputeAssertionGen();
6045 // Assertion Gen functions.
6046 void optAssertionGen(GenTreePtr tree);
6047 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6048 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6049 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6050 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6051 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6052 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6054 // Assertion creation functions.
6055 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6056 AssertionIndex optCreateAssertion(GenTreePtr op1,
6058 optAssertionKind assertionKind,
6059 AssertionDsc* assertion);
6060 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6062 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6063 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6064 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6066 void optPrintVnAssertionMapping();
6068 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6070 // Used for respective assertion propagations.
6071 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6072 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6073 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6074 bool optAssertionIsNonNull(GenTreePtr op,
6075 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6077 // Used for Relop propagation.
6078 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6079 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6080 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6082 // Assertion prop for lcl var functions.
6083 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6084 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6086 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6087 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6088 const GenTreePtr tree,
6089 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6090 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6092 // Assertion propagation functions.
6093 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6094 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6095 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6096 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6097 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6098 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6099 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6100 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6101 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6102 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6103 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6104 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6106 // Implied assertion functions.
6107 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6108 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6109 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6110 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6113 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6114 void optDebugCheckAssertion(AssertionDsc* assertion);
6115 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6117 void optAddCopies();
6118 #endif // ASSERTION_PROP
6120 /**************************************************************************
6122 *************************************************************************/
6125 struct LoopCloneVisitorInfo
6127 LoopCloneContext* context;
6130 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6131 : context(context), loopNum(loopNum), stmt(nullptr)
6136 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6137 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6138 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6139 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6140 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6141 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6142 void optObtainLoopCloningOpts(LoopCloneContext* context);
6143 bool optIsLoopClonable(unsigned loopInd);
6145 bool optCanCloneLoops();
6148 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6150 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6151 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6152 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6153 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6157 void optInsertLoopCloningStress(BasicBlock* head);
6159 #if COUNT_RANGECHECKS
6160 static unsigned optRangeChkRmv;
6161 static unsigned optRangeChkAll;
6170 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6175 RngChkDsc* rcdNextInBucket; // used by the hash table
6177 unsigned short rcdHashValue; // to make matching faster
6178 unsigned short rcdIndex; // 0..optRngChkCount-1
6180 GenTreePtr rcdTree; // the array index tree
6183 unsigned optRngChkCount;
6184 static const size_t optRngChkHashSize;
6186 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6187 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6189 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6192 bool optLoopsMarked;
6195 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6196 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6200 XX Does the register allocation and puts the remaining lclVars on the stack XX
6202 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6203 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6207 #ifndef LEGACY_BACKEND
6212 #else // LEGACY_BACKEND
6217 #endif // LEGACY_BACKEND
6219 #ifdef LEGACY_BACKEND
6221 void raAssignVars(); // register allocation
6222 #endif // LEGACY_BACKEND
6224 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6226 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6228 void raMarkStkVars();
6231 // Some things are used by both LSRA and regpredict allocators.
6233 FrameType rpFrameType;
6234 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6236 #ifdef LEGACY_BACKEND
6237 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6239 #endif // LEGACY_BACKEND
6241 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6243 #if FEATURE_FP_REGALLOC
6244 enum enumConfigRegisterFP
6246 CONFIG_REGISTER_FP_NONE = 0x0,
6247 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6248 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6249 CONFIG_REGISTER_FP_FULL = 0x3,
6251 enumConfigRegisterFP raConfigRegisterFP();
6252 #endif // FEATURE_FP_REGALLOC
6255 regMaskTP raConfigRestrictMaskFP();
6258 #ifndef LEGACY_BACKEND
6259 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6260 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6261 #else // LEGACY_BACKEND
6262 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6263 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6264 bool raNewBlocks; // True is we added killing blocks for FPU registers
6265 unsigned rpPasses; // Number of passes made by the register predicter
6266 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6267 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6268 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6269 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6270 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6271 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6272 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6273 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6274 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6275 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6276 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6277 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6279 bool rpRegAllocDone; // Set to true after we have completed register allocation
6281 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6283 void raSetupArgMasks(RegState* r);
6285 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6287 void raDumpVarIntf(); // Dump the variable to variable interference graph
6288 void raDumpRegIntf(); // Dump the variable to register interference graph
6290 void raAdjustVarIntf();
6292 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6294 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6296 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6297 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6299 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6301 static fgWalkPreFn rpMarkRegIntf;
6303 regMaskTP rpPredictAddressMode(
6304 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6306 void rpPredictRefAssign(unsigned lclNum);
6308 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6310 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6312 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6314 void rpPredictRegUse(); // Entry point
6316 unsigned raPredictTreeRegUse(GenTreePtr tree);
6317 unsigned raPredictListRegUse(GenTreePtr list);
6319 void raSetRegVarOrder(var_types regType,
6320 regNumber* customVarOrder,
6321 unsigned* customVarOrderSize,
6323 regMaskTP avoidReg);
6325 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6326 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6327 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6328 void raAddToStkPredict(unsigned val)
6330 unsigned newStkPredict = rpStkPredict + val;
6331 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6332 rpStkPredict = UINT_MAX - 1;
6334 rpStkPredict = newStkPredict;
6338 #if !FEATURE_FP_REGALLOC
6339 void raDispFPlifeInfo();
6343 regMaskTP genReturnRegForTree(GenTreePtr tree);
6344 #endif // LEGACY_BACKEND
6346 /* raIsVarargsStackArg is called by raMaskStkVars and by
6347 lvaSortByRefCount. It identifies the special case
6348 where a varargs function has a parameter passed on the
6349 stack, other than the special varargs handle. Such parameters
6350 require special treatment, because they cannot be tracked
6351 by the GC (their offsets in the stack are not known
6355 bool raIsVarargsStackArg(unsigned lclNum)
6359 LclVarDsc* varDsc = &lvaTable[lclNum];
6361 assert(varDsc->lvIsParam);
6363 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6365 #else // _TARGET_X86_
6369 #endif // _TARGET_X86_
6372 #ifdef LEGACY_BACKEND
6373 // Records the current prediction, if it's better than any previous recorded prediction.
6374 void rpRecordPrediction();
6375 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6376 void rpUseRecordedPredictionIfBetter();
6378 // Data members used in the methods above.
6379 unsigned rpBestRecordedStkPredict;
6380 struct VarRegPrediction
6382 bool m_isEnregistered;
6383 regNumberSmall m_regNum;
6384 regNumberSmall m_otherReg;
6386 VarRegPrediction* rpBestRecordedPrediction;
6387 #endif // LEGACY_BACKEND
6390 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6391 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6395 XX Get to the class and method info from the Execution Engine given XX
6396 XX tokens for the class and method XX
6398 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6399 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6403 /* These are the different addressing modes used to access a local var.
6404 * The JIT has to report the location of the locals back to the EE
6405 * for debugging purposes.
6411 VLT_REG_BYREF, // this type is currently only used for value types on X64
6414 VLT_STK_BYREF, // this type is currently only used for value types on X64
6428 siVarLocType vlType;
6431 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6433 // VLT_REG_BYREF -- the specified register contains the address of the variable
6441 // VLT_STK -- Any 32 bit value which is on the stack
6442 // eg. [ESP+0x20], or [EBP-0x28]
6443 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6444 // eg. mov EAX, [ESP+0x20]; [EAX]
6448 regNumber vlsBaseReg;
6449 NATIVE_OFFSET vlsOffset;
6452 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6461 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6462 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6470 regNumber vlrssBaseReg;
6471 NATIVE_OFFSET vlrssOffset;
6475 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6476 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6482 regNumber vlsrsBaseReg;
6483 NATIVE_OFFSET vlsrsOffset;
6489 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6490 // eg 2 DWords at [ESP+0x10]
6494 regNumber vls2BaseReg;
6495 NATIVE_OFFSET vls2Offset;
6498 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6499 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6506 // VLT_FIXED_VA -- fixed argument of a varargs function.
6507 // The argument location depends on the size of the variable
6508 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6509 // location of the first arg. This argument can then be accessed
6510 // relative to the position of the first arg
6514 unsigned vlfvOffset;
6521 void* rpValue; // pointer to the in-process
6522 // location of the value.
6528 bool vlIsInReg(regNumber reg);
6529 bool vlIsOnStk(regNumber reg, signed offset);
6532 /*************************************************************************/
6537 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6538 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6539 CORINFO_CALLINFO_FLAGS flags,
6540 CORINFO_CALL_INFO* pResult);
6541 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6543 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6544 CORINFO_ACCESS_FLAGS flags,
6545 CORINFO_FIELD_INFO* pResult);
6549 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6551 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6553 bool IsSuperPMIException(unsigned code)
6555 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6557 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6558 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6559 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6560 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6561 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6562 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6563 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6564 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6568 case EXCEPTIONCODE_DebugBreakorAV:
6569 case EXCEPTIONCODE_MC:
6570 case EXCEPTIONCODE_LWM:
6571 case EXCEPTIONCODE_SASM:
6572 case EXCEPTIONCODE_SSYM:
6573 case EXCEPTIONCODE_CALLUTILS:
6574 case EXCEPTIONCODE_TYPEUTILS:
6575 case EXCEPTIONCODE_ASSERT:
6582 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6583 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6585 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6586 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6589 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6590 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6591 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6593 // VOM info, method sigs
6595 void eeGetSig(unsigned sigTok,
6596 CORINFO_MODULE_HANDLE scope,
6597 CORINFO_CONTEXT_HANDLE context,
6598 CORINFO_SIG_INFO* retSig);
6600 void eeGetCallSiteSig(unsigned sigTok,
6601 CORINFO_MODULE_HANDLE scope,
6602 CORINFO_CONTEXT_HANDLE context,
6603 CORINFO_SIG_INFO* retSig);
6605 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6607 // Method entry-points, instrs
6609 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6611 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6613 CORINFO_EE_INFO eeInfo;
6614 bool eeInfoInitialized;
6616 CORINFO_EE_INFO* eeGetEEInfo();
6618 // Gets the offset of a SDArray's first element
6619 unsigned eeGetArrayDataOffset(var_types type);
6620 // Gets the offset of a MDArray's first element
6621 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6623 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6625 // Returns the page size for the target machine as reported by the EE.
6626 inline size_t eeGetPageSize()
6628 return eeGetEEInfo()->osPageSize;
6631 // Returns the frame size at which we will generate a loop to probe the stack.
6632 inline size_t getVeryLargeFrameSize()
6635 // The looping probe code is 40 bytes, whereas the straight-line probing for
6636 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6637 // or greater, to generate smaller code.
6638 return 2 * eeGetPageSize();
6640 return 3 * eeGetPageSize();
6644 //------------------------------------------------------------------------
6645 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6647 // It represents Abi and target specific registers for the parameter.
