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);
2764 #endif // defined(_TARGET_64BIT_)
2766 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2767 // already does this calculation. However, we also need to prevent mapping types if the var is a
2768 // dependently promoted struct field, which must remain its exact size within its parent struct.
2769 // However, we don't know this until late, so we may have already pretended the field is bigger
2771 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2780 #endif // defined(FEATURE_SIMD)
2782 BYTE* lvaGetGcLayout(unsigned varNum);
2783 bool lvaTypeIsGC(unsigned varNum);
2784 unsigned lvaGSSecurityCookie; // LclVar number
2785 bool lvaTempsHaveLargerOffsetThanVars();
2787 unsigned lvaSecurityObject; // variable representing the security object on the stack
2788 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2790 #if FEATURE_EH_FUNCLETS
2791 unsigned lvaPSPSym; // variable representing the PSPSym
2794 InlineInfo* impInlineInfo;
2795 InlineStrategy* m_inlineStrategy;
2797 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2798 Compiler* impInlineRoot();
2800 #if defined(DEBUG) || defined(INLINE_DATA)
2801 unsigned __int64 getInlineCycleCount()
2803 return m_compCycles;
2805 #endif // defined(DEBUG) || defined(INLINE_DATA)
2807 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2808 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2810 //=========================================================================
2812 //=========================================================================
2815 //---------------- Local variable ref-counting ----------------------------
2818 BasicBlock* lvaMarkRefsCurBlock;
2819 GenTreePtr lvaMarkRefsCurStmt;
2821 BasicBlock::weight_t lvaMarkRefsWeight;
2823 void lvaMarkLclRefs(GenTreePtr tree);
2825 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2826 void SetVolatileHint(LclVarDsc* varDsc);
2828 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2829 PerSsaArray lvMemoryPerSsaData;
2830 unsigned lvMemoryNumSsaNames;
2833 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2834 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2835 // not an SSA variable).
2836 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2838 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2839 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2841 assert(ssaNum < lvMemoryNumSsaNames);
2842 return &lvMemoryPerSsaData.GetRef(ssaNum);
2846 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2847 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2851 XX Imports the given method and converts it to semantic trees XX
2853 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2854 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2860 void impImport(BasicBlock* method);
2862 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2863 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2864 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2865 CORINFO_CLASS_HANDLE impGetStringClass();
2866 CORINFO_CLASS_HANDLE impGetObjectClass();
2868 //=========================================================================
2870 //=========================================================================
2873 //-------------------- Stack manipulation ---------------------------------
2875 unsigned impStkSize; // Size of the full stack
2877 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2879 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2881 struct SavedStack // used to save/restore stack contents.
2883 unsigned ssDepth; // number of values on stack
2884 StackEntry* ssTrees; // saved tree values
2887 bool impIsPrimitive(CorInfoType type);
2888 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2890 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2892 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2893 void impPushNullObjRefOnStack();
2894 StackEntry impPopStack();
2895 StackEntry& impStackTop(unsigned n = 0);
2896 unsigned impStackHeight();
2898 void impSaveStackState(SavedStack* savePtr, bool copy);
2899 void impRestoreStackState(SavedStack* savePtr);
2901 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2902 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2903 CORINFO_CALL_INFO* pCallInfo);
2905 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2907 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2909 bool impCanPInvokeInline();
2910 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2911 void impCheckForPInvokeCall(
2912 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2913 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2914 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2916 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2917 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2918 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2920 var_types impImportCall(OPCODE opcode,
2921 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2922 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2924 GenTreePtr newobjThis,
2926 CORINFO_CALL_INFO* callInfo,
2927 IL_OFFSET rawILOffset);
2929 void impDevirtualizeCall(GenTreeCall* call,
2931 CORINFO_METHOD_HANDLE* method,
2932 unsigned* methodFlags,
2933 CORINFO_CONTEXT_HANDLE* contextHandle,
2934 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2936 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2938 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2940 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2943 var_types impImportJitTestLabelMark(int numArgs);
2946 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2948 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2950 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2951 CORINFO_ACCESS_FLAGS access,
2952 CORINFO_FIELD_INFO* pFieldInfo,
2955 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2957 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2959 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2961 void impImportLeave(BasicBlock* block);
2962 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2963 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2964 CORINFO_CLASS_HANDLE clsHnd,
2965 CORINFO_METHOD_HANDLE method,
2966 CORINFO_SIG_INFO* sig,
2970 CorInfoIntrinsics* pIntrinsicID);
2971 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2972 CORINFO_SIG_INFO* sig,
2975 CorInfoIntrinsics intrinsicID);
2976 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2978 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2980 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2981 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2982 CORINFO_THIS_TRANSFORM transform);
2984 //----------------- Manipulating the trees and stmts ----------------------
2986 GenTreePtr impTreeList; // Trees for the BB being imported
2987 GenTreePtr impTreeLast; // The last tree for the current BB
2991 CHECK_SPILL_ALL = -1,
2992 CHECK_SPILL_NONE = -2
2996 void impBeginTreeList();
2997 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2998 void impEndTreeList(BasicBlock* block);
2999 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3000 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3001 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3002 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3003 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3004 void impAssignTempGen(unsigned tmp,
3007 GenTreePtr* pAfterStmt = nullptr,
3008 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3009 BasicBlock* block = nullptr);
3010 void impAssignTempGen(unsigned tmpNum,
3012 CORINFO_CLASS_HANDLE structHnd,
3014 GenTreePtr* pAfterStmt = nullptr,
3015 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3016 BasicBlock* block = nullptr);
3017 GenTreePtr impCloneExpr(GenTreePtr tree,
3019 CORINFO_CLASS_HANDLE structHnd,
3021 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3022 GenTreePtr impAssignStruct(GenTreePtr dest,
3024 CORINFO_CLASS_HANDLE structHnd,
3026 GenTreePtr* pAfterStmt = nullptr,
3027 BasicBlock* block = nullptr);
3028 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3030 CORINFO_CLASS_HANDLE structHnd,
3032 GenTreePtr* pAfterStmt = nullptr,
3033 BasicBlock* block = nullptr);
3035 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3036 CORINFO_CLASS_HANDLE structHnd,
3040 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3041 BYTE* gcLayout = nullptr,
3042 unsigned* numGCVars = nullptr,
3043 var_types* simdBaseType = nullptr);
3045 GenTreePtr impNormStructVal(GenTreePtr structVal,
3046 CORINFO_CLASS_HANDLE structHnd,
3048 bool forceNormalization = false);
3050 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3051 BOOL* pRuntimeLookup = nullptr,
3052 BOOL mustRestoreHandle = FALSE,
3053 BOOL importParent = FALSE);
3055 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3056 BOOL* pRuntimeLookup = nullptr,
3057 BOOL mustRestoreHandle = FALSE)
3059 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3062 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3063 CORINFO_LOOKUP* pLookup,
3065 void* compileTimeHandle);
3067 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3069 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3070 CORINFO_LOOKUP* pLookup,
3071 void* compileTimeHandle);
3073 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3075 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3076 CorInfoHelpFunc helper,
3078 GenTreeArgList* arg = nullptr,
3079 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3081 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3083 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3086 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3087 CORINFO_CLASS_HANDLE typeClass,
3091 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3092 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3093 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3094 static bool IsMathIntrinsic(GenTreePtr tree);
3097 //----------------- Importing the method ----------------------------------
3099 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3102 unsigned impCurOpcOffs;
3103 const char* impCurOpcName;
3104 bool impNestedStackSpill;
3106 // For displaying instrs with generated native code (-n:B)
3107 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3108 void impNoteLastILoffs();
3111 /* IL offset of the stmt currently being imported. It gets set to
3112 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3113 updated at IL offsets for which we have to report mapping info.
3114 It also includes flag bits, so use jitGetILoffs()
3115 to get the actual IL offset value.
3118 IL_OFFSETX impCurStmtOffs;
3119 void impCurStmtOffsSet(IL_OFFSET offs);
3121 void impNoteBranchOffs();
3123 unsigned impInitBlockLineInfo();
3125 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3126 bool impIsThis(GenTreePtr obj);
3127 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3128 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3129 bool impIsAnySTLOC(OPCODE opcode)
3131 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3132 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3135 GenTreeArgList* impPopList(unsigned count,
3137 CORINFO_SIG_INFO* sig,
3138 GenTreeArgList* prefixTree = nullptr);
3140 GenTreeArgList* impPopRevList(unsigned count,
3142 CORINFO_SIG_INFO* sig,
3143 unsigned skipReverseCount = 0);
3146 * Get current IL offset with stack-empty info incoporated
3148 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3150 //---------------- Spilling the importer stack ----------------------------
3152 // The maximum number of bytes of IL processed without clean stack state.
3153 // It allows to limit the maximum tree size and depth.
3154 static const unsigned MAX_TREE_SIZE = 200;
3155 bool impCanSpillNow(OPCODE prevOpcode);
3161 SavedStack pdSavedStack;
3162 ThisInitState pdThisPtrInit;
3165 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3166 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3168 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3169 ExpandArray<BYTE> impPendingBlockMembers;
3171 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3172 // Operates on the map in the top-level ancestor.
3173 BYTE impGetPendingBlockMember(BasicBlock* blk)
3175 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3178 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3179 // Operates on the map in the top-level ancestor.
3180 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3182 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3185 bool impCanReimport;
3187 bool impSpillStackEntry(unsigned level,
3191 bool bAssertOnRecursion,
3196 void impSpillStackEnsure(bool spillLeaves = false);
3197 void impEvalSideEffects();
3198 void impSpillSpecialSideEff();
3199 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3200 void impSpillValueClasses();
3201 void impSpillEvalStack();
3202 static fgWalkPreFn impFindValueClasses;
3203 void impSpillLclRefs(ssize_t lclNum);
3205 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3207 void impImportBlockCode(BasicBlock* block);
3209 void impReimportMarkBlock(BasicBlock* block);
3210 void impReimportMarkSuccessors(BasicBlock* block);
3212 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3214 void impImportBlockPending(BasicBlock* block);
3216 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3217 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3218 // for the block, but instead, just re-uses the block's existing EntryState.
3219 void impReimportBlockPending(BasicBlock* block);
3221 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3223 void impImportBlock(BasicBlock* block);
3225 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3226 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3227 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3228 // the variables that will be used -- and for all the predecessors of those successors, and the
3229 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3230 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3231 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3232 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3233 // of local variable numbers, so we represent them with the base local variable number), returns that.
3234 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3235 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3236 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3237 // on which kind of member of the clique the block is).
3238 unsigned impGetSpillTmpBase(BasicBlock* block);
3240 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3241 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3242 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3243 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3244 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3245 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3246 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3247 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3248 // successors receive a native int. Similarly float and double are unified to double.
3249 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3250 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3251 // predecessors, so they insert an upcast if needed).
3252 void impReimportSpillClique(BasicBlock* block);
3254 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3255 // block, and represent the predecessor and successor members of the clique currently being computed.
3256 // *** Access to these will need to be locked in a parallel compiler.
3257 ExpandArray<BYTE> impSpillCliquePredMembers;
3258 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3266 // Abstract class for receiving a callback while walking a spill clique
3267 class SpillCliqueWalker
3270 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3273 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3274 class SetSpillTempsBase : public SpillCliqueWalker
3279 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3282 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3285 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3286 class ReimportSpillClique : public SpillCliqueWalker
3291 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3294 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3297 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3298 // predecessor or successor within the spill clique
3299 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3301 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3302 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3303 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3304 void impRetypeEntryStateTemps(BasicBlock* blk);
3306 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3307 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3309 void impPushVar(GenTree* op, typeInfo tiRetVal);
3310 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3311 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3313 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3315 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3316 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3317 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3320 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3323 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3324 struct BlockListNode
3327 BlockListNode* m_next;
3328 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3331 void* operator new(size_t sz, Compiler* comp);
3333 BlockListNode* impBlockListNodeFreeList;
3335 BlockListNode* AllocBlockListNode();
3336 void FreeBlockListNode(BlockListNode* node);
3338 bool impIsValueType(typeInfo* pTypeInfo);
3339 var_types mangleVarArgsType(var_types type);
3342 regNumber getCallArgIntRegister(regNumber floatReg);
3343 regNumber getCallArgFloatRegister(regNumber intReg);
3344 #endif // FEATURE_VARARG
3347 static unsigned jitTotalMethodCompiled;
3351 static LONG jitNestingLevel;
3354 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3356 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3358 // STATIC inlining decision based on the IL code.
3359 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3360 CORINFO_METHOD_INFO* methInfo,
3362 InlineResult* inlineResult);
3364 void impCheckCanInline(GenTreePtr call,
3365 CORINFO_METHOD_HANDLE fncHandle,
3367 CORINFO_CONTEXT_HANDLE exactContextHnd,
3368 InlineCandidateInfo** ppInlineCandidateInfo,
3369 InlineResult* inlineResult);
3371 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3372 GenTreePtr curArgVal,
3374 InlineResult* inlineResult);
3376 void impInlineInitVars(InlineInfo* pInlineInfo);
3378 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3380 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3382 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3384 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3385 GenTreePtr variableBeingDereferenced,
3386 InlArgInfo* inlArgInfo);
3388 void impMarkInlineCandidate(GenTreePtr call,
3389 CORINFO_CONTEXT_HANDLE exactContextHnd,
3390 bool exactContextNeedsRuntimeLookup,
3391 CORINFO_CALL_INFO* callInfo);
3393 bool impTailCallRetTypeCompatible(var_types callerRetType,
3394 CORINFO_CLASS_HANDLE callerRetTypeClass,
3395 var_types calleeRetType,
3396 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3398 bool impIsTailCallILPattern(bool tailPrefixed,
3400 const BYTE* codeAddrOfNextOpcode,
3401 const BYTE* codeEnd,
3403 bool* IsCallPopRet = nullptr);
3405 bool impIsImplicitTailCallCandidate(
3406 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3408 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3411 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3412 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3416 XX Info about the basic-blocks, their contents and the flow analysis XX
3418 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3419 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3423 BasicBlock* fgFirstBB; // Beginning of the basic block list
3424 BasicBlock* fgLastBB; // End of the basic block list
3425 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3426 #if FEATURE_EH_FUNCLETS
3427 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3429 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3431 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3432 unsigned fgEdgeCount; // # of control flow edges between the BBs
3433 unsigned fgBBcount; // # of BBs in the method
3435 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3437 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3438 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3439 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3440 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3442 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3443 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3444 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3445 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3446 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3447 // index). The arrays are of size fgBBNumMax + 1.
3448 unsigned* fgDomTreePreOrder;
3449 unsigned* fgDomTreePostOrder;
3451 bool fgBBVarSetsInited;
3453 // Allocate array like T* a = new T[fgBBNumMax + 1];
3454 // Using helper so we don't keep forgetting +1.
3455 template <typename T>
3456 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3458 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3461 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3462 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3463 // cannot be meaningfully combined. Note that new blocks can be created with higher
3464 // block numbers without changing the basic block epoch. These blocks *cannot*
3465 // participate in a block set until the blocks are all renumbered, causing the epoch
3466 // to change. This is useful if continuing to use previous block sets is valuable.
3467 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3468 unsigned fgCurBBEpoch;
3470 unsigned GetCurBasicBlockEpoch()
3472 return fgCurBBEpoch;
3475 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3476 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3477 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3478 unsigned fgCurBBEpochSize;
3480 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3481 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3482 unsigned fgBBSetCountInSizeTUnits;
3484 void NewBasicBlockEpoch()
3486 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3488 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3490 fgCurBBEpochSize = fgBBNumMax + 1;
3491 fgBBSetCountInSizeTUnits =
3492 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3495 // All BlockSet objects are now invalid!
3496 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3497 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3501 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3502 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3503 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3504 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3506 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3507 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3508 // array of size_t bitsets), then print that out.
3509 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3516 void EnsureBasicBlockEpoch()
3518 if (fgCurBBEpochSize != fgBBNumMax + 1)
3520 NewBasicBlockEpoch();
3524 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3525 void fgEnsureFirstBBisScratch();
3526 bool fgFirstBBisScratch();
3527 bool fgBBisScratch(BasicBlock* block);
3529 void fgExtendEHRegionBefore(BasicBlock* block);
3530 void fgExtendEHRegionAfter(BasicBlock* block);
3532 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3534 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3536 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3539 BasicBlock* nearBlk,
3540 bool putInFilter = false,
3541 bool runRarely = false,
3542 bool insertAtEnd = false);
3544 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3546 bool runRarely = false,
3547 bool insertAtEnd = false);
3549 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3551 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3552 BasicBlock* afterBlk,
3553 unsigned xcptnIndex,
3554 bool putInTryRegion);
3556 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3557 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3558 void fgUnlinkBlock(BasicBlock* block);
3560 unsigned fgMeasureIR();
3562 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3563 bool fgMultipleNots;
3566 bool fgModified; // True if the flow graph has been modified recently
3567 bool fgComputePredsDone; // Have we computed the bbPreds list
3568 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3569 bool fgDomsComputed; // Have we computed the dominator sets?
3570 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3572 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3573 bool fgHasPostfix; // any postfix ++/-- found?
3574 unsigned fgIncrCount; // number of increment nodes found
3576 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3580 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3581 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3584 bool fgRemoveRestOfBlock; // true if we know that we will throw
3585 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3587 // There are two modes for ordering of the trees.
3588 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3589 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3590 // by traversing the tree according to the order of the operands.
3591 // - In FGOrderLinear, the dominant ordering is the linear order.
3598 FlowGraphOrder fgOrder;
3600 // The following are boolean flags that keep track of the state of internal data structures
3602 bool fgStmtListThreaded;
3603 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3604 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3605 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3606 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3607 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3608 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3609 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3610 // This is derived from the profile data
3611 // or is BB_UNITY_WEIGHT when we don't have profile data
3613 #if FEATURE_EH_FUNCLETS
3614 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3615 #endif // FEATURE_EH_FUNCLETS
3617 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3618 // since fgMorphTree can be called from several places
3619 bool fgExpandInline; // indicates that we are creating tree for the inliner
3621 bool impBoxTempInUse; // the temp below is valid and available
3622 unsigned impBoxTemp; // a temporary that is used for boxing
3625 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3626 // and we are trying to compile again in a "safer", minopts mode?
3630 unsigned impInlinedCodeSize;
3633 //-------------------------------------------------------------------------
3639 void fgTransformFatCalli();
3643 void fgRemoveEmptyTry();
3645 void fgRemoveEmptyFinally();
3647 void fgMergeFinallyChains();
3649 void fgCloneFinally();
3651 void fgCleanupContinuation(BasicBlock* continuation);
3653 void fgUpdateFinallyTargetFlags();
3655 void fgClearAllFinallyTargetBits();
3657 void fgAddFinallyTargetFlags();
3659 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3660 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3661 // when this is necessary.
3662 bool fgNeedToAddFinallyTargetBits;
3663 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3665 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3666 BasicBlock* handler,
3667 BlockToBlockMap& continuationMap);
3669 GenTreePtr fgGetCritSectOfStaticMethod();
3671 #if FEATURE_EH_FUNCLETS
3673 void fgAddSyncMethodEnterExit();
3675 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3677 void fgConvertSyncReturnToLeave(BasicBlock* block);
3679 #endif // FEATURE_EH_FUNCLETS
3681 void fgAddReversePInvokeEnterExit();
3683 bool fgMoreThanOneReturnBlock();
3685 // The number of separate return points in the method.
3686 unsigned fgReturnCount;
3688 void fgAddInternal();
3690 bool fgFoldConditional(BasicBlock* block);
3692 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3693 void fgMorphBlocks();
3695 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3697 void fgCheckArgCnt();
3698 void fgSetOptions();
3701 static fgWalkPreFn fgAssertNoQmark;
3702 void fgPreExpandQmarkChecks(GenTreePtr expr);
3703 void fgPostExpandQmarkChecks();
3704 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3707 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3709 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3710 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3711 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3712 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3713 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3715 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3716 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3717 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3718 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3720 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3721 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3722 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3723 void fgExpandQmarkNodes();
3727 // Do "simple lowering." This functionality is (conceptually) part of "general"
3728 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3729 void fgSimpleLowering();
3731 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3733 GenTreePtr fgInitThisClass();
3735 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3737 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3739 inline bool backendRequiresLocalVarLifetimes()
3741 #if defined(LEGACY_BACKEND)
3744 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3748 void fgLocalVarLiveness();
3750 void fgLocalVarLivenessInit();
3752 #ifdef LEGACY_BACKEND
3753 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3755 void fgPerNodeLocalVarLiveness(GenTree* node);
3757 void fgPerBlockLocalVarLiveness();
3759 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3761 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3763 // This is used in the liveness computation, as a temporary. When we use the
3764 // arbitrary-length VarSet representation, it is better not to allocate a new one
3766 VARSET_TP fgMarkIntfUnionVS;
3768 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3770 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3772 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3774 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3776 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3778 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3780 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3782 void fgComputeLife(VARSET_TP& life,
3783 GenTreePtr startNode,
3785 VARSET_VALARG_TP volatileVars,
3786 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3788 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3790 bool fgRemoveDeadStore(GenTree** pTree,
3792 VARSET_VALARG_TP life,
3794 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3796 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3798 // For updating liveset during traversal AFTER fgComputeLife has completed
3799 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3800 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3802 // Returns the set of live variables after endTree,
3803 // assuming that liveSet is the set of live variables BEFORE tree.