6649 class VirtualStubParamInfo
6652 VirtualStubParamInfo(bool isCoreRTABI)
6654 #if defined(_TARGET_X86_)
6657 #elif defined(_TARGET_AMD64_)
6668 #elif defined(_TARGET_ARM_)
6671 #elif defined(_TARGET_ARM64_)
6675 #error Unsupported or unset target architecture
6678 #ifdef LEGACY_BACKEND
6679 #if defined(_TARGET_X86_)
6680 predict = PREDICT_REG_EAX;
6681 #elif defined(_TARGET_ARM_)
6682 predict = PREDICT_REG_R4;
6684 #error Unsupported or unset target architecture
6686 #endif // LEGACY_BACKEND
6689 regNumber GetReg() const
6694 _regMask_enum GetRegMask() const
6699 #ifdef LEGACY_BACKEND
6700 rpPredictReg GetPredict() const
6708 _regMask_enum regMask;
6710 #ifdef LEGACY_BACKEND
6711 rpPredictReg predict;
6715 VirtualStubParamInfo* virtualStubParamInfo;
6717 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6719 return eeGetEEInfo()->targetAbi == abi;
6722 inline bool generateCFIUnwindCodes()
6724 #ifdef UNIX_AMD64_ABI
6725 return IsTargetAbi(CORINFO_CORERT_ABI);
6733 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6735 // Debugging support - Line number info
6737 void eeGetStmtOffsets();
6739 unsigned eeBoundariesCount;
6741 struct boundariesDsc
6743 UNATIVE_OFFSET nativeIP;
6745 unsigned sourceReason;
6746 } * eeBoundaries; // Boundaries to report to EE
6747 void eeSetLIcount(unsigned count);
6748 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6752 static void eeDispILOffs(IL_OFFSET offs);
6753 static void eeDispLineInfo(const boundariesDsc* line);
6754 void eeDispLineInfos();
6757 // Debugging support - Local var info
6761 unsigned eeVarsCount;
6763 struct VarResultInfo
6765 UNATIVE_OFFSET startOffset;
6766 UNATIVE_OFFSET endOffset;
6770 void eeSetLVcount(unsigned count);
6771 void eeSetLVinfo(unsigned which,
6772 UNATIVE_OFFSET startOffs,
6773 UNATIVE_OFFSET length,
6778 const siVarLoc& loc);
6782 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6783 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6786 // ICorJitInfo wrappers
6788 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6790 void eeAllocUnwindInfo(BYTE* pHotCode,
6796 CorJitFuncKind funcKind);
6798 void eeSetEHcount(unsigned cEH);
6800 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6802 WORD eeGetRelocTypeHint(void* target);
6804 // ICorStaticInfo wrapper functions
6806 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6808 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6810 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6813 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6814 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6815 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6816 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6818 template <typename ParamType>
6819 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6821 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6824 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6826 // Utility functions
6828 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6831 const wchar_t* eeGetCPString(size_t stringHandle);
6834 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6836 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6837 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6839 static fgWalkPreFn CountSharedStaticHelper;
6840 static bool IsSharedStaticHelper(GenTreePtr tree);
6841 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6842 static bool IsGcSafePoint(GenTreePtr tree);
6844 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6845 // returns true/false if 'field' is a Jit Data offset
6846 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6847 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6848 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6850 /*****************************************************************************/
6855 enum TEMP_USAGE_TYPE
6861 static var_types tmpNormalizeType(var_types type);
6862 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6863 void tmpRlsTemp(TempDsc* temp);
6864 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6867 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6868 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6872 bool tmpAllFree() const;
6875 #ifndef LEGACY_BACKEND
6876 void tmpPreAllocateTemps(var_types type, unsigned count);
6877 #endif // !LEGACY_BACKEND
6880 #ifdef LEGACY_BACKEND
6881 unsigned tmpIntSpillMax; // number of int-sized spill temps
6882 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6883 #endif // LEGACY_BACKEND
6885 unsigned tmpCount; // Number of temps
6886 unsigned tmpSize; // Size of all the temps
6889 // Used by RegSet::rsSpillChk()
6890 unsigned tmpGetCount; // Temps which haven't been released yet
6893 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6895 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6896 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6899 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6900 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6904 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6905 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6909 CodeGenInterface* codeGen;
6911 // The following holds information about instr offsets in terms of generated code.
6915 IPmappingDsc* ipmdNext; // next line# record
6916 IL_OFFSETX ipmdILoffsx; // the instr offset
6917 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6918 bool ipmdIsLabel; // Can this code be a branch label?
6921 // Record the instr offset mapping to the generated code
6923 IPmappingDsc* genIPmappingList;
6924 IPmappingDsc* genIPmappingLast;
6926 // Managed RetVal - A side hash table meant to record the mapping from a
6927 // GT_CALL node to its IL offset. This info is used to emit sequence points
6928 // that can be used by debugger to determine the native offset at which the
6929 // managed RetVal will be available.
6931 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6932 // favor of a side table for two reasons: 1) We need IL offset for only those
6933 // GT_CALL nodes (created during importation) that correspond to an IL call and
6934 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6935 // structure and IL offset is needed only when generating debuggable code. Therefore
6936 // it is desirable to avoid memory size penalty in retail scenarios.
6937 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6938 CallSiteILOffsetTable;
6939 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6941 unsigned genReturnLocal; // Local number for the return value when applicable.
6942 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6944 // The following properties are part of CodeGenContext. Getters are provided here for
6945 // convenience and backward compatibility, but the properties can only be set by invoking
6946 // the setter on CodeGenContext directly.
6948 __declspec(property(get = getEmitter)) emitter* genEmitter;
6949 emitter* getEmitter()
6951 return codeGen->getEmitter();
6954 const bool isFramePointerUsed()
6956 return codeGen->isFramePointerUsed();
6959 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6960 bool getInterruptible()
6962 return codeGen->genInterruptible;
6964 void setInterruptible(bool value)
6966 codeGen->setInterruptible(value);
6970 const bool genDoubleAlign()
6972 return codeGen->doDoubleAlign();
6974 DWORD getCanDoubleAlign();
6975 bool shouldDoubleAlign(unsigned refCntStk,
6977 unsigned refCntWtdReg,
6978 unsigned refCntStkParam,
6979 unsigned refCntWtdStkDbl);
6980 #endif // DOUBLE_ALIGN
6982 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6983 bool getFullPtrRegMap()
6985 return codeGen->genFullPtrRegMap;
6987 void setFullPtrRegMap(bool value)
6989 codeGen->setFullPtrRegMap(value);
6992 // Things that MAY belong either in CodeGen or CodeGenContext
6994 #if FEATURE_EH_FUNCLETS
6995 FuncInfoDsc* compFuncInfos;
6996 unsigned short compCurrFuncIdx;
6997 unsigned short compFuncInfoCount;
6999 unsigned short compFuncCount()
7001 assert(fgFuncletsCreated);
7002 return compFuncInfoCount;
7005 #else // !FEATURE_EH_FUNCLETS
7007 // This is a no-op when there are no funclets!
7008 void genUpdateCurrentFunclet(BasicBlock* block)
7013 FuncInfoDsc compFuncInfoRoot;
7015 static const unsigned compCurrFuncIdx = 0;
7017 unsigned short compFuncCount()
7022 #endif // !FEATURE_EH_FUNCLETS
7024 FuncInfoDsc* funCurrentFunc();
7025 void funSetCurrentFunc(unsigned funcIdx);
7026 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7027 unsigned int funGetFuncIdx(BasicBlock* block);
7031 VARSET_TP compCurLife; // current live variables
7032 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7034 template <bool ForCodeGen>
7035 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7037 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7039 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7042 template <bool ForCodeGen>
7043 void compUpdateLife(GenTreePtr tree);
7045 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7046 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7047 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7048 template <bool ForCodeGen>
7049 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7051 template <bool ForCodeGen>
7052 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7054 // Gets a register mask that represent the kill set for a helper call since
7055 // not all JIT Helper calls follow the standard ABI on the target architecture.
7056 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7058 // Gets a register mask that represent the kill set for a NoGC helper call.
7059 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7062 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7063 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7064 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7065 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7066 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7067 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7068 #endif // _TARGET_ARM_
7070 // 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
7072 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7074 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7075 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7076 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7077 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7078 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7079 // for the tracked var indices of the field vars, as in a live var set).
7080 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7082 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7084 if (m_promotedStructDeathVars == nullptr)
7086 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7088 return m_promotedStructDeathVars;
7092 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7093 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7097 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7098 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7101 #if !defined(__GNUC__)
7102 #pragma region Unwind information
7107 // Infrastructure functions: start/stop/reserve/emit.
7110 void unwindBegProlog();
7111 void unwindEndProlog();
7112 void unwindBegEpilog();
7113 void unwindEndEpilog();
7114 void unwindReserve();
7115 void unwindEmit(void* pHotCode, void* pColdCode);
7118 // Specific unwind information functions: called by code generation to indicate a particular
7119 // prolog or epilog unwindable instruction has been generated.
7122 void unwindPush(regNumber reg);
7123 void unwindAllocStack(unsigned size);
7124 void unwindSetFrameReg(regNumber reg, unsigned offset);
7125 void unwindSaveReg(regNumber reg, unsigned offset);
7127 #if defined(_TARGET_ARM_)
7128 void unwindPushMaskInt(regMaskTP mask);
7129 void unwindPushMaskFloat(regMaskTP mask);
7130 void unwindPopMaskInt(regMaskTP mask);
7131 void unwindPopMaskFloat(regMaskTP mask);
7132 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7133 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7134 // called via unwindPadding().
7135 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7136 // instruction and the current location.
7137 #endif // _TARGET_ARM_
7139 #if defined(_TARGET_ARM64_)
7141 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7142 // instruction and the current location.
7143 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7144 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7145 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7146 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7147 void unwindSaveNext(); // unwind code: save_next
7148 void unwindReturn(regNumber reg); // ret lr
7149 #endif // defined(_TARGET_ARM64_)
7152 // Private "helper" functions for the unwind implementation.
7156 #if FEATURE_EH_FUNCLETS
7157 void unwindGetFuncLocations(FuncInfoDsc* func,
7158 bool getHotSectionData,
7159 /* OUT */ emitLocation** ppStartLoc,
7160 /* OUT */ emitLocation** ppEndLoc);
7161 #endif // FEATURE_EH_FUNCLETS
7163 void unwindReserveFunc(FuncInfoDsc* func);
7164 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7166 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7168 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7169 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7171 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7173 #if defined(_TARGET_AMD64_)
7175 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7177 void unwindBegPrologWindows();
7178 void unwindPushWindows(regNumber reg);
7179 void unwindAllocStackWindows(unsigned size);
7180 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7181 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7183 #ifdef UNIX_AMD64_ABI
7184 void unwindBegPrologCFI();
7185 void unwindPushCFI(regNumber reg);
7186 void unwindAllocStackCFI(unsigned size);
7187 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7188 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7189 int mapRegNumToDwarfReg(regNumber reg);
7190 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7191 #endif // UNIX_AMD64_ABI
7192 #elif defined(_TARGET_ARM_)
7194 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7195 void unwindPushPopMaskFloat(regMaskTP mask);
7196 void unwindSplit(FuncInfoDsc* func);
7198 #endif // _TARGET_ARM_
7200 #if !defined(__GNUC__)
7201 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7205 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7206 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7210 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7211 XX that contains the distinguished, well-known SIMD type definitions). XX
7213 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7214 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7217 // Get highest available instruction set for floating point codegen
7218 InstructionSet getFloatingPointInstructionSet()
7220 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7223 return InstructionSet_AVX;
7228 return InstructionSet_SSE3_4;
7232 assert(canUseSSE2());
7233 return InstructionSet_SSE2;
7235 assert(!"getFPInstructionSet() is not implemented for target arch");
7237 return InstructionSet_NONE;
7241 // Get highest available instruction set for SIMD codegen
7242 InstructionSet getSIMDInstructionSet()
7244 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7245 return getFloatingPointInstructionSet();
7247 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7249 return InstructionSet_NONE;
7255 // Should we support SIMD intrinsics?
7258 // Have we identified any SIMD types?
7259 // This is currently used by struct promotion to avoid getting type information for a struct
7260 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7262 bool _usesSIMDTypes;
7263 bool usesSIMDTypes()
7265 return _usesSIMDTypes;
7267 void setUsesSIMDTypes(bool value)
7269 _usesSIMDTypes = value;
7272 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7273 // that require indexed access to the individual fields of the vector, which is not well supported
7274 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7275 unsigned lvaSIMDInitTempVarNum;
7278 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7279 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7280 CORINFO_CLASS_HANDLE SIMDIntHandle;
7281 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7282 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7283 CORINFO_CLASS_HANDLE SIMDShortHandle;
7284 CORINFO_CLASS_HANDLE SIMDByteHandle;
7285 CORINFO_CLASS_HANDLE SIMDLongHandle;
7286 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7287 CORINFO_CLASS_HANDLE SIMDULongHandle;
7288 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7289 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7290 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7291 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7293 // Get the handle for a SIMD type.
7294 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7296 if (simdBaseType == TYP_FLOAT)
7301 return SIMDVector2Handle;
7303 return SIMDVector3Handle;
7305 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7307 return SIMDVector4Handle;
7316 assert(simdType == getSIMDVectorType());
7317 switch (simdBaseType)
7320 return SIMDFloatHandle;
7322 return SIMDDoubleHandle;
7324 return SIMDIntHandle;
7326 return SIMDUShortHandle;
7328 return SIMDUShortHandle;
7330 return SIMDUByteHandle;
7332 return SIMDShortHandle;
7334 return SIMDByteHandle;
7336 return SIMDLongHandle;
7338 return SIMDUIntHandle;
7340 return SIMDULongHandle;
7342 assert(!"Didn't find a class handle for simdType");
7344 return NO_CLASS_HANDLE;
7348 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7349 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7350 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7352 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7353 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7354 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7355 bool isSIMDTypeLocal(GenTree* tree)
7357 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7360 // Returns true if the type of the tree is a byref of TYP_SIMD
7361 bool isAddrOfSIMDType(GenTree* tree)
7363 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7365 switch (tree->OperGet())
7368 return varTypeIsSIMD(tree->gtGetOp1());
7370 case GT_LCL_VAR_ADDR:
7371 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7374 return isSIMDTypeLocal(tree);
7381 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7383 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7384 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7385 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7388 // Returns base type of a TYP_SIMD local.
7389 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7390 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7392 if (isSIMDTypeLocal(tree))
7394 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7400 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7402 return info.compCompHnd->isInSIMDModule(clsHnd);
7405 bool isSIMDClass(typeInfo* pTypeInfo)
7407 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7410 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7411 // if it is not a SIMD type or is an unsupported base type.
7412 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7414 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7416 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7419 // Get SIMD Intrinsic info given the method handle.
7420 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7421 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7422 CORINFO_METHOD_HANDLE methodHnd,
7423 CORINFO_SIG_INFO* sig,
7426 var_types* baseType,
7427 unsigned* sizeBytes);
7429 // Pops and returns GenTree node from importers type stack.
7430 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7431 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7433 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7434 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7436 // Creates a GT_SIMD tree for Select operation
7437 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7439 unsigned simdVectorSize,
7444 // Creates a GT_SIMD tree for Min/Max operation
7445 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7446 CORINFO_CLASS_HANDLE typeHnd,
7448 unsigned simdVectorSize,
7452 // Transforms operands and returns the SIMD intrinsic to be applied on
7453 // transformed operands to obtain given relop result.
7454 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7455 CORINFO_CLASS_HANDLE typeHnd,
7456 unsigned simdVectorSize,
7457 var_types* baseType,
7461 // Creates a GT_SIMD tree for Abs intrinsic.
7462 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7464 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7465 // Transforms operands and returns the SIMD intrinsic to be applied on
7466 // transformed operands to obtain == comparison result.
7467 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7468 unsigned simdVectorSize,
7472 // Transforms operands and returns the SIMD intrinsic to be applied on
7473 // transformed operands to obtain > comparison result.
7474 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7475 unsigned simdVectorSize,
7479 // Transforms operands and returns the SIMD intrinsic to be applied on
7480 // transformed operands to obtain >= comparison result.
7481 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7482 unsigned simdVectorSize,
7486 // Transforms operands and returns the SIMD intrinsic to be applied on
7487 // transformed operands to obtain >= comparison result in case of int32
7488 // and small int base type vectors.
7489 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7490 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7491 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7493 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7494 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7495 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7496 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7497 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7499 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7500 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7501 GenTreePtr newobjThis,
7502 CORINFO_CLASS_HANDLE clsHnd,
7503 CORINFO_METHOD_HANDLE method,
7504 CORINFO_SIG_INFO* sig,
7507 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7509 // Whether SIMD vector occupies part of SIMD register.
7510 // SSE2: vector2f/3f are considered sub register SIMD types.
7511 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7512 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7514 unsigned sizeBytes = 0;
7515 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7516 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7519 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7521 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7524 // Get the type for the hardware SIMD vector.
7525 // This is the maximum SIMD type supported for this target.
7526 var_types getSIMDVectorType()
7528 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7535 assert(canUseSSE2());
7539 assert(!"getSIMDVectorType() unimplemented on target arch");
7544 // Get the size of the SIMD type in bytes
7545 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7547 unsigned sizeBytes = 0;
7548 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7552 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7553 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7555 // Get the the number of elements of basetype of SIMD vector given by its type handle
7556 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7558 // Get preferred alignment of SIMD type.
7559 int getSIMDTypeAlignment(var_types simdType);
7561 // Get the number of bytes in a SIMD Vector for the current compilation.
7562 unsigned getSIMDVectorRegisterByteLength()
7564 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7567 return YMM_REGSIZE_BYTES;
7571 assert(canUseSSE2());
7572 return XMM_REGSIZE_BYTES;
7575 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7580 // The minimum and maximum possible number of bytes in a SIMD vector.
7581 unsigned int maxSIMDStructBytes()
7583 return getSIMDVectorRegisterByteLength();
7585 unsigned int minSIMDStructBytes()
7587 return emitTypeSize(TYP_SIMD8);
7590 #ifdef FEATURE_AVX_SUPPORT
7591 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7592 static const unsigned maxPossibleSIMDStructBytes = 32;
7593 #else // !FEATURE_AVX_SUPPORT
7594 static const unsigned maxPossibleSIMDStructBytes = 16;
7595 #endif // !FEATURE_AVX_SUPPORT
7597 // Returns the codegen type for a given SIMD size.
7598 var_types getSIMDTypeForSize(unsigned size)
7600 var_types simdType = TYP_UNDEF;
7603 simdType = TYP_SIMD8;
7605 else if (size == 12)
7607 simdType = TYP_SIMD12;
7609 else if (size == 16)
7611 simdType = TYP_SIMD16;
7613 #ifdef FEATURE_AVX_SUPPORT
7614 else if (size == 32)
7616 simdType = TYP_SIMD32;
7618 #endif // FEATURE_AVX_SUPPORT
7621 noway_assert(!"Unexpected size for SIMD type");
7626 unsigned getSIMDInitTempVarNum()
7628 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7630 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7631 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7633 return lvaSIMDInitTempVarNum;
7636 #endif // FEATURE_SIMD
7639 //------------------------------------------------------------------------
7640 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7642 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7643 // candidate for enregistration.
7645 unsigned largestEnregisterableStructSize()
7648 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7649 if (vectorRegSize > TARGET_POINTER_SIZE)
7651 return vectorRegSize;
7654 #endif // FEATURE_SIMD
7656 return TARGET_POINTER_SIZE;
7661 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7662 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7663 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7665 // Is this var is of type simd struct?
7666 bool lclVarIsSIMDType(unsigned varNum)
7668 LclVarDsc* varDsc = lvaTable + varNum;
7669 return varDsc->lvIsSIMDType();
7672 // Is this Local node a SIMD local?
7673 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7675 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7678 // Returns true if the TYP_SIMD locals on stack are aligned at their
7679 // preferred byte boundary specified by getSIMDTypeAlignment().
7681 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7682 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7683 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7684 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7685 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7686 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7687 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7690 bool isSIMDTypeLocalAligned(unsigned varNum)
7692 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7693 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7696 int off = lvaFrameAddress(varNum, &ebpBased);
7697 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7698 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7699 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7702 #endif // FEATURE_SIMD
7707 // Whether SSE2 is available
7708 bool canUseSSE2() const
7710 #ifdef _TARGET_XARCH_
7711 return opts.compCanUseSSE2;
7717 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7718 bool CanUseSSE3_4() const
7720 #ifdef _TARGET_XARCH_
7721 return opts.compCanUseSSE3_4;
7727 bool canUseAVX() const
7729 #ifdef FEATURE_AVX_SUPPORT
7730 return opts.compCanUseAVX;
7737 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7742 XX Generic info about the compilation and the method being compiled. XX
7743 XX It is responsible for driving the other phases. XX
7744 XX It is also responsible for all the memory management. XX
7746 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7747 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7751 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7753 InlineResult* compInlineResult; // The result of importing the inlinee method.
7755 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7756 bool compJmpOpUsed; // Does the method do a JMP
7757 bool compLongUsed; // Does the method use TYP_LONG
7758 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7759 bool compTailCallUsed; // Does the method do a tailcall
7760 bool compLocallocUsed; // Does the method use localloc.
7761 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7762 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7763 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7765 // NOTE: These values are only reliable after
7766 // the importing is completely finished.
7768 #ifdef LEGACY_BACKEND
7769 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7770 // we can iterate over these efficiently.
7773 #if CPU_USES_BLOCK_MOVE
7774 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7778 // State information - which phases have completed?