3804 // Requires that fgComputeLife has completed, and that tree is in the same
3805 // statement as endTree, and that it comes before endTree in execution order
3807 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3809 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3810 while (tree != nullptr && tree != endTree->gtNext)
3812 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3813 tree = tree->gtNext;
3815 assert(tree == endTree->gtNext);
3819 void fgInterBlockLocalVarLiveness();
3821 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3822 // "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
3823 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3824 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3825 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3826 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3827 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3829 if (m_opAsgnVarDefSsaNums == nullptr)
3831 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3833 return m_opAsgnVarDefSsaNums;
3836 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3837 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3838 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3840 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3842 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3843 // Except: assumes that lcl is a def, and if it is
3844 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3845 // rather than the "use" SSA number recorded in the tree "lcl".
3846 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3848 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3849 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3850 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3851 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3852 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3854 // (byref addrS1 = &s1,
3855 // *(addrS1 * offsetof(f0)) = s2f0,
3857 // *(addrS1 * offsetof(fn)) = s2fn)
3859 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3860 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3861 // give it SSA names and value numbers?
3863 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3864 // end with an instance of the structure below, whose fields are described in the declaration.
3865 struct IndirectAssignmentAnnotation
3867 unsigned m_lclNum; // The local num that is being indirectly assigned.
3868 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3869 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3870 // be the singleton field sequence "g". The individual assignments would
3871 // further append the fields of "s.g" to that.
3872 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3873 // structure has a single field).
3874 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3875 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3878 IndirectAssignmentAnnotation(unsigned lclNum,
3879 FieldSeqNode* fldSeq,
3881 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3882 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3883 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3887 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3888 NodeToIndirAssignMap;
3889 NodeToIndirAssignMap* m_indirAssignMap;
3890 NodeToIndirAssignMap* GetIndirAssignMap()
3892 if (m_indirAssignMap == nullptr)
3894 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3895 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3896 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3898 return m_indirAssignMap;
3901 // Performs SSA conversion.
3904 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3905 void fgResetForSsa();
3907 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3909 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3910 inline bool fgExcludeFromSsa(unsigned lclNum);
3912 // The value numbers for this compilation.
3913 ValueNumStore* vnStore;
3916 ValueNumStore* GetValueNumStore()
3921 // Do value numbering (assign a value number to each
3923 void fgValueNumber();
3925 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3926 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3927 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3928 // match the element type of the array or fldSeq. When this type doesn't match
3929 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3931 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3934 FieldSeqNode* fldSeq,
3938 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3939 // has been parsed to yield the other input arguments. If evaluation of the address
3940 // can raise exceptions, those should be captured in the exception set "excVN."
3941 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3942 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3943 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3944 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3945 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3947 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3948 CORINFO_CLASS_HANDLE elemTypeEq,
3952 FieldSeqNode* fldSeq);
3954 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3955 // by evaluating the array index expression "tree". Returns the value number resulting from
3956 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3957 // "GT_IND" that does the dereference, and it is given the returned value number.
3958 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3960 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3961 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3963 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3965 // Utility functions for fgValueNumber.
3967 // Perform value-numbering for the trees in "blk".
3968 void fgValueNumberBlock(BasicBlock* blk);
3970 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3971 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3972 // assumed for the memoryKind at the start "entryBlk".
3973 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3975 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3976 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3977 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3979 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3981 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3983 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3984 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3985 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3987 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3988 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3990 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3991 // value in that SSA #.
3992 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3994 // The input 'tree' is a leaf node that is a constant
3995 // Assign the proper value number to the tree
3996 void fgValueNumberTreeConst(GenTreePtr tree);
3998 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3999 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4001 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4003 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
4005 // Does value-numbering for a block assignment.
4006 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4008 // Does value-numbering for a cast tree.
4009 void fgValueNumberCastTree(GenTreePtr tree);
4011 // Does value-numbering for an intrinsic tree.
4012 void fgValueNumberIntrinsic(GenTreePtr tree);
4014 // Does value-numbering for a call. We interpret some helper calls.
4015 void fgValueNumberCall(GenTreeCall* call);
4017 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4018 void fgUpdateArgListVNs(GenTreeArgList* args);
4020 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4021 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4023 // Requires "helpCall" to be a helper call. Assigns it a value number;
4024 // we understand the semantics of some of the calls. Returns "true" if
4025 // the call may modify the heap (we assume arbitrary memory side effects if so).
4026 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4028 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4029 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4031 // These are the current value number for the memory implicit variables while
4032 // doing value numbering. These are the value numbers under the "liberal" interpretation
4033 // of memory values; the "conservative" interpretation needs no VN, since every access of
4034 // memory yields an unknown value.
4035 ValueNum fgCurMemoryVN[MemoryKindCount];
4037 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4038 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4039 // is 1, and the rest is an encoding of "elemTyp".
4040 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4042 if (elemStructType != nullptr)
4044 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4045 varTypeIsIntegral(elemTyp));
4046 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4047 return elemStructType;
4051 elemTyp = varTypeUnsignedToSigned(elemTyp);
4052 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4055 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4056 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4057 // the struct type of the element).
4058 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4060 size_t clsHndVal = size_t(clsHnd);
4061 if (clsHndVal & 0x1)
4063 return var_types(clsHndVal >> 1);
4071 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4072 var_types getJitGCType(BYTE gcType);
4074 enum structPassingKind
4076 SPK_Unknown, // Invalid value, never returned
4077 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4078 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4079 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4080 // parameters registers are used, then the stack will be used)
4081 // for X86 passed on the stack, for ARM32 passed in registers
4082 // or the stack or split between registers and the stack.
4083 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4085 }; // The struct is passed/returned by reference to a copy/buffer.
4087 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4088 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4089 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4090 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4092 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4094 // Get the type that is used to pass 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 getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4098 structPassingKind* wbPassStruct,
4099 unsigned structSize = 0);
4101 // Get the type that is used to return values of the given struct type.
4102 // If you have already retrieved the struct size then pass it as the optional third argument
4104 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4105 structPassingKind* wbPassStruct = nullptr,
4106 unsigned structSize = 0);
4109 // Print a representation of "vnp" or "vn" on standard output.
4110 // If "level" is non-zero, we also print out a partial expansion of the value.
4111 void vnpPrint(ValueNumPair vnp, unsigned level);
4112 void vnPrint(ValueNum vn, unsigned level);
4115 // Dominator computation member functions
4116 // Not exposed outside Compiler
4118 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4120 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4122 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4123 // flow graph. We first assume the fields bbIDom on each
4124 // basic block are invalid. This computation is needed later
4125 // by fgBuildDomTree to build the dominance tree structure.
4126 // Based on: A Simple, Fast Dominance Algorithm
4127 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4129 void fgCompDominatedByExceptionalEntryBlocks();
4131 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4132 // Note: this is relatively slow compared to calling fgDominate(),
4133 // especially if dealing with a single block versus block check.
4135 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4137 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4139 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4141 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4143 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4145 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4146 // processed in topological sort, this function takes care of that.
4148 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4150 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4151 // Returns this as a set.
4153 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4154 // root nodes. Returns this as a set.
4157 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4160 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4161 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4164 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4165 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4166 // && postOrder(A) >= postOrder(B) making the computation O(1).
4167 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4169 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4171 void fgUpdateChangedFlowGraph();
4174 // Compute the predecessors of the blocks in the control flow graph.
4175 void fgComputePreds();
4177 // Remove all predecessor information.
4178 void fgRemovePreds();
4180 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4181 // before the full predecessors lists are computed.
4182 void fgComputeCheapPreds();
4185 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4187 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4197 // Initialize the per-block variable sets (used for liveness analysis).
4198 void fgInitBlockVarSets();
4200 // true if we've gone through and created GC Poll calls.
4201 bool fgGCPollsCreated;
4202 void fgMarkGCPollBlocks();
4203 void fgCreateGCPolls();
4204 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4206 // Requires that "block" is a block that returns from
4207 // a finally. Returns the number of successors (jump targets of
4208 // of blocks in the covered "try" that did a "LEAVE".)
4209 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4211 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4212 // a finally. Returns its "i"th successor (jump targets of
4213 // of blocks in the covered "try" that did a "LEAVE".)
4214 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4215 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4218 // Factor out common portions of the impls of the methods above.
4219 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4222 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4223 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4224 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4225 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4226 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4227 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4228 // we leave the entry associated with the block, but it will no longer be accessed.)
4229 struct SwitchUniqueSuccSet
4231 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4232 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4235 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4236 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4237 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4238 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4241 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4242 BlockToSwitchDescMap;
4245 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4246 // iteration over only the distinct successors.
4247 BlockToSwitchDescMap* m_switchDescMap;
4250 BlockToSwitchDescMap* GetSwitchDescMap()
4252 if (m_switchDescMap == nullptr)
4254 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4256 return m_switchDescMap;
4259 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4260 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4261 // we don't accidentally look up and return the wrong switch data.
4262 void InvalidateUniqueSwitchSuccMap()
4264 m_switchDescMap = nullptr;
4267 // Requires "switchBlock" to be a block that ends in a switch. Returns
4268 // the corresponding SwitchUniqueSuccSet.
4269 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4271 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4272 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4273 // remove it from "this", and ensure that "to" is a member.