7779 // These are kept together for easy discoverability
7781 bool bRangeAllowStress;
7782 bool compCodeGenDone;
7783 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7784 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7785 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7786 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7789 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7790 bool fgLocalVarLivenessChanged;
7792 bool compStackProbePrologDone;
7794 #ifndef LEGACY_BACKEND
7796 #endif // !LEGACY_BACKEND
7797 bool compRationalIRForm;
7799 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7801 bool compGeneratingProlog;
7802 bool compGeneratingEpilog;
7803 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7804 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7805 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7806 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7807 bool getNeedsGSSecurityCookie() const
7809 return compNeedsGSSecurityCookie;
7811 void setNeedsGSSecurityCookie()
7813 compNeedsGSSecurityCookie = true;
7816 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7817 // frame layout calculations, this is the level we are currently
7820 //---------------------------- JITing options -----------------------------
7833 JitFlags* jitFlags; // all flags passed from the EE
7834 unsigned compFlags; // method attributes
7836 codeOptimize compCodeOpt; // what type of code optimizations
7840 #ifdef _TARGET_XARCH_
7841 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7842 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7844 #ifdef FEATURE_AVX_SUPPORT
7845 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7846 #endif // FEATURE_AVX_SUPPORT
7847 #endif // _TARGET_XARCH_
7849 // optimize maximally and/or favor speed over size?
7851 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7852 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7853 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7854 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7855 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7857 // Maximun number of locals before turning off the inlining
7858 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7861 unsigned instrCount;
7862 unsigned lvRefCount;
7863 bool compMinOptsIsSet;
7865 bool compMinOptsIsUsed;
7867 inline bool MinOpts()
7869 assert(compMinOptsIsSet);
7870 compMinOptsIsUsed = true;
7873 inline bool IsMinOptsSet()
7875 return compMinOptsIsSet;
7878 inline bool MinOpts()
7882 inline bool IsMinOptsSet()
7884 return compMinOptsIsSet;
7887 inline void SetMinOpts(bool val)
7889 assert(!compMinOptsIsUsed);
7890 assert(!compMinOptsIsSet || (compMinOpts == val));
7892 compMinOptsIsSet = true;
7895 // true if the CLFLG_* for an optimization is set.
7896 inline bool OptEnabled(unsigned optFlag)
7898 return !!(compFlags & optFlag);
7901 #ifdef FEATURE_READYTORUN_COMPILER
7902 inline bool IsReadyToRun()
7904 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7907 inline bool IsReadyToRun()
7913 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7914 // PInvoke transitions inline (e.g. when targeting CoreRT).
7915 inline bool ShouldUsePInvokeHelpers()
7917 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7920 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7922 inline bool IsReversePInvoke()
7924 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7927 // true if we must generate code compatible with JIT32 quirks
7928 inline bool IsJit32Compat()
7930 #if defined(_TARGET_X86_)
7931 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7937 // true if we must generate code compatible with Jit64 quirks
7938 inline bool IsJit64Compat()
7940 #if defined(_TARGET_AMD64_)
7941 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7942 #elif !defined(FEATURE_CORECLR)
7949 bool compScopeInfo; // Generate the LocalVar info ?
7950 bool compDbgCode; // Generate debugger-friendly code?
7951 bool compDbgInfo; // Gather debugging info?
7954 #ifdef PROFILING_SUPPORTED
7955 bool compNoPInvokeInlineCB;
7957 static const bool compNoPInvokeInlineCB;
7961 bool compGcChecks; // Check arguments and return values to ensure they are sane
7962 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7963 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7967 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7968 // to be allocated on the stack.
7969 // It will be set to true in the following cases:
7970 // 1. When the method being compiled has a declarative security
7971 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7972 // This is also the case when we inject a prolog and epilog in the method.
7974 // 2. When the method being compiled has imperative security (i.e. the method
7975 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7977 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7979 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7980 // which gets reported as a GC root to stackwalker.
7981 // (See also ICodeManager::GetAddrOfSecurityObject.)
7986 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7987 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7991 #ifdef UNIX_AMD64_ABI
7992 // This flag is indicating if there is a need to align the frame.
7993 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7994 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7995 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7996 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7997 // there are calls and making sure the frame alignment logic is executed.
7998 bool compNeedToAlignFrame;
7999 #endif // UNIX_AMD64_ABI
8001 bool compProcedureSplitting; // Separate cold code from hot code
8003 bool genFPorder; // Preserve FP order (operations are non-commutative)
8004 bool genFPopt; // Can we do frame-pointer-omission optimization?
8005 bool altJit; // True if we are an altjit and are compiling this method
8008 bool optRepeat; // Repeat optimizer phases k times
8012 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8013 bool dspCode; // Display native code generated
8014 bool dspEHTable; // Display the EH table reported to the VM
8015 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8016 bool dspEmit; // Display emitter output
8017 bool dspLines; // Display source-code lines intermixed with native code output
8018 bool dmpHex; // Display raw bytes in hex of native code output
8019 bool varNames; // Display variables names in native code output
8020 bool disAsm; // Display native code as it is generated
8021 bool disAsmSpilled; // Display native code when any register spilling occurs
8022 bool disDiffable; // Makes the Disassembly code 'diff-able'
8023 bool disAsm2; // Display native code after it is generated using external disassembler
8024 bool dspOrder; // Display names of each of the methods that we ngen/jit
8025 bool dspUnwind; // Display the unwind info output
8026 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8027 bool compLongAddress; // Force using large pseudo instructions for long address
8028 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8029 bool dspGCtbls; // Display the GC tables
8033 bool doLateDisasm; // Run the late disassembler
8034 #endif // LATE_DISASM
8036 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8037 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8038 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8039 static const bool dspGCtbls = true;
8042 // We need stack probes to guarantee that we won't trigger a stack overflow
8043 // when calling unmanaged code until they get a chance to set up a frame, because
8044 // the EE will have no idea where it is.
8046 // We will only be doing this currently for hosted environments. Unfortunately
8047 // we need to take care of stubs, so potentially, we will have to do the probes
8048 // for any call. We have a plan for not needing for stubs though
8049 bool compNeedStackProbes;
8051 #ifdef PROFILING_SUPPORTED
8052 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8053 // This option helps make the JIT behave as if it is running under a profiler.
8054 bool compJitELTHookEnabled;
8055 #endif // PROFILING_SUPPORTED
8057 #if FEATURE_TAILCALL_OPT
8058 // Whether opportunistic or implicit tail call optimization is enabled.
8059 bool compTailCallOpt;
8060 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8061 bool compTailCallLoopOpt;
8065 static const bool compUseSoftFP = true;
8066 #else // !ARM_SOFTFP
8067 static const bool compUseSoftFP = false;
8070 GCPollType compGCPollType;
8074 static bool s_pAltJitExcludeAssembliesListInitialized;
8075 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8080 template <typename T>
8083 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8086 template <typename T>
8089 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8092 static int dspTreeID(GenTree* tree)
8094 return tree->gtTreeID;
8096 static void printTreeID(GenTree* tree)
8098 if (tree == nullptr)
8104 printf("[%06d]", dspTreeID(tree));
8111 #define STRESS_MODES \
8115 /* "Variations" stress areas which we try to mix up with each other. */ \
8116 /* These should not be exhaustively used as they might */ \
8117 /* hide/trivialize other areas */ \
8120 STRESS_MODE(DBL_ALN) \
8121 STRESS_MODE(LCL_FLDS) \
8122 STRESS_MODE(UNROLL_LOOPS) \
8123 STRESS_MODE(MAKE_CSE) \
8124 STRESS_MODE(LEGACY_INLINE) \
8125 STRESS_MODE(CLONE_EXPR) \
8126 STRESS_MODE(USE_FCOMI) \
8127 STRESS_MODE(USE_CMOV) \
8129 STRESS_MODE(BB_PROFILE) \
8130 STRESS_MODE(OPT_BOOLS_GC) \
8131 STRESS_MODE(REMORPH_TREES) \
8132 STRESS_MODE(64RSLT_MUL) \
8133 STRESS_MODE(DO_WHILE_LOOPS) \
8134 STRESS_MODE(MIN_OPTS) \
8135 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8136 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8137 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8138 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8139 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8140 STRESS_MODE(NULL_OBJECT_CHECK) \
8141 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8142 STRESS_MODE(RANDOM_INLINE) \
8143 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8144 STRESS_MODE(GENERIC_VARN) \
8146 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8148 STRESS_MODE(COUNT_VARN) \
8150 /* "Check" stress areas that can be exhaustively used if we */ \
8151 /* dont care about performance at all */ \
8153 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8154 STRESS_MODE(CHK_FLOW_UPDATE) \
8155 STRESS_MODE(EMITTER) \
8156 STRESS_MODE(CHK_REIMPORT) \
8157 STRESS_MODE(FLATFP) \
8158 STRESS_MODE(GENERIC_CHECK) \
8163 #define STRESS_MODE(mode) STRESS_##mode,
8170 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8171 BYTE compActiveStressModes[STRESS_COUNT];
8174 #define MAX_STRESS_WEIGHT 100
8176 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8180 bool compInlineStress()
8182 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8185 bool compRandomInlineStress()
8187 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8192 bool compTailCallStress()
8195 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8201 codeOptimize compCodeOpt()
8204 // Switching between size & speed has measurable throughput impact
8205 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8206 // DEBUG, but should generate identical code between CHK & RET builds,
8207 // so that's not acceptable.
8208 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8209 // Investigate the cause of the throughput regression.
8211 return opts.compCodeOpt;
8213 return BLENDED_CODE;
8217 //--------------------- Info about the procedure --------------------------
8221 COMP_HANDLE compCompHnd;
8222 CORINFO_MODULE_HANDLE compScopeHnd;
8223 CORINFO_CLASS_HANDLE compClassHnd;
8224 CORINFO_METHOD_HANDLE compMethodHnd;
8225 CORINFO_METHOD_INFO* compMethodInfo;
8227 BOOL hasCircularClassConstraints;
8228 BOOL hasCircularMethodConstraints;
8230 #if defined(DEBUG) || defined(LATE_DISASM)
8231 const char* compMethodName;
8232 const char* compClassName;
8233 const char* compFullName;
8234 #endif // defined(DEBUG) || defined(LATE_DISASM)
8236 #if defined(DEBUG) || defined(INLINE_DATA)
8237 // Method hash is logcally const, but computed
8239 mutable unsigned compMethodHashPrivate;
8240 unsigned compMethodHash() const;
8241 #endif // defined(DEBUG) || defined(INLINE_DATA)
8243 #ifdef PSEUDORANDOM_NOP_INSERTION
8244 // things for pseudorandom nop insertion
8245 unsigned compChecksum;
8249 // The following holds the FLG_xxxx flags for the method we're compiling.
8252 // The following holds the class attributes for the method we're compiling.
8253 unsigned compClassAttr;
8255 const BYTE* compCode;
8256 IL_OFFSET compILCodeSize; // The IL code size
8257 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8258 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8259 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8260 // (2) the code is hot/cold split, and we issued less code than we expected
8261 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8263 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8264 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8265 bool compIsContextful : 1; // contextful method
8266 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8267 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8268 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8269 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8270 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8272 var_types compRetType; // Return type of the method as declared in IL
8273 var_types compRetNativeType; // Normalized return type as per target arch ABI
8274 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8275 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8276 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8277 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8278 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8279 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8280 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8281 unsigned compMaxStack;
8282 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8283 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8285 unsigned compCallUnmanaged; // count of unmanaged calls
8286 unsigned compLvFrameListRoot; // lclNum for the Frame root
8287 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8288 // You should generally use compHndBBtabCount instead: it is the
8289 // current number of EH clauses (after additions like synchronized
8290 // methods and funclets, and removals like unreachable code deletion).
8292 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8293 // and the VM expects that, or the JIT is a "self-host" compiler
8294 // (e.g., x86 hosted targeting x86) and the VM expects that.
8296 /* The following holds IL scope information about local variables.
8299 unsigned compVarScopesCount;
8300 VarScopeDsc* compVarScopes;
8302 /* The following holds information about instr offsets for
8303 * which we need to report IP-mappings
8306 IL_OFFSET* compStmtOffsets; // sorted
8307 unsigned compStmtOffsetsCount;
8308 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8310 #define CPU_X86 0x0100 // The generic X86 CPU
8311 #define CPU_X86_PENTIUM_4 0x0110
8313 #define CPU_X64 0x0200 // The generic x64 CPU
8314 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8315 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8317 #define CPU_ARM 0x0300 // The generic ARM CPU
8319 unsigned genCPU; // What CPU are we running on
8322 // Returns true if the method being compiled returns a non-void and non-struct value.
8323 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8324 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8325 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8326 // Methods returning such structs are considered to return non-struct return value and
8327 // this method returns true in that case.
8328 bool compMethodReturnsNativeScalarType()
8330 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8333 // Returns true if the method being compiled returns RetBuf addr as its return value
8334 bool compMethodReturnsRetBufAddr()
8336 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8337 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8339 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8340 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8341 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8342 // methods with hidden RetBufArg.
8344 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8345 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8346 // returning the address of RetBuf.
8348 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8349 // to be returned in RAX.
8350 CLANG_FORMAT_COMMENT_ANCHOR;
8352 #ifdef _TARGET_AMD64_
8353 return (info.compRetBuffArg != BAD_VAR_NUM);
8354 #else // !_TARGET_AMD64_
8355 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8356 #endif // !_TARGET_AMD64_
8359 // Returns true if the method returns a value in more than one return register
8360 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8361 // TODO-ARM64: Does this apply for ARM64 too?
8362 bool compMethodReturnsMultiRegRetType()
8364 #if FEATURE_MULTIREG_RET
8365 #if defined(_TARGET_X86_)
8366 // On x86 only 64-bit longs are returned in multiple registers
8367 return varTypeIsLong(info.compRetNativeType);
8368 #else // targets: X64-UNIX, ARM64 or ARM32
8369 // On all other targets that support multireg return values:
8370 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8371 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8372 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8373 #endif // TARGET_XXX
8375 #else // not FEATURE_MULTIREG_RET
8377 // For this architecture there are no multireg returns
8380 #endif // FEATURE_MULTIREG_RET
8383 #if FEATURE_MULTIREG_ARGS
8384 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8385 // return the gcPtr layout for the pointers sized fields
8386 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8387 #endif // FEATURE_MULTIREG_ARGS
8389 // Returns true if the method being compiled returns a value
8390 bool compMethodHasRetVal()
8392 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8393 compMethodReturnsMultiRegRetType();
8398 void compDispLocalVars();
8402 //-------------------------- Global Compiler Data ------------------------------------
8405 static unsigned s_compMethodsCount; // to produce unique label names
8406 unsigned compGenTreeID;
8409 BasicBlock* compCurBB; // the current basic block in process
8410 GenTreePtr compCurStmt; // the current statement in process
8412 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8415 // The following is used to create the 'method JIT info' block.
8416 size_t compInfoBlkSize;
8417 BYTE* compInfoBlkAddr;
8419 EHblkDsc* compHndBBtab; // array of EH data
8420 unsigned compHndBBtabCount; // element count of used elements in EH data array
8421 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8423 #if defined(_TARGET_X86_)
8425 //-------------------------------------------------------------------------
8426 // Tracking of region covered by the monitor in synchronized methods
8427 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8428 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8430 #endif // !_TARGET_X86_
8432 Phases previousCompletedPhase; // the most recently completed phase
8434 //-------------------------------------------------------------------------
8435 // The following keeps track of how many bytes of local frame space we've
8436 // grabbed so far in the current function, and how many argument bytes we
8437 // need to pop when we return.
8440 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8442 // Count of callee-saved regs we pushed in the prolog.
8443 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8444 // In case of Amd64 this doesn't include float regs saved on stack.
8445 unsigned compCalleeRegsPushed;
8447 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8448 // Mask of callee saved float regs on stack.
8449 regMaskTP compCalleeFPRegsSavedMask;
8451 #ifdef _TARGET_AMD64_
8452 // Quirk for VS debug-launch scenario to work:
8453 // Bytes of padding between save-reg area and locals.
8454 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8455 unsigned compVSQuirkStackPaddingNeeded;
8456 bool compQuirkForPPPflag;
8459 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8461 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8462 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8463 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8465 //-------------------------------------------------------------------------
8467 static void compStartup(); // One-time initialization
8468 static void compShutdown(); // One-time finalization
8470 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8473 static void compDisplayStaticSizes(FILE* fout);
8475 //------------ Some utility functions --------------
8477 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8478 void** ppIndirection); /* OUT */
8480 // Several JIT/EE interface functions return a CorInfoType, and also return a
8481 // class handle as an out parameter if the type is a value class. Returns the
8482 // size of the type these describe.
8483 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8486 // Components used by the compiler may write unit test suites, and
8487 // have them run within this method. They will be run only once per process, and only
8488 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8489 // These should fail by asserting.
8490 void compDoComponentUnitTestsOnce();
8493 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8494 CORINFO_MODULE_HANDLE classPtr,
8495 COMP_HANDLE compHnd,
8496 CORINFO_METHOD_INFO* methodInfo,
8497 void** methodCodePtr,
8498 ULONG* methodCodeSize,
8499 JitFlags* compileFlags);
8500 void compCompileFinish();
8501 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8502 COMP_HANDLE compHnd,
8503 CORINFO_METHOD_INFO* methodInfo,
8504 void** methodCodePtr,
8505 ULONG* methodCodeSize,
8506 JitFlags* compileFlags,
8507 CorInfoInstantiationVerification instVerInfo);
8509 ArenaAllocator* compGetAllocator();
8511 #if MEASURE_MEM_ALLOC
8513 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8517 unsigned allocCnt; // # of allocs
8518 UINT64 allocSz; // total size of those alloc.
8519 UINT64 allocSzMax; // Maximum single allocation.
8520 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8521 UINT64 nraTotalSizeAlloc;
8522 UINT64 nraTotalSizeUsed;
8524 static const char* s_CompMemKindNames[]; // Names of the kinds.
8526 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8528 for (int i = 0; i < CMK_Count; i++)
8530 allocSzByKind[i] = 0;
8533 MemStats(const MemStats& ms)
8534 : allocCnt(ms.allocCnt)
8535 , allocSz(ms.allocSz)
8536 , allocSzMax(ms.allocSzMax)
8537 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8538 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8540 for (int i = 0; i < CMK_Count; i++)
8542 allocSzByKind[i] = ms.allocSzByKind[i];
8546 // Until we have ubiquitous constructors.
8549 this->MemStats::MemStats();
8552 void AddAlloc(size_t sz, CompMemKind cmk)
8556 if (sz > allocSzMax)
8560 allocSzByKind[cmk] += sz;
8563 void Print(FILE* f); // Print these stats to f.
8564 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8566 MemStats genMemStats;
8568 struct AggregateMemStats : public MemStats
8572 AggregateMemStats() : MemStats(), nMethods(0)
8576 void Add(const MemStats& ms)
8579 allocCnt += ms.allocCnt;
8580 allocSz += ms.allocSz;
8581 allocSzMax = max(allocSzMax, ms.allocSzMax);
8582 for (int i = 0; i < CMK_Count; i++)
8584 allocSzByKind[i] += ms.allocSzByKind[i];
8586 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8587 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8590 void Print(FILE* f); // Print these stats to jitstdout.
8593 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8594 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8595 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8597 #endif // MEASURE_MEM_ALLOC
8599 #if LOOP_HOIST_STATS
8600 unsigned m_loopsConsidered;
8601 bool m_curLoopHasHoistedExpression;
8602 unsigned m_loopsWithHoistedExpressions;
8603 unsigned m_totalHoistedExpressions;
8605 void AddLoopHoistStats();
8606 void PrintPerMethodLoopHoistStats();
8608 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8609 static unsigned s_loopsConsidered;
8610 static unsigned s_loopsWithHoistedExpressions;
8611 static unsigned s_totalHoistedExpressions;
8613 static void PrintAggregateLoopHoistStats(FILE* f);
8614 #endif // LOOP_HOIST_STATS
8616 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8617 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8618 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8619 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8620 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8621 void compFreeMem(void*);
8623 bool compIsForImportOnly();
8624 bool compIsForInlining();
8625 bool compDonotInline();
8628 const char* compLocalVarName(unsigned varNum, unsigned offs);
8629 VarName compVarName(regNumber reg, bool isFloatReg = false);
8630 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8631 const char* compRegPairName(regPairNo regPair);
8632 const char* compRegNameForSize(regNumber reg, size_t size);
8633 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8634 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8635 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8638 //-------------------------------------------------------------------------
8640 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8642 struct VarScopeMapInfo
8644 VarScopeListNode* head;
8645 VarScopeListNode* tail;
8646 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8648 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8655 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8656 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8658 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8659 VarNumToScopeDscMap;
8661 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8662 VarNumToScopeDscMap* compVarScopeMap;
8664 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8666 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8668 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8670 void compInitVarScopeMap();
8672 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8673 // enter scope, sorted by instr offset
8674 unsigned compNextEnterScope;
8676 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8677 // go out of scope, sorted by instr offset
8678 unsigned compNextExitScope;
8680 void compInitScopeLists();
8682 void compResetScopeLists();
8684 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8686 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8688 void compProcessScopesUntil(unsigned offset,
8690 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8691 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8694 void compDispScopeLists();
8697 bool compIsProfilerHookNeeded();
8699 //-------------------------------------------------------------------------
8700 /* Statistical Data Gathering */
8702 void compJitStats(); // call this function and enable
8703 // various ifdef's below for statistical data
8706 void compCallArgStats();
8707 static void compDispCallArgStats(FILE* fout);
8710 //-------------------------------------------------------------------------
8717 ArenaAllocator* compAllocator;
8720 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8721 // suitable for use by utilcode collection types.