4274 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4276 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4277 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4279 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4281 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4283 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4285 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4287 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4289 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4291 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4293 void fgRemoveBlockAsPred(BasicBlock* block);
4295 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4297 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4299 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4301 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4303 flowList* fgAddRefPred(BasicBlock* block,
4304 BasicBlock* blockPred,
4305 flowList* oldEdge = nullptr,
4306 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4309 void fgFindBasicBlocks();
4311 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4313 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4315 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4316 bool putInTryRegion,
4317 BasicBlock* startBlk,
4319 BasicBlock* nearBlk,
4320 BasicBlock* jumpBlk,
4323 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4325 void fgRemoveEmptyBlocks();
4327 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4329 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4331 void fgCreateLoopPreHeader(unsigned lnum);
4333 void fgUnreachableBlock(BasicBlock* block);
4335 void fgRemoveConditionalJump(BasicBlock* block);
4337 BasicBlock* fgLastBBInMainFunction();
4339 BasicBlock* fgEndBBAfterMainFunction();
4341 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4343 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4345 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4347 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4349 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4351 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4353 bool fgRenumberBlocks();
4355 bool fgExpandRarelyRunBlocks();
4357 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4359 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4361 enum FG_RELOCATE_TYPE
4363 FG_RELOCATE_TRY, // relocate the 'try' region
4364 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4366 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4368 #if FEATURE_EH_FUNCLETS
4369 #if defined(_TARGET_ARM_)
4370 void fgClearFinallyTargetBit(BasicBlock* block);
4371 #endif // defined(_TARGET_ARM_)
4372 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4373 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4374 void fgInsertFuncletPrologBlock(BasicBlock* block);
4375 void fgCreateFuncletPrologBlocks();
4376 void fgCreateFunclets();
4377 #else // !FEATURE_EH_FUNCLETS
4378 bool fgRelocateEHRegions();
4379 #endif // !FEATURE_EH_FUNCLETS
4381 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4383 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4385 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4387 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4389 bool fgOptimizeEmptyBlock(BasicBlock* block);
4391 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4393 bool fgOptimizeBranch(BasicBlock* bJump);
4395 bool fgOptimizeSwitchBranches(BasicBlock* block);
4397 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4399 bool fgOptimizeSwitchJumps();
4401 void fgPrintEdgeWeights();
4403 void fgComputeEdgeWeights();
4405 void fgReorderBlocks();
4407 void fgDetermineFirstColdBlock();
4409 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4411 bool fgUpdateFlowGraph(bool doTailDup = false);
4413 void fgFindOperOrder();
4415 // method that returns if you should split here
4416 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4418 void fgSetBlockOrder();
4420 void fgRemoveReturnBlock(BasicBlock* block);
4422 /* Helper code that has been factored out */
4423 inline void fgConvertBBToThrowBB(BasicBlock* block);
4425 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4426 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4427 GenTreePtr fgMakeTmpArgNode(
4428 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4430 // The following check for loops that don't execute calls
4431 bool fgLoopCallMarked;
4433 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4434 void fgLoopCallMark();
4436 void fgMarkLoopHead(BasicBlock* block);
4438 unsigned fgGetCodeEstimate(BasicBlock* block);
4441 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4442 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4443 bool fgDumpFlowGraph(Phases phase);
4445 #endif // DUMP_FLOWGRAPHS
4450 void fgDispBBLiveness(BasicBlock* block);
4451 void fgDispBBLiveness();
4452 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4453 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4454 void fgDispBasicBlocks(bool dumpTrees = false);
4455 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4456 void fgDumpBlock(BasicBlock* block);
4457 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4459 static fgWalkPreFn fgStress64RsltMulCB;
4460 void fgStress64RsltMul();
4461 void fgDebugCheckUpdate();
4462 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4463 void fgDebugCheckBlockLinks();
4464 void fgDebugCheckLinks(bool morphTrees = false);
4465 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4466 void fgDebugCheckFlags(GenTreePtr tree);
4467 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4468 void fgDebugCheckTryFinallyExits();
4471 #ifdef LEGACY_BACKEND
4472 static void fgOrderBlockOps(GenTreePtr tree,
4476 GenTreePtr* opsPtr, // OUT
4477 regMaskTP* regsPtr); // OUT
4478 #endif // LEGACY_BACKEND
4480 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4481 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4483 inline bool fgIsInlining()
4485 return fgExpandInline;
4488 void fgTraverseRPO();
4490 //--------------------- Walking the trees in the IR -----------------------
4495 fgWalkPreFn* wtprVisitorFn;
4496 fgWalkPostFn* wtpoVisitorFn;
4497 void* pCallbackData; // user-provided data
4498 bool wtprLclsOnly; // whether to only visit lclvar nodes
4499 GenTreePtr parent; // parent of current node, provided to callback
4500 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4502 bool printModified; // callback can use this
4506 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4507 fgWalkPreFn* visitor,
4508 void* pCallBackData = nullptr,
4509 bool lclVarsOnly = false,
4510 bool computeStack = false);
4512 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4513 fgWalkPreFn* preVisitor,
4514 fgWalkPostFn* postVisitor,
4515 void* pCallBackData = nullptr);
4517 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4521 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4522 fgWalkPostFn* visitor,
4523 void* pCallBackData = nullptr,
4524 bool computeStack = false);
4526 // An fgWalkPreFn that looks for expressions that have inline throws in
4527 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4528 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4529 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4530 // properly propagated to parent trees). It returns WALK_CONTINUE
4532 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4533 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4534 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4536 /**************************************************************************
4538 *************************************************************************/
4541 friend class SsaBuilder;
4542 friend struct ValueNumberState;
4544 //--------------------- Detect the basic blocks ---------------------------
4546 BasicBlock** fgBBs; // Table of pointers to the BBs
4548 void fgInitBBLookup();
4549 BasicBlock* fgLookupBB(unsigned addr);
4551 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4553 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4555 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4557 void fgLinkBasicBlocks();
4559 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4561 void fgCheckBasicBlockControlFlow();
4563 void fgControlFlowPermitted(BasicBlock* blkSrc,
4564 BasicBlock* blkDest,
4565 BOOL IsLeave = false /* is the src a leave block */);
4567 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4569 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4571 void fgAdjustForAddressExposedOrWrittenThis();
4573 bool fgProfileData_ILSizeMismatch;
4574 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4575 ULONG fgProfileBufferCount;
4576 ULONG fgNumProfileRuns;
4578 unsigned fgStressBBProf()
4581 unsigned result = JitConfig.JitStressBBProf();
4584 if (compStressCompile(STRESS_BB_PROFILE, 15))
4595 bool fgHaveProfileData();
4596 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4597 void fgInstrumentMethod();
4600 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4601 // or if we have some fake profile data for the stress mode
4602 bool fgIsUsingProfileWeights()
4604 return (fgHaveProfileData() || fgStressBBProf());
4607 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4608 // or BB_UNITY_WEIGHT when we aren't using profile data.
4609 unsigned fgProfileRunsCount()
4611 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4614 //-------- Insert a statement at the start or end of a basic block --------
4618 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4622 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4624 public: // Used by linear scan register allocation
4625 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4628 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4629 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4631 public: // Used by linear scan register allocation
4632 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4635 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4637 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4639 // Create a new temporary variable to hold the result of *ppTree,
4640 // and transform the graph accordingly.
4641 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4642 GenTree* fgMakeMultiUse(GenTree** ppTree);
4645 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4646 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4647 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4649 //-------- Determine the order in which the trees will be evaluated -------
4651 unsigned fgTreeSeqNum;
4652 GenTree* fgTreeSeqLst;
4653 GenTree* fgTreeSeqBeg;
4655 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4656 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4657 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4658 void fgSetStmtSeq(GenTree* tree);
4659 void fgSetBlockOrder(BasicBlock* block);
4661 //------------------------- Morphing --------------------------------------
4663 unsigned fgPtrArgCntCur;
4664 unsigned fgPtrArgCntMax;
4665 hashBv* fgOutgoingArgTemps;
4666 hashBv* fgCurrentlyInUseArgTemps;
4668 bool compCanEncodePtrArgCntMax();
4670 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4673 void fgMoveOpsLeft(GenTreePtr tree);
4676 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4678 bool fgIsThrow(GenTreePtr tree);
4680 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4681 bool fgIsBlockCold(BasicBlock* block);
4683 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4685 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4687 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4689 bool fgMorphRelopToQmark(GenTreePtr tree);
4691 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4692 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4693 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4694 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4695 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4696 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4697 // small; hence the other fields of MorphAddrContext.
4698 enum MorphAddrContextKind
4703 struct MorphAddrContext
4705 MorphAddrContextKind m_kind;
4706 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4707 // top-level indirection and here have been constants.
4708 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4709 // In that case, is the sum of those constant offsets.
4711 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4716 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4717 static MorphAddrContext s_CopyBlockMAC;
4720 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4721 var_types* baseTypeOut,
4723 unsigned* simdSizeOut,
4724 bool ignoreUsedInSIMDIntrinsic = false);
4725 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4726 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4727 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4728 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4730 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4731 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4732 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4734 #endif // FEATURE_SIMD
4735 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4736 GenTreePtr fgMorphCast(GenTreePtr tree);
4737 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4738 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4740 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4743 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4744 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4746 void fgFixupStructReturn(GenTreePtr call);
4747 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4748 bool fgAddrCouldBeNull(GenTreePtr addr);
4749 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4750 bool fgCanFastTailCall(GenTreeCall* call);
4751 void fgMorphTailCall(GenTreeCall* call);
4752 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4753 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4754 fgArgTabEntryPtr argTabEntry,
4756 IL_OFFSETX callILOffset,
4757 GenTreePtr tmpAssignmentInsertionPoint,
4758 GenTreePtr paramAssignmentInsertionPoint);
4759 static int fgEstimateCallStackSize(GenTreeCall* call);
4760 GenTreePtr fgMorphCall(GenTreeCall* call);
4761 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4762 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4764 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4765 static fgWalkPreFn fgFindNonInlineCandidate;
4767 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4768 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4769 CORINFO_RESOLVED_TOKEN* ldftnToken);
4770 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4771 void fgAssignSetVarDef(GenTreePtr tree);
4772 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4773 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4774 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4775 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4776 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4777 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4778 void fgMorphUnsafeBlk(GenTreeObj* obj);
4779 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4780 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4781 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4782 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4783 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4784 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4785 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4787 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4788 GenTreePtr fgMorphConst(GenTreePtr tree);
4791 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4794 #if LOCAL_ASSERTION_PROP
4795 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4796 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4798 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4800 GenTreeStmt* fgMorphStmt;
4802 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4803 // used when morphing big offset.
4805 //----------------------- Liveness analysis -------------------------------
4807 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4808 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4810 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4811 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4812 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4814 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4816 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4818 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4819 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4821 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4822 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4824 void fgExtendDbgScopes();
4825 void fgExtendDbgLifetimes();
4828 void fgDispDebugScopes();
4831 //-------------------------------------------------------------------------
4833 // The following keeps track of any code we've added for things like array
4834 // range checking or explicit calls to enable GC, and so on.
4839 AddCodeDsc* acdNext;
4840 BasicBlock* acdDstBlk; // block to which we jump
4842 SpecialCodeKind acdKind; // what kind of a special block is this?
4843 unsigned short acdStkLvl;
4847 static unsigned acdHelper(SpecialCodeKind codeKind);
4849 AddCodeDsc* fgAddCodeList;
4851 bool fgRngChkThrowAdded;
4852 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4854 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4856 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4859 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4862 bool fgIsCodeAdded();
4864 bool fgIsThrowHlpBlk(BasicBlock* block);
4865 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4867 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4869 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4870 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4871 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4872 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4873 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4875 #if FEATURE_MULTIREG_RET
4876 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4877 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4878 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4879 #endif // FEATURE_MULTIREG_RET
4881 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4884 static fgWalkPreFn fgDebugCheckInlineCandidates;
4886 void CheckNoFatPointerCandidatesLeft();
4887 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4890 void fgPromoteStructs();
4891 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4892 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4894 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4895 void fgMarkImplicitByRefArgs();
4897 // Change implicit byrefs' types from struct to pointer, and for any that were
4898 // promoted, create new promoted struct temps.
4899 void fgRetypeImplicitByRefArgs();
4901 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4902 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4903 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4905 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4906 void fgMarkDemotedImplicitByRefArgs();
4908 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4909 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4910 void fgMarkAddressExposedLocals();
4911 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4913 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4915 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4917 // The given local variable, required to be a struct variable, is being assigned via
4918 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4919 // the variable is not enregistered, and is therefore not promoted independently.
4920 void fgLclFldAssign(unsigned lclNum);
4922 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4923 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4924 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4925 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4928 bool fgPrintInlinedMethods;
4931 bool fgIsBigOffset(size_t offset);
4933 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4934 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4935 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4936 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4937 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4940 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4941 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4945 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4946 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4953 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4956 void optRemoveRangeCheck(
4957 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4958 bool optIsRangeCheckRemovable(GenTreePtr tree);
4961 static fgWalkPreFn optValidRangeCheckIndex;
4962 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4965 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4967 /**************************************************************************
4969 *************************************************************************/
4972 // Do hoisting for all loops.
4973 void optHoistLoopCode();
4975 // To represent sets of VN's that have already been hoisted in outer loops.
4976 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4977 typedef VNToBoolMap VNSet;
4979 struct LoopHoistContext
4982 // The set of variables hoisted in the current loop (or nullptr if there are none).
4983 VNSet* m_pHoistedInCurLoop;
4986 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4987 VNSet m_hoistedInParentLoops;
4988 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4989 // Previous decisions on loop-invariance of value numbers in the current loop.
4990 VNToBoolMap m_curLoopVnInvariantCache;
4992 VNSet* GetHoistedInCurLoop(Compiler* comp)
4994 if (m_pHoistedInCurLoop == nullptr)
4996 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4998 return m_pHoistedInCurLoop;
5001 VNSet* ExtractHoistedInCurLoop()
5003 VNSet* res = m_pHoistedInCurLoop;
5004 m_pHoistedInCurLoop = nullptr;
5008 LoopHoistContext(Compiler* comp)
5009 : m_pHoistedInCurLoop(nullptr)
5010 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5011 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5016 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5017 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5018 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5019 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5021 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5022 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5023 // "m_hoistedInParentLoops".
5025 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5027 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5028 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5029 // expressions to "hoistInLoop".
5030 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5032 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5033 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5035 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5036 // that are invariant in loop "lnum" (an index into the optLoopTable)
5037 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5038 // expressions to "hoistInLoop".
5039 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5040 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5041 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5042 bool optHoistLoopExprsForTree(GenTreePtr tree,
5044 LoopHoistContext* hoistCtxt,
5045 bool* firstBlockAndBeforeSideEffect,
5047 bool* pCctorDependent);
5049 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5050 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5052 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5053 // Constants and init values are always loop invariant.
5054 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5055 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5057 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5058 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5059 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5060 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5061 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5063 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5064 // in the loop table.
5065 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5067 // Records the set of "side effects" of all loops: fields (object instance and static)
5068 // written to, and SZ-array element type equivalence classes updated.
5069 void optComputeLoopSideEffects();
5072 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5073 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5074 // static) written to, and SZ-array element type equivalence classes updated.
5075 void optComputeLoopNestSideEffects(unsigned lnum);
5077 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5078 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5080 // Hoist the expression "expr" out of loop "lnum".
5081 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5084 void optOptimizeBools();
5087 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5089 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5092 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5094 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5095 // the loop into a "do-while" loop
5096 // Also finds all natural loops and records them in the loop table
5098 // Optionally clone loops in the loop table.
5099 void optCloneLoops();
5101 // Clone loop "loopInd" in the loop table.
5102 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5104 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5105 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5106 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5108 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5110 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5113 // This enumeration describes what is killed by a call.
5117 CALLINT_NONE, // no interference (most helpers)
5118 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5119 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5120 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5121 CALLINT_ALL, // kills everything (normal method call)
5125 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5126 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5127 // in bbNext order; we use comparisons on the bbNum to decide order.)
5128 // The blocks that define the body are
5129 // first <= top <= entry <= bottom .
5130 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5131 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5132 // Compiler::optFindNaturalLoops().
5135 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5136 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5137 // loop, but not the outer loop.)
5138 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5140 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5141 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5142 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5144 callInterf lpAsgCall; // "callInterf" for calls in the loop
5145 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5146 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5148 unsigned short lpFlags; // Mask of the LPFLG_* constants
5150 unsigned char lpExitCnt; // number of exits from the loop
5152 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5153 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5154 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5155 // (Actually, an "immediately" nested loop --
5156 // no other child of this loop is a parent of lpChild.)
5157 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5158 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5159 // by following "lpChild" then "lpSibling" links.
5161 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5162 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5164 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5165 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5166 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5168 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5169 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5171 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5172 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5173 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5174 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5176 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5177 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5178 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5180 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5181 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5182 // type are assigned to.
5184 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5185 // memory side effects. If this is set, the fields below
5186 // may not be accurate (since they become irrelevant.)
5187 bool lpContainsCall; // True if executing the loop body *may* execute a call
5189 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5190 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5192 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5194 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5195 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5197 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5199 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5200 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5202 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5203 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5205 JitSimplerHashBehavior>
5207 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5208 // instance fields modified
5211 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5212 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5214 JitSimplerHashBehavior>
5216 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5217 // arrays of that type are modified
5220 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5221 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5223 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5224 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5225 // (shifted left, with a low-order bit set to distinguish.)
5226 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5227 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5229 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5231 GenTreePtr lpIterTree; // The "i <op>= const" tree
5232 unsigned lpIterVar(); // iterator variable #
5233 int lpIterConst(); // the constant with which the iterator is incremented
5234 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5235 void VERIFY_lpIterTree();
5237 var_types lpIterOperType(); // For overflow instructions
5240 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5241 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5245 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5247 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5248 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5249 void VERIFY_lpTestTree();
5251 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5252 GenTreePtr lpIterator(); // the iterator node in the loop test
5253 GenTreePtr lpLimit(); // the limit node in the loop test
5255 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5256 // LPFLG_CONST_LIMIT
5257 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5259 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5260 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5261 // LPFLG_ARRLEN_LIMIT
5263 // Returns "true" iff "*this" contains the blk.
5264 bool lpContains(BasicBlock* blk)
5266 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5268 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5269 // to be equal, but requiring bottoms to be different.)
5270 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5272 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5275 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5276 // bottoms to be different.)
5277 bool lpContains(const LoopDsc& lp2)
5279 return lpContains(lp2.lpFirst, lp2.lpBottom);
5282 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5283 // (allowing firsts to be equal, but requiring bottoms to be different.)
5284 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5286 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5289 // Returns "true" iff "*this" is (properly) contained by "lp2"
5290 // (allowing firsts to be equal, but requiring bottoms to be different.)
5291 bool lpContainedBy(const LoopDsc& lp2)
5293 return lpContains(lp2.lpFirst, lp2.lpBottom);
5296 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5297 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5299 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5301 // Returns "true" iff "*this" is disjoint from "lp2".
5302 bool lpDisjoint(const LoopDsc& lp2)
5304 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5306 // Returns "true" iff the loop is well-formed (see code for defn).
5309 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5310 lpEntry->bbNum <= lpBottom->bbNum &&
5311 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5316 bool fgMightHaveLoop(); // returns true if there are any backedges
5317 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5320 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5321 unsigned char optLoopCount; // number of tracked loops
5324 unsigned optCallCount; // number of calls made in the method
5325 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5326 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5327 unsigned optLoopsCloned; // number of loops cloned in the current method.
5330 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5331 void optPrintLoopInfo(unsigned loopNum,
5333 BasicBlock* lpFirst,
5335 BasicBlock* lpEntry,
5336 BasicBlock* lpBottom,
5337 unsigned char lpExitCnt,
5339 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5340 void optPrintLoopInfo(unsigned lnum);
5341 void optPrintLoopRecording(unsigned lnum);
5343 void optCheckPreds();
5346 void optSetBlockWeights();
5348 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5350 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5352 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5354 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5355 unsigned optIsLoopIncrTree(GenTreePtr incr);
5356 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5357 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5358 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5359 bool optExtractInitTestIncr(BasicBlock* head,
5364 GenTreePtr* ppIncr);
5366 void optRecordLoop(BasicBlock* head,
5372 unsigned char exitCnt);
5374 void optFindNaturalLoops();
5376 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5377 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5378 bool optCanonicalizeLoopNest(unsigned char loopInd);
5380 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5381 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5382 bool optCanonicalizeLoop(unsigned char loopInd);
5384 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5385 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5386 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5387 bool optLoopContains(unsigned l1, unsigned l2);
5389 // Requires "loopInd" to be a valid index into the loop table.
5390 // Updates the loop table by changing loop "loopInd", whose head is required
5391 // to be "from", to be "to". Also performs this transformation for any
5392 // loop nested in "loopInd" that shares the same head as "loopInd".
5393 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5395 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5396 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5397 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5399 // Marks the containsCall information to "lnum" and any parent loops.
5400 void AddContainsCallAllContainingLoops(unsigned lnum);
5401 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5402 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5403 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5404 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5405 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5406 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5408 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5409 // of "from".) Copies the jump destination from "from" to "to".
5410 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5412 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5413 unsigned optLoopDepth(unsigned lnum)
5415 unsigned par = optLoopTable[lnum].lpParent;
5416 if (par == BasicBlock::NOT_IN_LOOP)
5422 return 1 + optLoopDepth(par);
5426 void fgOptWhileLoop(BasicBlock* block);
5428 bool optComputeLoopRep(int constInit,
5431 genTreeOps iterOper,
5433 genTreeOps testOper,
5436 unsigned* iterCount);
5437 #if FEATURE_STACK_FP_X87
5440 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5441 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5442 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5443 #endif // FEATURE_STACK_FP_X87
5446 static fgWalkPreFn optIsVarAssgCB;
5449 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5451 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5453 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5455 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5457 /**************************************************************************
5458 * Optimization conditions
5459 *************************************************************************/
5461 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5462 bool optPentium4(void);
5463 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5464 bool optAvoidIntMult(void);
5469 // The following is the upper limit on how many expressions we'll keep track
5470 // of for the CSE analysis.
5472 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5474 static const int MIN_CSE_COST = 2;
5476 // Keeps tracked cse indices
5477 BitVecTraits* cseTraits;
5480 /* Generic list of nodes - used by the CSE logic */
5488 typedef struct treeLst* treeLstPtr;
5492 treeStmtLst* tslNext;
5493 GenTreePtr tslTree; // tree node
5494 GenTreePtr tslStmt; // statement containing the tree
5495 BasicBlock* tslBlock; // block containing the statement
5498 typedef struct treeStmtLst* treeStmtLstPtr;
5500 // The following logic keeps track of expressions via a simple hash table.
5504 CSEdsc* csdNextInBucket; // used by the hash table
5506 unsigned csdHashValue; // the orginal hashkey
5508 unsigned csdIndex; // 1..optCSECandidateCount
5509 char csdLiveAcrossCall; // 0 or 1
5511 unsigned short csdDefCount; // definition count
5512 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5514 unsigned csdDefWtCnt; // weighted def count
5515 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5517 GenTreePtr csdTree; // treenode containing the 1st occurance
5518 GenTreePtr csdStmt; // stmt containing the 1st occurance
5519 BasicBlock* csdBlock; // block containing the 1st occurance
5521 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5522 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5524 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5525 // number, this will reflect it; otherwise, NoVN.