8722 IAllocator* compAsIAllocator;
8724 #if MEASURE_MEM_ALLOC
8725 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8726 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8727 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8729 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8731 #endif // MEASURE_MEM_ALLOC
8733 void compFunctionTraceStart();
8734 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8737 size_t compMaxUncheckedOffsetForNullObject;
8739 void compInitOptions(JitFlags* compileFlags);
8741 void compSetProcessor();
8742 void compInitDebuggingInfo();
8743 void compSetOptimizationLevel();
8744 #ifdef _TARGET_ARMARCH_
8745 bool compRsvdRegCheck(FrameLayoutState curState);
8747 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8749 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8750 void ResetOptAnnotations();
8752 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8753 void RecomputeLoopInfo();
8755 #ifdef PROFILING_SUPPORTED
8756 // Data required for generating profiler Enter/Leave/TailCall hooks
8758 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8759 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8760 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8763 #ifdef _TARGET_AMD64_
8764 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8767 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8768 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8770 IAllocator* getAllocator()
8772 return compAsIAllocator;
8775 #if MEASURE_MEM_ALLOC
8776 IAllocator* getAllocatorBitset()
8778 return compAsIAllocatorBitset;
8780 IAllocator* getAllocatorGC()
8782 return compAsIAllocatorGC;
8784 IAllocator* getAllocatorLoopHoist()
8786 return compAsIAllocatorLoopHoist;
8788 #else // !MEASURE_MEM_ALLOC
8789 IAllocator* getAllocatorBitset()
8791 return compAsIAllocator;
8793 IAllocator* getAllocatorGC()
8795 return compAsIAllocator;
8797 IAllocator* getAllocatorLoopHoist()
8799 return compAsIAllocator;
8801 #endif // !MEASURE_MEM_ALLOC
8804 IAllocator* getAllocatorDebugOnly()
8806 #if MEASURE_MEM_ALLOC
8807 return compAsIAllocatorDebugOnly;
8808 #else // !MEASURE_MEM_ALLOC
8809 return compAsIAllocator;
8810 #endif // !MEASURE_MEM_ALLOC
8815 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8816 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8820 XX Checks for type compatibility and merges types XX
8822 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8823 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8827 // Set to TRUE if verification cannot be skipped for this method
8828 // If we detect unverifiable code, we will lazily check
8829 // canSkipMethodVerification() to see if verification is REALLY needed.
8830 BOOL tiVerificationNeeded;
8832 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8833 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8834 BOOL tiIsVerifiableCode;
8836 // Set to TRUE if runtime callout is needed for this method
8837 BOOL tiRuntimeCalloutNeeded;
8839 // Set to TRUE if security prolog/epilog callout is needed for this method
8840 // Note: This flag is different than compNeedSecurityCheck.
8841 // compNeedSecurityCheck means whether or not a security object needs
8842 // to be allocated on the stack, which is currently true for EnC as well.
8843 // tiSecurityCalloutNeeded means whether or not security callouts need
8844 // to be inserted in the jitted code.
8845 BOOL tiSecurityCalloutNeeded;
8847 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8848 // This support is necessary to suport attributes that are not described in
8849 // for example, signatures. For example, the permanent home byref (byref that
8850 // points to the gc heap), isn't a property of method signatures, therefore,
8851 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8852 // but when deciding if we need to reimport a block, we need to take these
8854 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8856 // Returns TRUE if child is equal to or a subtype of parent.
8857 // normalisedForStack indicates that both types are normalised for the stack
8858 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8860 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8861 // *pDest is modified to represent the merged type. Sets "*changed" to true
8862 // if this changes "*pDest".
8863 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8865 // Set pDest from the primitive value type.
8866 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8868 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8871 // <BUGNUM> VSW 471305
8872 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8873 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8874 // We use a "short" as we need to push/pop this scope.
8876 short compRegSetCheckLevel;
8880 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8881 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8883 XX IL verification stuff XX
8886 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8887 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8891 // The following is used to track liveness of local variables, initialization
8892 // of valueclass constructors, and type safe use of IL instructions.
8894 // dynamic state info needed for verification
8895 EntryState verCurrentState;
8897 // this ptr of object type .ctors are considered intited only after
8898 // the base class ctor is called, or an alternate ctor is called.
8899 // An uninited this ptr can be used to access fields, but cannot
8900 // be used to call a member function.
8901 BOOL verTrackObjCtorInitState;
8903 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8905 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8906 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8907 void verInitCurrentState();
8908 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8910 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8911 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8912 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8914 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8915 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8916 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8917 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8918 typeInfo verMakeTypeInfo(CorInfoType ciType,
8919 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8920 BOOL verIsSDArray(typeInfo ti);
8921 typeInfo verGetArrayElemType(typeInfo ti);
8923 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8924 BOOL verNeedsVerification();
8925 BOOL verIsByRefLike(const typeInfo& ti);
8926 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8928 // generic type variables range over types that satisfy IsBoxable
8929 BOOL verIsBoxable(const typeInfo& ti);
8931 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8932 DEBUGARG(unsigned line));
8933 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8934 DEBUGARG(unsigned line));
8935 bool verCheckTailCallConstraint(OPCODE opcode,
8936 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8937 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8938 // on a type parameter?
8939 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8940 // return false to the caller.
8941 // If false, it will throw.
8943 bool verIsBoxedValueType(typeInfo ti);
8945 void verVerifyCall(OPCODE opcode,
8946 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8947 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8949 bool readonlyCall, // is this a "readonly." call?
8950 const BYTE* delegateCreateStart,
8951 const BYTE* codeAddr,
8952 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8954 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8956 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8957 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8958 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8959 const CORINFO_FIELD_INFO& fieldInfo,
8960 const typeInfo* tiThis,
8962 BOOL allowPlainStructAsThis = FALSE);
8963 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8964 void verVerifyThisPtrInitialised();
8965 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8967 // Register allocator
8968 void raInitStackFP();
8969 void raEnregisterVarsPrePassStackFP();
8970 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8971 void raEnregisterVarsPostPassStackFP();
8972 void raGenerateFPRefCounts();
8973 void raEnregisterVarsStackFP();
8974 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8976 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8977 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8979 // returns true if enregistering v1 would save more mem accesses than v2
8980 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8983 void raDumpHeightsStackFP();
8984 void raDumpVariableRegIntfFloat();
8987 #if FEATURE_STACK_FP_X87
8989 // Currently, we use FP transition blocks in only 2 situations:
8991 // -conditional jump on longs where FP stack differs with target: it's not strictly
8992 // necessary, but its low frequency and the code would get complicated if we try to
8993 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8994 // minimize the way we generate the jump code.
8995 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8996 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8998 // However, transition blocks have 2 problems
9000 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9001 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9002 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9003 // in the right place without preordering them), this causes us to have to generate the transition
9004 // blocks in the cold area if we want procedure splitting.
9007 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9008 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9009 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9010 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9011 // a big change in the exception.
9013 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9014 // optimizations. For these 2 cases:
9016 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9017 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9018 // a switch statement.
9020 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9021 // current procedure splitting and exception code have.
9022 bool compMayHaveTransitionBlocks;
9024 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9026 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9028 unsigned raCntStkStackFP;
9029 unsigned raCntWtdStkDblStackFP;
9030 unsigned raCntStkParamDblStackFP;
9032 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9033 // TODO: Do we want to put this in LclVarDsc?
9034 unsigned raPayloadStackFP[lclMAX_TRACKED];
9035 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9037 // Useful for debugging
9038 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9040 #endif // FEATURE_STACK_FP_X87
9043 // One line log function. Default level is 0. Increasing it gives you
9044 // more log information
9046 // levels are currently unused: #define JITDUMP(level,...) ();
9047 void JitLogEE(unsigned level, const char* fmt, ...);
9049 bool compDebugBreak;
9051 bool compJitHaltMethod();
9056 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9057 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9059 XX GS Security checks for unsafe buffers XX
9061 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9062 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9065 struct ShadowParamVarInfo
9067 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9068 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9070 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9072 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9073 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9074 // slots and update all trees to refer to shadow slots is done immediately after
9075 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9076 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9077 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9078 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9079 // creating a shadow slot even though this routine returns true.
9081 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9082 // required. There are two cases under which a reg arg could potentially be used from its
9084 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9085 // b) LSRA spills it
9087 // Possible solution to address case (a)
9088 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9089 // in this routine. Note that live out of exception handler is something we may not be
9090 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9091 // Therefore, for methods with exception handling and need GS cookie check we might have
9092 // to take conservative approach.
9094 // Possible solution to address case (b)
9095 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9096 // create a new spill temp if the method needs GS cookie check.
9097 return varDsc->lvIsParam;
9098 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9099 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9106 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9111 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9112 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9113 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9115 void gsGSChecksInitCookie(); // Grabs cookie variable
9116 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9117 bool gsFindVulnerableParams(); // Shadow param analysis code
9118 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9120 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9121 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9123 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9124 // This can be overwritten by setting complus_JITInlineSize env variable.
9126 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9129 #ifdef FEATURE_JIT_METHOD_PERF
9130 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9131 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9133 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9134 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9136 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9138 #if MEASURE_CLRAPI_CALLS
9139 // Thin wrappers that call into JitTimer (if present).
9140 inline void CLRApiCallEnter(unsigned apix);
9141 inline void CLRApiCallLeave(unsigned apix);
9144 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9145 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9150 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9151 // These variables are associated with maintaining SQM data about compile time.
9152 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9153 // in the current compilation.
9154 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9155 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9156 // the inlining phase in the current compilation.
9157 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9159 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9160 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9161 // type-loading and class initialization).
9162 void RecordStateAtEndOfInlining();
9163 // Assumes being called at the end of compilation. Update the SQM state.
9164 void RecordStateAtEndOfCompilation();
9166 #ifdef FEATURE_CLRSQM
9167 // Does anything SQM related necessary at process shutdown time.
9168 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9169 #endif // FEATURE_CLRSQM
9172 #if FUNC_INFO_LOGGING
9173 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9174 // filename to write it to.
9175 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9176 #endif // FUNC_INFO_LOGGING
9178 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9180 // Is the compilation in a full trust context?
9181 bool compIsFullTrust();
9184 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9185 #endif // MEASURE_NOWAY
9187 #ifndef FEATURE_TRACELOGGING
9188 // Should we actually fire the noway assert body and the exception handler?
9189 bool compShouldThrowOnNoway();
9190 #else // FEATURE_TRACELOGGING
9191 // Should we actually fire the noway assert body and the exception handler?
9192 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9194 // Telemetry instance to use per method compilation.
9195 JitTelemetry compJitTelemetry;
9197 // Get common parameters that have to be logged with most telemetry data.
9198 void compGetTelemetryDefaults(const char** assemblyName,
9199 const char** scopeName,
9200 const char** methodName,
9201 unsigned* methodHash);
9202 #endif // !FEATURE_TRACELOGGING
9206 NodeToTestDataMap* m_nodeTestData;
9208 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9209 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9210 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9211 // Current kept in this.
9213 NodeToTestDataMap* GetNodeTestData()
9215 Compiler* compRoot = impInlineRoot();
9216 if (compRoot->m_nodeTestData == nullptr)
9218 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9220 return compRoot->m_nodeTestData;
9223 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9225 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9226 // currently occur in the AST graph.
9227 NodeToIntMap* FindReachableNodesInNodeTestData();
9229 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9230 // test data, associate that data with "to".
9231 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9233 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9234 // have annotations, attach similar annotations to the corresponding nodes in "to".