5528 static const size_t s_optCSEhashSize;
5529 CSEdsc** optCSEhash;
5532 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5534 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5535 // re-numbered with the bound to improve range check elimination
5537 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5538 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5542 CSEdsc* optCSEfindDsc(unsigned index);
5543 void optUnmarkCSE(GenTreePtr tree);
5545 // user defined callback data for the tree walk function optCSE_MaskHelper()
5546 struct optCSE_MaskData
5548 EXPSET_TP CSE_defMask;
5549 EXPSET_TP CSE_useMask;
5552 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5553 static fgWalkPreFn optCSE_MaskHelper;
5555 // This function walks all the node for an given tree
5556 // and return the mask of CSE definitions and uses for the tree
5558 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5560 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5561 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5562 bool optCSE_canSwap(GenTree* tree);
5564 static fgWalkPostFn optPropagateNonCSE;
5565 static fgWalkPreFn optHasNonCSEChild;
5567 static fgWalkPreFn optUnmarkCSEs;
5569 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5570 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5572 void optCleanupCSEs();
5575 void optEnsureClearCSEInfo();
5578 #endif // FEATURE_ANYCSE
5580 #if FEATURE_VALNUM_CSE
5581 /**************************************************************************
5582 * Value Number based CSEs
5583 *************************************************************************/
5586 void optOptimizeValnumCSEs();
5589 void optValnumCSE_Init();
5590 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5591 unsigned optValnumCSE_Locate();
5592 void optValnumCSE_InitDataFlow();
5593 void optValnumCSE_DataFlow();
5594 void optValnumCSE_Availablity();
5595 void optValnumCSE_Heuristic();
5596 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5598 #endif // FEATURE_VALNUM_CSE
5601 bool optDoCSE; // True when we have found a duplicate CSE tree
5602 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5603 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5604 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5605 unsigned optCSEstart; // The first local variable number that is a CSE
5606 unsigned optCSEcount; // The total count of CSE's introduced.
5607 unsigned optCSEweight; // The weight of the current block when we are
5608 // scanning for CSE expressions
5610 bool optIsCSEcandidate(GenTreePtr tree);
5612 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5614 bool lclNumIsTrueCSE(unsigned lclNum) const
5616 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5619 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5621 bool lclNumIsCSE(unsigned lclNum) const
5623 return lvaTable[lclNum].lvIsCSE;
5627 bool optConfigDisableCSE();
5628 bool optConfigDisableCSE2();
5630 void optOptimizeCSEs();
5632 #endif // FEATURE_ANYCSE
5640 unsigned ivaVar; // Variable we are interested in, or -1
5641 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5642 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5643 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5644 callInterf ivaMaskCall; // What kind of calls are there?
5647 static callInterf optCallInterf(GenTreeCall* call);
5650 // VN based copy propagation.
5651 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5652 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5653 LclNumToGenTreePtrStack;
5655 // Kill set to track variables with intervening definitions.
5656 VARSET_TP optCopyPropKillSet;
5658 // Copy propagation functions.
5659 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5660 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5661 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5662 bool optIsSsaLocal(GenTreePtr tree);
5663 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5664 void optVnCopyProp();
5666 /**************************************************************************
5667 * Early value propagation
5668 *************************************************************************/
5674 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5678 static unsigned GetHashCode(SSAName ssaNm)
5680 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5683 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5685 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5689 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5690 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5691 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5692 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5693 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5694 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5696 bool doesMethodHaveFatPointer()
5698 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5701 void setMethodHasFatPointer()
5703 optMethodFlags |= OMF_HAS_FATPOINTER;
5706 void clearMethodHasFatPointer()
5708 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5711 void addFatPointerCandidate(GenTreeCall* call)
5713 setMethodHasFatPointer();
5714 call->SetFatPointerCandidate();
5717 unsigned optMethodFlags;
5719 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5720 // No throughput diff was found with backward walk bound between 3-8.
5721 static const int optEarlyPropRecurBound = 5;
5723 enum class optPropKind
5731 bool gtIsVtableRef(GenTreePtr tree);
5732 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5733 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5734 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5735 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5736 bool optEarlyPropRewriteTree(GenTreePtr tree);
5737 bool optDoEarlyPropForBlock(BasicBlock* block);
5738 bool optDoEarlyPropForFunc();
5739 void optEarlyProp();
5740 void optFoldNullCheck(GenTreePtr tree);
5741 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5744 /**************************************************************************
5745 * Value/Assertion propagation
5746 *************************************************************************/
5748 // Data structures for assertion prop
5749 BitVecTraits* apTraits;
5752 enum optAssertionKind
5769 O1K_CONSTANT_LOOP_BND,
5790 optAssertionKind assertionKind;
5793 unsigned lclNum; // assigned to or property of this local var number
5801 struct AssertionDscOp1
5803 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5810 struct AssertionDscOp2
5812 optOp2Kind kind; // a const or copy assignment
5816 ssize_t iconVal; // integer
5817 unsigned iconFlags; // gtFlags
5819 struct Range // integer subrange
5833 bool IsCheckedBoundArithBound()
5835 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5837 bool IsCheckedBoundBound()
5839 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5841 bool IsConstantBound()
5843 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5844 op1.kind == O1K_CONSTANT_LOOP_BND);
5846 bool IsBoundsCheckNoThrow()
5848 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5851 bool IsCopyAssertion()
5853 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5856 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5858 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5859 a1->op2.kind == a2->op2.kind;
5862 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5864 if (kind == OAK_EQUAL)
5866 return kind2 == OAK_NOT_EQUAL;
5868 else if (kind == OAK_NOT_EQUAL)
5870 return kind2 == OAK_EQUAL;
5875 static ssize_t GetLowerBoundForIntegralType(var_types type)
5895 static ssize_t GetUpperBoundForIntegralType(var_types type)
5919 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5921 if (op1.kind != that->op1.kind)
5925 else if (op1.kind == O1K_ARR_BND)
5928 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5932 return ((vnBased && (op1.vn == that->op1.vn)) ||
5933 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5937 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5939 if (op2.kind != that->op2.kind)
5945 case O2K_IND_CNS_INT:
5947 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5949 case O2K_CONST_LONG:
5950 return (op2.lconVal == that->op2.lconVal);
5952 case O2K_CONST_DOUBLE:
5953 // exact match because of positive and negative zero.
5954 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5956 case O2K_LCLVAR_COPY:
5958 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5959 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5962 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5965 // we will return false
5969 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5975 bool Complementary(AssertionDsc* that, bool vnBased)
5977 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5978 HasSameOp2(that, vnBased);
5981 bool Equals(AssertionDsc* that, bool vnBased)
5983 if (assertionKind != that->assertionKind)
5987 else if (assertionKind == OAK_NO_THROW)
5989 assert(op2.kind == O2K_INVALID);
5990 return HasSameOp1(that, vnBased);
5994 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
6000 static fgWalkPreFn optAddCopiesCallback;
6001 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
6002 unsigned optAddCopyLclNum;
6003 GenTreePtr optAddCopyAsgnNode;
6005 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6006 bool optAssertionPropagated; // set to true if we modified the trees
6007 bool optAssertionPropagatedCurrentStmt;
6009 GenTreePtr optAssertionPropCurrentTree;
6011 AssertionIndex* optComplementaryAssertionMap;
6012 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6013 // using the value of a local var) for each local var
6014 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6015 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6016 AssertionIndex optMaxAssertionCount;
6019 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6020 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6021 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6022 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6023 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6024 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6026 AssertionIndex GetAssertionCount()
6028 return optAssertionCount;
6030 ASSERT_TP* bbJtrueAssertionOut;
6031 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6032 ValueNumToAssertsMap;
6033 ValueNumToAssertsMap* optValueNumToAsserts;
6035 // Assertion prop helpers.
6036 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6037 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6038 void optAssertionInit(bool isLocalProp);
6039 void optAssertionTraitsInit(AssertionIndex assertionCount);
6040 #if LOCAL_ASSERTION_PROP
6041 void optAssertionReset(AssertionIndex limit);
6042 void optAssertionRemove(AssertionIndex index);
6045 // Assertion prop data flow functions.
6046 void optAssertionPropMain();
6047 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6048 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6049 ASSERT_TP* optInitAssertionDataflowFlags();
6050 ASSERT_TP* optComputeAssertionGen();
6052 // Assertion Gen functions.
6053 void optAssertionGen(GenTreePtr tree);
6054 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6055 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6056 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6057 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6058 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6059 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6061 // Assertion creation functions.
6062 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6063 AssertionIndex optCreateAssertion(GenTreePtr op1,
6065 optAssertionKind assertionKind,
6066 AssertionDsc* assertion);
6067 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6069 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6070 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6071 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6073 void optPrintVnAssertionMapping();
6075 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6077 // Used for respective assertion propagations.
6078 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6079 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6080 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6081 bool optAssertionIsNonNull(GenTreePtr op,
6082 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6084 // Used for Relop propagation.
6085 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6086 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6087 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6089 // Assertion prop for lcl var functions.
6090 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6091 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6093 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6094 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6095 const GenTreePtr tree,
6096 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6097 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6099 // Assertion propagation functions.
6100 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6101 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6102 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6103 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6104 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6105 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6106 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6107 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6108 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6109 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6110 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6111 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6113 // Implied assertion functions.
6114 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6115 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6116 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6117 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6120 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6121 void optDebugCheckAssertion(AssertionDsc* assertion);
6122 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6124 void optAddCopies();
6125 #endif // ASSERTION_PROP
6127 /**************************************************************************
6129 *************************************************************************/
6132 struct LoopCloneVisitorInfo
6134 LoopCloneContext* context;
6137 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6138 : context(context), loopNum(loopNum), stmt(nullptr)
6143 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6144 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6145 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6146 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6147 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6148 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6149 void optObtainLoopCloningOpts(LoopCloneContext* context);
6150 bool optIsLoopClonable(unsigned loopInd);
6152 bool optCanCloneLoops();
6155 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6157 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6158 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6159 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6160 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6164 void optInsertLoopCloningStress(BasicBlock* head);
6166 #if COUNT_RANGECHECKS
6167 static unsigned optRangeChkRmv;
6168 static unsigned optRangeChkAll;
6177 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6182 RngChkDsc* rcdNextInBucket; // used by the hash table
6184 unsigned short rcdHashValue; // to make matching faster
6185 unsigned short rcdIndex; // 0..optRngChkCount-1
6187 GenTreePtr rcdTree; // the array index tree
6190 unsigned optRngChkCount;
6191 static const size_t optRngChkHashSize;
6193 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6194 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6196 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6199 bool optLoopsMarked;
6202 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6203 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6207 XX Does the register allocation and puts the remaining lclVars on the stack XX
6209 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6210 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6214 #ifndef LEGACY_BACKEND
6219 #else // LEGACY_BACKEND
6224 #endif // LEGACY_BACKEND
6226 #ifdef LEGACY_BACKEND
6228 void raAssignVars(); // register allocation
6229 #endif // LEGACY_BACKEND
6231 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6233 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6235 void raMarkStkVars();
6238 // Some things are used by both LSRA and regpredict allocators.
6240 FrameType rpFrameType;
6241 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6243 #ifdef LEGACY_BACKEND
6244 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6246 #endif // LEGACY_BACKEND
6248 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6250 #if FEATURE_FP_REGALLOC
6251 enum enumConfigRegisterFP
6253 CONFIG_REGISTER_FP_NONE = 0x0,
6254 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6255 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6256 CONFIG_REGISTER_FP_FULL = 0x3,
6258 enumConfigRegisterFP raConfigRegisterFP();
6259 #endif // FEATURE_FP_REGALLOC
6262 regMaskTP raConfigRestrictMaskFP();
6265 #ifndef LEGACY_BACKEND
6266 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6267 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6268 #else // LEGACY_BACKEND
6269 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6270 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6271 bool raNewBlocks; // True is we added killing blocks for FPU registers
6272 unsigned rpPasses; // Number of passes made by the register predicter
6273 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6274 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6275 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6276 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6277 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6278 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6279 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6280 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6281 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6282 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6283 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6284 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6286 bool rpRegAllocDone; // Set to true after we have completed register allocation
6288 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6290 void raSetupArgMasks(RegState* r);
6292 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6294 void raDumpVarIntf(); // Dump the variable to variable interference graph
6295 void raDumpRegIntf(); // Dump the variable to register interference graph
6297 void raAdjustVarIntf();
6299 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6301 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6303 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6304 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6306 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6308 static fgWalkPreFn rpMarkRegIntf;
6310 regMaskTP rpPredictAddressMode(
6311 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6313 void rpPredictRefAssign(unsigned lclNum);
6315 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6317 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6319 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6321 void rpPredictRegUse(); // Entry point
6323 unsigned raPredictTreeRegUse(GenTreePtr tree);
6324 unsigned raPredictListRegUse(GenTreePtr list);
6326 void raSetRegVarOrder(var_types regType,
6327 regNumber* customVarOrder,
6328 unsigned* customVarOrderSize,
6330 regMaskTP avoidReg);
6332 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6333 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6334 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6335 void raAddToStkPredict(unsigned val)
6337 unsigned newStkPredict = rpStkPredict + val;
6338 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6339 rpStkPredict = UINT_MAX - 1;
6341 rpStkPredict = newStkPredict;
6345 #if !FEATURE_FP_REGALLOC
6346 void raDispFPlifeInfo();
6350 regMaskTP genReturnRegForTree(GenTreePtr tree);
6351 #endif // LEGACY_BACKEND
6353 /* raIsVarargsStackArg is called by raMaskStkVars and by
6354 lvaSortByRefCount. It identifies the special case
6355 where a varargs function has a parameter passed on the
6356 stack, other than the special varargs handle. Such parameters
6357 require special treatment, because they cannot be tracked
6358 by the GC (their offsets in the stack are not known
6362 bool raIsVarargsStackArg(unsigned lclNum)
6366 LclVarDsc* varDsc = &lvaTable[lclNum];
6368 assert(varDsc->lvIsParam);
6370 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6372 #else // _TARGET_X86_
6376 #endif // _TARGET_X86_
6379 #ifdef LEGACY_BACKEND
6380 // Records the current prediction, if it's better than any previous recorded prediction.
6381 void rpRecordPrediction();
6382 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6383 void rpUseRecordedPredictionIfBetter();
6385 // Data members used in the methods above.
6386 unsigned rpBestRecordedStkPredict;
6387 struct VarRegPrediction
6389 bool m_isEnregistered;
6390 regNumberSmall m_regNum;
6391 regNumberSmall m_otherReg;
6393 VarRegPrediction* rpBestRecordedPrediction;
6394 #endif // LEGACY_BACKEND
6397 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6398 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6402 XX Get to the class and method info from the Execution Engine given XX
6403 XX tokens for the class and method XX
6405 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6406 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6410 /* These are the different addressing modes used to access a local var.
6411 * The JIT has to report the location of the locals back to the EE
6412 * for debugging purposes.
6418 VLT_REG_BYREF, // this type is currently only used for value types on X64
6421 VLT_STK_BYREF, // this type is currently only used for value types on X64
6435 siVarLocType vlType;
6438 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6440 // VLT_REG_BYREF -- the specified register contains the address of the variable
6448 // VLT_STK -- Any 32 bit value which is on the stack
6449 // eg. [ESP+0x20], or [EBP-0x28]
6450 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6451 // eg. mov EAX, [ESP+0x20]; [EAX]
6455 regNumber vlsBaseReg;
6456 NATIVE_OFFSET vlsOffset;
6459 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6468 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6469 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6477 regNumber vlrssBaseReg;
6478 NATIVE_OFFSET vlrssOffset;
6482 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6483 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6489 regNumber vlsrsBaseReg;
6490 NATIVE_OFFSET vlsrsOffset;
6496 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6497 // eg 2 DWords at [ESP+0x10]
6501 regNumber vls2BaseReg;
6502 NATIVE_OFFSET vls2Offset;
6505 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6506 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6513 // VLT_FIXED_VA -- fixed argument of a varargs function.
6514 // The argument location depends on the size of the variable
6515 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6516 // location of the first arg. This argument can then be accessed
6517 // relative to the position of the first arg
6521 unsigned vlfvOffset;
6528 void* rpValue; // pointer to the in-process
6529 // location of the value.
6535 bool vlIsInReg(regNumber reg);
6536 bool vlIsOnStk(regNumber reg, signed offset);
6539 /*************************************************************************/
6544 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6545 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6546 CORINFO_CALLINFO_FLAGS flags,
6547 CORINFO_CALL_INFO* pResult);
6548 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6550 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6551 CORINFO_ACCESS_FLAGS flags,
6552 CORINFO_FIELD_INFO* pResult);
6556 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6558 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6560 bool IsSuperPMIException(unsigned code)
6562 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6564 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6565 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6566 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6567 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6568 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6569 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6570 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6571 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6575 case EXCEPTIONCODE_DebugBreakorAV:
6576 case EXCEPTIONCODE_MC:
6577 case EXCEPTIONCODE_LWM:
6578 case EXCEPTIONCODE_SASM:
6579 case EXCEPTIONCODE_SSYM:
6580 case EXCEPTIONCODE_CALLUTILS:
6581 case EXCEPTIONCODE_TYPEUTILS:
6582 case EXCEPTIONCODE_ASSERT:
6589 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6590 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6592 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6593 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6596 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6597 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6598 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6600 // VOM info, method sigs
6602 void eeGetSig(unsigned sigTok,
6603 CORINFO_MODULE_HANDLE scope,
6604 CORINFO_CONTEXT_HANDLE context,
6605 CORINFO_SIG_INFO* retSig);
6607 void eeGetCallSiteSig(unsigned sigTok,
6608 CORINFO_MODULE_HANDLE scope,
6609 CORINFO_CONTEXT_HANDLE context,
6610 CORINFO_SIG_INFO* retSig);
6612 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6614 // Method entry-points, instrs
6616 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6618 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6620 CORINFO_EE_INFO eeInfo;
6621 bool eeInfoInitialized;
6623 CORINFO_EE_INFO* eeGetEEInfo();
6625 // Gets the offset of a SDArray's first element
6626 unsigned eeGetArrayDataOffset(var_types type);
6627 // Gets the offset of a MDArray's first element
6628 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6630 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6632 // Returns the page size for the target machine as reported by the EE.
6633 inline size_t eeGetPageSize()
6635 return eeGetEEInfo()->osPageSize;
6638 // Returns the frame size at which we will generate a loop to probe the stack.
6639 inline size_t getVeryLargeFrameSize()
6642 // The looping probe code is 40 bytes, whereas the straight-line probing for
6643 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6644 // or greater, to generate smaller code.
6645 return 2 * eeGetPageSize();
6647 return 3 * eeGetPageSize();
6651 //------------------------------------------------------------------------
6652 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6654 // It represents Abi and target specific registers for the parameter.
6656 class VirtualStubParamInfo
6659 VirtualStubParamInfo(bool isCoreRTABI)
6661 #if defined(_TARGET_X86_)
6664 #elif defined(_TARGET_AMD64_)
6675 #elif defined(_TARGET_ARM_)
6678 #elif defined(_TARGET_ARM64_)
6682 #error Unsupported or unset target architecture
6685 #ifdef LEGACY_BACKEND
6686 #if defined(_TARGET_X86_)
6687 predict = PREDICT_REG_EAX;
6688 #elif defined(_TARGET_ARM_)
6689 predict = PREDICT_REG_R4;
6691 #error Unsupported or unset target architecture
6693 #endif // LEGACY_BACKEND
6696 regNumber GetReg() const
6701 _regMask_enum GetRegMask() const
6706 #ifdef LEGACY_BACKEND
6707 rpPredictReg GetPredict() const
6715 _regMask_enum regMask;
6717 #ifdef LEGACY_BACKEND
6718 rpPredictReg predict;
6722 VirtualStubParamInfo* virtualStubParamInfo;
6724 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6726 return eeGetEEInfo()->targetAbi == abi;
6729 inline bool generateCFIUnwindCodes()
6731 #ifdef UNIX_AMD64_ABI
6732 return IsTargetAbi(CORINFO_CORERT_ABI);
6740 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6742 // Debugging support - Line number info
6744 void eeGetStmtOffsets();
6746 unsigned eeBoundariesCount;
6748 struct boundariesDsc
6750 UNATIVE_OFFSET nativeIP;
6752 unsigned sourceReason;
6753 } * eeBoundaries; // Boundaries to report to EE
6754 void eeSetLIcount(unsigned count);
6755 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6759 static void eeDispILOffs(IL_OFFSET offs);
6760 static void eeDispLineInfo(const boundariesDsc* line);
6761 void eeDispLineInfos();
6764 // Debugging support - Local var info
6768 unsigned eeVarsCount;
6770 struct VarResultInfo
6772 UNATIVE_OFFSET startOffset;
6773 UNATIVE_OFFSET endOffset;
6777 void eeSetLVcount(unsigned count);
6778 void eeSetLVinfo(unsigned which,
6779 UNATIVE_OFFSET startOffs,
6780 UNATIVE_OFFSET length,
6785 const siVarLoc& loc);
6789 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6790 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6793 // ICorJitInfo wrappers
6795 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6797 void eeAllocUnwindInfo(BYTE* pHotCode,
6803 CorJitFuncKind funcKind);
6805 void eeSetEHcount(unsigned cEH);
6807 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6809 WORD eeGetRelocTypeHint(void* target);
6811 // ICorStaticInfo wrapper functions
6813 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6815 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6817 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6820 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6821 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6822 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6823 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6825 template <typename ParamType>
6826 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6828 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6831 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6833 // Utility functions
6835 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6838 const wchar_t* eeGetCPString(size_t stringHandle);
6841 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6843 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6844 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6846 static fgWalkPreFn CountSharedStaticHelper;
6847 static bool IsSharedStaticHelper(GenTreePtr tree);
6848 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6849 static bool IsGcSafePoint(GenTreePtr tree);
6851 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6852 // returns true/false if 'field' is a Jit Data offset
6853 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6854 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6855 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6857 /*****************************************************************************/
6862 enum TEMP_USAGE_TYPE
6868 static var_types tmpNormalizeType(var_types type);
6869 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6870 void tmpRlsTemp(TempDsc* temp);
6871 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6874 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6875 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6879 bool tmpAllFree() const;
6882 #ifndef LEGACY_BACKEND
6883 void tmpPreAllocateTemps(var_types type, unsigned count);
6884 #endif // !LEGACY_BACKEND
6887 #ifdef LEGACY_BACKEND
6888 unsigned tmpIntSpillMax; // number of int-sized spill temps
6889 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6890 #endif // LEGACY_BACKEND
6892 unsigned tmpCount; // Number of temps
6893 unsigned tmpSize; // Size of all the temps
6896 // Used by RegSet::rsSpillChk()
6897 unsigned tmpGetCount; // Temps which haven't been released yet
6900 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6902 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6903 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6906 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6907 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6911 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6912 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6916 CodeGenInterface* codeGen;
6918 // The following holds information about instr offsets in terms of generated code.