9235 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9237 // These are the methods that test that the various conditions implied by the
9238 // test attributes are satisfied.
9239 void JitTestCheckSSA(); // SSA builder tests.
9240 void JitTestCheckVN(); // Value numbering tests.
9243 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9245 FieldSeqStore* m_fieldSeqStore;
9247 FieldSeqStore* GetFieldSeqStore()
9249 Compiler* compRoot = impInlineRoot();
9250 if (compRoot->m_fieldSeqStore == nullptr)
9252 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9253 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9254 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9256 return compRoot->m_fieldSeqStore;
9259 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9261 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9262 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9263 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9264 // attach the field sequence directly to the address node.
9265 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9267 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9269 // Don't need to worry about inlining here
9270 if (m_zeroOffsetFieldMap == nullptr)
9272 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9274 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9275 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9277 return m_zeroOffsetFieldMap;
9280 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9281 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9282 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9283 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9284 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9285 // record the the field sequence using the ZeroOffsetFieldMap described above.
9287 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9288 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9289 // CoreRT. Such case is handled same as the default case.
9290 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9292 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9294 NodeToArrayInfoMap* m_arrayInfoMap;
9296 NodeToArrayInfoMap* GetArrayInfoMap()
9298 Compiler* compRoot = impInlineRoot();
9299 if (compRoot->m_arrayInfoMap == nullptr)
9301 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9302 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9303 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9305 return compRoot->m_arrayInfoMap;
9308 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9310 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9311 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9312 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9313 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9315 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9317 // Use the same map for GCHeap and ByrefExposed when their states match.
9318 memoryKind = ByrefExposed;
9321 assert(memoryKind < MemoryKindCount);
9322 Compiler* compRoot = impInlineRoot();
9323 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9325 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9326 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9327 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9329 return compRoot->m_memorySsaMap[memoryKind];
9332 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9333 CORINFO_CLASS_HANDLE m_refAnyClass;
9334 CORINFO_FIELD_HANDLE GetRefanyDataField()
9336 if (m_refAnyClass == nullptr)
9338 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9340 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9342 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9344 if (m_refAnyClass == nullptr)
9346 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9348 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9352 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9354 #if ALLVARSET_COUNTOPS
9355 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9358 static HelperCallProperties s_helperCallProperties;
9360 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9361 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9362 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9364 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9367 unsigned __int8* offset0,
9368 unsigned __int8* offset1);
9369 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9370 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9372 void fgMorphMultiregStructArgs(GenTreeCall* call);
9373 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9375 }; // end of class Compiler
9377 // Inline methods of CompAllocator.
9378 void* CompAllocator::Alloc(size_t sz)
9380 #if MEASURE_MEM_ALLOC
9381 return m_comp->compGetMem(sz, m_cmk);
9383 return m_comp->compGetMem(sz);
9387 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9389 #if MEASURE_MEM_ALLOC
9390 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9392 return m_comp->compGetMemArray(elems, elemSize);
9396 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9397 inline LclVarDsc::LclVarDsc(Compiler* comp)
9398 : // Initialize the ArgRegs to REG_STK.
9399 // The morph will do the right thing to change
9400 // to the right register if passed in register.
9403 #if FEATURE_MULTIREG_ARGS
9404 _lvOtherArgReg(REG_STK)
9406 #endif // FEATURE_MULTIREG_ARGS
9408 lvRefBlks(BlockSetOps::UninitVal())
9410 #endif // ASSERTION_PROP
9411 lvPerSsaData(comp->getAllocator())
9415 //---------------------------------------------------------------------------------------------------------------------
9416 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9418 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9419 // shown in parentheses):
9421 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9422 // of a misnomer, as the first entry will always be the current node.
9424 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9425 // argument before visiting the node's operands.
9427 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9428 // argument after visiting the node's operands.
9430 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9431 // `DoPreOrder` must be true if this option is true.
9433 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9434 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9435 // visited before the first).
9437 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9439 // A simple pre-order visitor might look something like the following:
9441 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9446 // DoPreOrder = true
9449 // unsigned m_count;
9451 // CountingVisitor(Compiler* compiler)
9452 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9456 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9462 // This visitor would then be used like so:
9464 // CountingVisitor countingVisitor(compiler);
9465 // countingVisitor.WalkTree(root);
9467 template <typename TVisitor>
9468 class GenTreeVisitor
9471 typedef Compiler::fgWalkResult fgWalkResult;
9475 ComputeStack = false,
9477 DoPostOrder = false,
9478 DoLclVarsOnly = false,
9479 UseExecutionOrder = false,
9482 Compiler* m_compiler;
9483 ArrayStack<GenTree*> m_ancestors;
9485 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9487 assert(compiler != nullptr);
9489 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9490 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9493 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9495 return fgWalkResult::WALK_CONTINUE;
9498 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9500 return fgWalkResult::WALK_CONTINUE;
9504 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9506 assert(use != nullptr);
9508 GenTree* node = *use;
9510 if (TVisitor::ComputeStack)
9512 m_ancestors.Push(node);
9515 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9516 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9518 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9519 if (result == fgWalkResult::WALK_ABORT)
9525 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9531 switch (node->OperGet())
9536 case GT_LCL_VAR_ADDR:
9537 case GT_LCL_FLD_ADDR:
9538 if (TVisitor::DoLclVarsOnly)
9540 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9541 if (result == fgWalkResult::WALK_ABORT)
9557 case GT_MEMORYBARRIER:
9562 case GT_START_NONGC:
9564 #if !FEATURE_EH_FUNCLETS
9566 #endif // !FEATURE_EH_FUNCLETS
9568 #ifndef LEGACY_BACKEND
9570 #endif // LEGACY_BACKEND
9573 case GT_CLS_VAR_ADDR:
9577 case GT_PINVOKE_PROLOG:
9578 case GT_PINVOKE_EPILOG:
9582 // Lclvar unary operators
9583 case GT_STORE_LCL_VAR:
9584 case GT_STORE_LCL_FLD:
9585 if (TVisitor::DoLclVarsOnly)
9587 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9588 if (result == fgWalkResult::WALK_ABORT)
9595 // Standard unary operators
9623 GenTreeUnOp* const unOp = node->AsUnOp();
9624 if (unOp->gtOp1 != nullptr)
9626 result = WalkTree(&unOp->gtOp1, unOp);
9627 if (result == fgWalkResult::WALK_ABORT)
9638 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9640 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9641 if (result == fgWalkResult::WALK_ABORT)
9645 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9646 if (result == fgWalkResult::WALK_ABORT)
9650 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9651 if (result == fgWalkResult::WALK_ABORT)
9658 case GT_ARR_BOUNDS_CHECK:
9661 #endif // FEATURE_SIMD
9663 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9665 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9666 if (result == fgWalkResult::WALK_ABORT)
9670 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9671 if (result == fgWalkResult::WALK_ABORT)
9680 GenTreeField* const field = node->AsField();
9682 if (field->gtFldObj != nullptr)
9684 result = WalkTree(&field->gtFldObj, field);
9685 if (result == fgWalkResult::WALK_ABORT)
9695 GenTreeArrElem* const arrElem = node->AsArrElem();
9697 result = WalkTree(&arrElem->gtArrObj, arrElem);
9698 if (result == fgWalkResult::WALK_ABORT)
9703 const unsigned rank = arrElem->gtArrRank;
9704 for (unsigned dim = 0; dim < rank; dim++)
9706 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9707 if (result == fgWalkResult::WALK_ABORT)
9717 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9719 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9720 if (result == fgWalkResult::WALK_ABORT)
9724 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9725 if (result == fgWalkResult::WALK_ABORT)
9729 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9730 if (result == fgWalkResult::WALK_ABORT)
9739 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9741 GenTree** op1Use = &dynBlock->gtOp1;
9742 GenTree** op2Use = &dynBlock->gtDynamicSize;
9744 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9746 std::swap(op1Use, op2Use);
9749 result = WalkTree(op1Use, dynBlock);
9750 if (result == fgWalkResult::WALK_ABORT)
9754 result = WalkTree(op2Use, dynBlock);
9755 if (result == fgWalkResult::WALK_ABORT)
9762 case GT_STORE_DYN_BLK:
9764 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9766 GenTree** op1Use = &dynBlock->gtOp1;
9767 GenTree** op2Use = &dynBlock->gtOp2;
9768 GenTree** op3Use = &dynBlock->gtDynamicSize;
9770 if (TVisitor::UseExecutionOrder)
9772 if (dynBlock->IsReverseOp())
9774 std::swap(op1Use, op2Use);
9776 if (dynBlock->gtEvalSizeFirst)
9778 std::swap(op3Use, op2Use);
9779 std::swap(op2Use, op1Use);
9783 result = WalkTree(op1Use, dynBlock);
9784 if (result == fgWalkResult::WALK_ABORT)
9788 result = WalkTree(op2Use, dynBlock);
9789 if (result == fgWalkResult::WALK_ABORT)
9793 result = WalkTree(op3Use, dynBlock);
9794 if (result == fgWalkResult::WALK_ABORT)
9803 GenTreeCall* const call = node->AsCall();
9805 if (call->gtCallObjp != nullptr)
9807 result = WalkTree(&call->gtCallObjp, call);
9808 if (result == fgWalkResult::WALK_ABORT)
9814 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9816 result = WalkTree(args->pCurrent(), call);
9817 if (result == fgWalkResult::WALK_ABORT)
9823 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9825 result = WalkTree(args->pCurrent(), call);
9826 if (result == fgWalkResult::WALK_ABORT)
9832 if (call->gtCallType == CT_INDIRECT)
9834 if (call->gtCallCookie != nullptr)
9836 result = WalkTree(&call->gtCallCookie, call);
9837 if (result == fgWalkResult::WALK_ABORT)
9843 result = WalkTree(&call->gtCallAddr, call);
9844 if (result == fgWalkResult::WALK_ABORT)
9850 if (call->gtControlExpr != nullptr)
9852 result = WalkTree(&call->gtControlExpr, call);
9853 if (result == fgWalkResult::WALK_ABORT)
9865 assert(node->OperIsBinary());
9867 GenTreeOp* const op = node->AsOp();
9869 GenTree** op1Use = &op->gtOp1;
9870 GenTree** op2Use = &op->gtOp2;
9872 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9874 std::swap(op1Use, op2Use);
9877 if (*op1Use != nullptr)
9879 result = WalkTree(op1Use, op);
9880 if (result == fgWalkResult::WALK_ABORT)
9886 if (*op2Use != nullptr)
9888 result = WalkTree(op2Use, op);
9889 if (result == fgWalkResult::WALK_ABORT)
9899 // Finally, visit the current node
9900 if (TVisitor::DoPostOrder)
9902 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9905 if (TVisitor::ComputeStack)
9914 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9915 class GenericTreeWalker final
9916 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
9921 ComputeStack = computeStack,
9922 DoPreOrder = doPreOrder,
9923 DoPostOrder = doPostOrder,
9924 DoLclVarsOnly = doLclVarsOnly,
9925 UseExecutionOrder = useExecutionOrder,
9929 Compiler::fgWalkData* m_walkData;
9932 GenericTreeWalker(Compiler::fgWalkData* walkData)
9933 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
9935 , m_walkData(walkData)
9937 assert(walkData != nullptr);
9941 walkData->parentStack = &this->m_ancestors;
9945 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9947 m_walkData->parent = user;
9948 return m_walkData->wtprVisitorFn(use, m_walkData);
9951 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9953 m_walkData->parent = user;
9954 return m_walkData->wtpoVisitorFn(use, m_walkData);
9958 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
9964 DoLclVarsOnly = true
9967 IncLclVarRefCountsVisitor(Compiler* compiler);
9968 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9970 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9973 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
9979 DoLclVarsOnly = true
9982 DecLclVarRefCountsVisitor(Compiler* compiler);
9983 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9985 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9989 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9990 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9992 XX Miscellaneous Compiler stuff XX
9994 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9995 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9998 // Values used to mark the types a stack slot is used for
10000 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10001 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10002 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10003 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10004 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10005 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10006 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10007 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10009 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10011 /*****************************************************************************
10013 * Variables to keep track of total code amounts.
10018 extern size_t grossVMsize;
10019 extern size_t grossNCsize;
10020 extern size_t totalNCsize;
10022 extern unsigned genMethodICnt;
10023 extern unsigned genMethodNCnt;
10024 extern size_t gcHeaderISize;
10025 extern size_t gcPtrMapISize;
10026 extern size_t gcHeaderNSize;
10027 extern size_t gcPtrMapNSize;
10029 #endif // DISPLAY_SIZES
10031 /*****************************************************************************
10033 * Variables to keep track of basic block counts (more data on 1 BB methods)
10036 #if COUNT_BASIC_BLOCKS
10037 extern Histogram bbCntTable;
10038 extern Histogram bbOneBBSizeTable;
10041 /*****************************************************************************
10043 * Used by optFindNaturalLoops to gather statistical information such as
10044 * - total number of natural loops
10045 * - number of loops with 1, 2, ... exit conditions
10046 * - number of loops that have an iterator (for like)
10047 * - number of loops that have a constant iterator
10052 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10053 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10054 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10055 extern unsigned totalLoopCount; // counts the total number of natural loops
10056 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10057 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10058 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10059 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10061 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10062 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10063 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10064 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10065 extern Histogram loopCountTable; // Histogram of loop counts
10066 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10068 #endif // COUNT_LOOPS
10070 /*****************************************************************************
10071 * variables to keep track of how many iterations we go in a dataflow pass
10076 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10077 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10079 #endif // DATAFLOW_ITER
10081 #if MEASURE_BLOCK_SIZE
10082 extern size_t genFlowNodeSize;
10083 extern size_t genFlowNodeCnt;
10084 #endif // MEASURE_BLOCK_SIZE
10086 #if MEASURE_NODE_SIZE
10087 struct NodeSizeStats
10091 genTreeNodeCnt = 0;
10092 genTreeNodeSize = 0;
10093 genTreeNodeActualSize = 0;
10096 // Count of tree nodes allocated.
10097 unsigned __int64 genTreeNodeCnt;
10099 // The size we allocate.
10100 unsigned __int64 genTreeNodeSize;
10102 // The actual size of the node. Note that the actual size will likely be smaller
10103 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10104 // a smaller node to a larger one. TODO-Cleanup: add stats on
10105 // SetOper()/ChangeOper() usage to quantify this.
10106 unsigned __int64 genTreeNodeActualSize;
10108 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10109 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10110 extern Histogram genTreeNcntHist;
10111 extern Histogram genTreeNsizHist;
10112 #endif // MEASURE_NODE_SIZE
10114 /*****************************************************************************
10115 * Count fatal errors (including noway_asserts).
10119 extern unsigned fatal_badCode;
10120 extern unsigned fatal_noWay;
10121 extern unsigned fatal_NOMEM;
10122 extern unsigned fatal_noWayAssertBody;
10124 extern unsigned fatal_noWayAssertBodyArgs;
10126 extern unsigned fatal_NYI;
10127 #endif // MEASURE_FATAL
10129 /*****************************************************************************
10133 #ifdef _TARGET_XARCH_
10135 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10136 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10137 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10139 const instruction INS_AND = INS_and;
10140 const instruction INS_OR = INS_or;
10141 const instruction INS_XOR = INS_xor;
10142 const instruction INS_NEG = INS_neg;
10143 const instruction INS_TEST = INS_test;
10144 const instruction INS_MUL = INS_imul;
10145 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10146 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10147 const instruction INS_BREAKPOINT = INS_int3;
10148 const instruction INS_ADDC = INS_adc;
10149 const instruction INS_SUBC = INS_sbb;
10150 const instruction INS_NOT = INS_not;
10154 #ifdef _TARGET_ARM_
10156 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10157 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10158 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10160 const instruction INS_AND = INS_and;
10161 const instruction INS_OR = INS_orr;
10162 const instruction INS_XOR = INS_eor;
10163 const instruction INS_NEG = INS_rsb;
10164 const instruction INS_TEST = INS_tst;
10165 const instruction INS_MUL = INS_mul;
10166 const instruction INS_MULADD = INS_mla;
10167 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10168 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10169 const instruction INS_BREAKPOINT = INS_bkpt;
10170 const instruction INS_ADDC = INS_adc;
10171 const instruction INS_SUBC = INS_sbc;
10172 const instruction INS_NOT = INS_mvn;
10174 const instruction INS_ABS = INS_vabs;
10175 const instruction INS_ROUND = INS_invalid;
10176 const instruction INS_SQRT = INS_vsqrt;
10180 #ifdef _TARGET_ARM64_
10182 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10183 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10184 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10186 const instruction INS_AND = INS_and;
10187 const instruction INS_OR = INS_orr;
10188 const instruction INS_XOR = INS_eor;
10189 const instruction INS_NEG = INS_neg;
10190 const instruction INS_TEST = INS_tst;
10191 const instruction INS_MUL = INS_mul;
10192 const instruction INS_MULADD = INS_madd;
10193 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10194 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10195 const instruction INS_BREAKPOINT = INS_bkpt;
10196 const instruction INS_ADDC = INS_adc;
10197 const instruction INS_SUBC = INS_sbc;
10198 const instruction INS_NOT = INS_mvn;
10200 const instruction INS_ABS = INS_fabs;
10201 const instruction INS_ROUND = INS_frintn;
10202 const instruction INS_SQRT = INS_fsqrt;
10206 /*****************************************************************************/
10208 extern const BYTE genTypeSizes[];
10209 extern const BYTE genTypeAlignments[];
10210 extern const BYTE genTypeStSzs[];
10211 extern const BYTE genActualTypes[];
10213 /*****************************************************************************/
10215 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10216 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10218 #ifdef _TARGET_ARM_
10219 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10220 #elif defined(_TARGET_ARM64_)
10221 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10224 /*****************************************************************************/
10226 #define REG_CORRUPT regNumber(REG_NA + 1)
10227 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10228 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10230 /*****************************************************************************/
10232 extern BasicBlock dummyBB;
10234 /*****************************************************************************/
10235 /*****************************************************************************/
10237 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10238 // nodes of a statement in execution order.
10239 // __stmt: a GT_STMT type GenTree*
10240 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10242 #define foreach_treenode_execution_order(__node, __stmt) \
10243 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10245 // foreach_block: An iterator over all blocks in the function.
10246 // __compiler: the Compiler* object
10247 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10249 #define foreach_block(__compiler, __block) \
10250 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10252 /*****************************************************************************/
10253 /*****************************************************************************/
10257 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10259 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10260 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10262 XX Debugging helpers XX
10264 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10265 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10268 /*****************************************************************************/
10269 /* The following functions are intended to be called from the debugger, to dump
10270 * various data structures. The can be used in the debugger Watch or Quick Watch
10271 * windows. They are designed to be short to type and take as few arguments as
10272 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10273 * See the function definition comment for more details.
10276 void cBlock(Compiler* comp, BasicBlock* block);
10277 void cBlocks(Compiler* comp);
10278 void cBlocksV(Compiler* comp);
10279 void cTree(Compiler* comp, GenTree* tree);
10280 void cTrees(Compiler* comp);
10281 void cEH(Compiler* comp);
10282 void cVar(Compiler* comp, unsigned lclNum);
10283 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10284 void cVars(Compiler* comp);
10285 void cVarsFinal(Compiler* comp);
10286 void cBlockPreds(Compiler* comp, BasicBlock* block);
10287 void cReach(Compiler* comp);
10288 void cDoms(Compiler* comp);
10289 void cLiveness(Compiler* comp);
10290 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10292 void cFuncIR(Compiler* comp);
10293 void cBlockIR(Compiler* comp, BasicBlock* block);
10294 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10295 void cTreeIR(Compiler* comp, GenTree* tree);
10296 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10297 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10298 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10299 int cOperandIR(Compiler* comp, GenTree* operand);
10300 int cLeafIR(Compiler* comp, GenTree* tree);
10301 int cIndirIR(Compiler* comp, GenTree* tree);
10302 int cListIR(Compiler* comp, GenTree* list);
10303 int cSsaNumIR(Compiler* comp, GenTree* tree);
10304 int cValNumIR(Compiler* comp, GenTree* tree);
10305 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10307 void dBlock(BasicBlock* block);
10310 void dTree(GenTree* tree);
10313 void dVar(unsigned lclNum);
10314 void dVarDsc(LclVarDsc* varDsc);
10317 void dBlockPreds(BasicBlock* block);
10321 void dCVarSet(VARSET_VALARG_TP vars);
10323 void dVarSet(VARSET_VALARG_TP vars);
10324 void dRegMask(regMaskTP mask);
10327 void dBlockIR(BasicBlock* block);
10328 void dTreeIR(GenTree* tree);
10329 void dLoopIR(Compiler::LoopDsc* loop);
10330 void dLoopNumIR(unsigned loopNum);
10331 int dTabStopIR(int curr, int tabstop);
10332 int dTreeTypeIR(GenTree* tree);
10333 int dTreeKindsIR(GenTree* tree);
10334 int dTreeFlagsIR(GenTree* tree);
10335 int dOperandIR(GenTree* operand);
10336 int dLeafIR(GenTree* tree);
10337 int dIndirIR(GenTree* tree);
10338 int dListIR(GenTree* list);
10339 int dSsaNumIR(GenTree* tree);
10340 int dValNumIR(GenTree* tree);
10341 int dDependsIR(GenTree* comma);
10344 GenTree* dFindTree(GenTree* tree, unsigned id);
10345 GenTree* dFindTree(unsigned id);
10346 GenTreeStmt* dFindStmt(unsigned id);
10347 BasicBlock* dFindBlock(unsigned bbNum);
10351 #include "compiler.hpp" // All the shared inline functions
10353 /*****************************************************************************/
10354 #endif //_COMPILER_H_
10355 /*****************************************************************************/