6922 IPmappingDsc* ipmdNext; // next line# record
6923 IL_OFFSETX ipmdILoffsx; // the instr offset
6924 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6925 bool ipmdIsLabel; // Can this code be a branch label?
6928 // Record the instr offset mapping to the generated code
6930 IPmappingDsc* genIPmappingList;
6931 IPmappingDsc* genIPmappingLast;
6933 // Managed RetVal - A side hash table meant to record the mapping from a
6934 // GT_CALL node to its IL offset. This info is used to emit sequence points
6935 // that can be used by debugger to determine the native offset at which the
6936 // managed RetVal will be available.
6938 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6939 // favor of a side table for two reasons: 1) We need IL offset for only those
6940 // GT_CALL nodes (created during importation) that correspond to an IL call and
6941 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6942 // structure and IL offset is needed only when generating debuggable code. Therefore
6943 // it is desirable to avoid memory size penalty in retail scenarios.
6944 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6945 CallSiteILOffsetTable;
6946 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6948 unsigned genReturnLocal; // Local number for the return value when applicable.
6949 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6951 // The following properties are part of CodeGenContext. Getters are provided here for
6952 // convenience and backward compatibility, but the properties can only be set by invoking
6953 // the setter on CodeGenContext directly.
6955 __declspec(property(get = getEmitter)) emitter* genEmitter;
6956 emitter* getEmitter()
6958 return codeGen->getEmitter();
6961 const bool isFramePointerUsed()
6963 return codeGen->isFramePointerUsed();
6966 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6967 bool getInterruptible()
6969 return codeGen->genInterruptible;
6971 void setInterruptible(bool value)
6973 codeGen->setInterruptible(value);
6977 const bool genDoubleAlign()
6979 return codeGen->doDoubleAlign();
6981 DWORD getCanDoubleAlign();
6982 bool shouldDoubleAlign(unsigned refCntStk,
6984 unsigned refCntWtdReg,
6985 unsigned refCntStkParam,
6986 unsigned refCntWtdStkDbl);
6987 #endif // DOUBLE_ALIGN
6989 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6990 bool getFullPtrRegMap()
6992 return codeGen->genFullPtrRegMap;
6994 void setFullPtrRegMap(bool value)
6996 codeGen->setFullPtrRegMap(value);
6999 // Things that MAY belong either in CodeGen or CodeGenContext
7001 #if FEATURE_EH_FUNCLETS
7002 FuncInfoDsc* compFuncInfos;
7003 unsigned short compCurrFuncIdx;
7004 unsigned short compFuncInfoCount;
7006 unsigned short compFuncCount()
7008 assert(fgFuncletsCreated);
7009 return compFuncInfoCount;
7012 #else // !FEATURE_EH_FUNCLETS
7014 // This is a no-op when there are no funclets!
7015 void genUpdateCurrentFunclet(BasicBlock* block)
7020 FuncInfoDsc compFuncInfoRoot;
7022 static const unsigned compCurrFuncIdx = 0;
7024 unsigned short compFuncCount()
7029 #endif // !FEATURE_EH_FUNCLETS
7031 FuncInfoDsc* funCurrentFunc();
7032 void funSetCurrentFunc(unsigned funcIdx);
7033 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7034 unsigned int funGetFuncIdx(BasicBlock* block);
7038 VARSET_TP compCurLife; // current live variables
7039 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7041 template <bool ForCodeGen>
7042 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7044 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7046 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7049 template <bool ForCodeGen>
7050 void compUpdateLife(GenTreePtr tree);
7052 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7053 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7054 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7055 template <bool ForCodeGen>
7056 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7058 template <bool ForCodeGen>
7059 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7061 // Gets a register mask that represent the kill set for a helper call since
7062 // not all JIT Helper calls follow the standard ABI on the target architecture.
7063 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7065 // Gets a register mask that represent the kill set for a NoGC helper call.
7066 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7069 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7070 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7071 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7072 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7073 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7074 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7075 #endif // _TARGET_ARM_
7077 // 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
7079 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7081 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7082 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7083 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7084 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7085 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7086 // for the tracked var indices of the field vars, as in a live var set).
7087 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7089 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7091 if (m_promotedStructDeathVars == nullptr)
7093 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7095 return m_promotedStructDeathVars;
7099 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7100 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7104 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7105 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7108 #if !defined(__GNUC__)
7109 #pragma region Unwind information
7114 // Infrastructure functions: start/stop/reserve/emit.
7117 void unwindBegProlog();
7118 void unwindEndProlog();
7119 void unwindBegEpilog();
7120 void unwindEndEpilog();
7121 void unwindReserve();
7122 void unwindEmit(void* pHotCode, void* pColdCode);
7125 // Specific unwind information functions: called by code generation to indicate a particular
7126 // prolog or epilog unwindable instruction has been generated.
7129 void unwindPush(regNumber reg);
7130 void unwindAllocStack(unsigned size);
7131 void unwindSetFrameReg(regNumber reg, unsigned offset);
7132 void unwindSaveReg(regNumber reg, unsigned offset);
7134 #if defined(_TARGET_ARM_)
7135 void unwindPushMaskInt(regMaskTP mask);
7136 void unwindPushMaskFloat(regMaskTP mask);
7137 void unwindPopMaskInt(regMaskTP mask);
7138 void unwindPopMaskFloat(regMaskTP mask);
7139 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7140 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7141 // called via unwindPadding().
7142 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7143 // instruction and the current location.
7144 #endif // _TARGET_ARM_
7146 #if defined(_TARGET_ARM64_)
7148 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7149 // instruction and the current location.
7150 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7151 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7152 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7153 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7154 void unwindSaveNext(); // unwind code: save_next
7155 void unwindReturn(regNumber reg); // ret lr
7156 #endif // defined(_TARGET_ARM64_)
7159 // Private "helper" functions for the unwind implementation.
7163 #if FEATURE_EH_FUNCLETS
7164 void unwindGetFuncLocations(FuncInfoDsc* func,
7165 bool getHotSectionData,
7166 /* OUT */ emitLocation** ppStartLoc,
7167 /* OUT */ emitLocation** ppEndLoc);
7168 #endif // FEATURE_EH_FUNCLETS
7170 void unwindReserveFunc(FuncInfoDsc* func);
7171 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7173 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7175 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7176 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7178 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7180 #if defined(_TARGET_AMD64_)
7182 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7184 void unwindBegPrologWindows();
7185 void unwindPushWindows(regNumber reg);
7186 void unwindAllocStackWindows(unsigned size);
7187 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7188 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7190 #ifdef UNIX_AMD64_ABI
7191 void unwindBegPrologCFI();
7192 void unwindPushCFI(regNumber reg);
7193 void unwindAllocStackCFI(unsigned size);
7194 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7195 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7196 int mapRegNumToDwarfReg(regNumber reg);
7197 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7198 #endif // UNIX_AMD64_ABI
7199 #elif defined(_TARGET_ARM_)
7201 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7202 void unwindPushPopMaskFloat(regMaskTP mask);
7203 void unwindSplit(FuncInfoDsc* func);
7205 #endif // _TARGET_ARM_
7207 #if !defined(__GNUC__)
7208 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7212 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7213 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7217 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7218 XX that contains the distinguished, well-known SIMD type definitions). XX
7220 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7221 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7224 // Get highest available instruction set for floating point codegen
7225 InstructionSet getFloatingPointInstructionSet()
7227 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7230 return InstructionSet_AVX;
7235 return InstructionSet_SSE3_4;
7239 assert(canUseSSE2());
7240 return InstructionSet_SSE2;
7242 assert(!"getFPInstructionSet() is not implemented for target arch");
7244 return InstructionSet_NONE;
7248 // Get highest available instruction set for SIMD codegen
7249 InstructionSet getSIMDInstructionSet()
7251 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7252 return getFloatingPointInstructionSet();
7254 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7256 return InstructionSet_NONE;
7262 // Should we support SIMD intrinsics?
7265 // Have we identified any SIMD types?
7266 // This is currently used by struct promotion to avoid getting type information for a struct
7267 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7269 bool _usesSIMDTypes;
7270 bool usesSIMDTypes()
7272 return _usesSIMDTypes;
7274 void setUsesSIMDTypes(bool value)
7276 _usesSIMDTypes = value;
7279 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7280 // that require indexed access to the individual fields of the vector, which is not well supported
7281 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7282 unsigned lvaSIMDInitTempVarNum;
7285 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7286 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7287 CORINFO_CLASS_HANDLE SIMDIntHandle;
7288 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7289 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7290 CORINFO_CLASS_HANDLE SIMDShortHandle;
7291 CORINFO_CLASS_HANDLE SIMDByteHandle;
7292 CORINFO_CLASS_HANDLE SIMDLongHandle;
7293 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7294 CORINFO_CLASS_HANDLE SIMDULongHandle;
7295 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7296 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7297 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7298 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7300 // Get the handle for a SIMD type.
7301 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7303 if (simdBaseType == TYP_FLOAT)
7308 return SIMDVector2Handle;
7310 return SIMDVector3Handle;
7312 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7314 return SIMDVector4Handle;
7323 assert(simdType == getSIMDVectorType());
7324 switch (simdBaseType)
7327 return SIMDFloatHandle;
7329 return SIMDDoubleHandle;
7331 return SIMDIntHandle;
7333 return SIMDUShortHandle;
7335 return SIMDUShortHandle;
7337 return SIMDUByteHandle;
7339 return SIMDShortHandle;
7341 return SIMDByteHandle;
7343 return SIMDLongHandle;
7345 return SIMDUIntHandle;
7347 return SIMDULongHandle;
7349 assert(!"Didn't find a class handle for simdType");
7351 return NO_CLASS_HANDLE;
7355 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7356 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7357 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7359 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7360 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7361 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7362 bool isSIMDTypeLocal(GenTree* tree)
7364 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7367 // Returns true if the type of the tree is a byref of TYP_SIMD
7368 bool isAddrOfSIMDType(GenTree* tree)
7370 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7372 switch (tree->OperGet())
7375 return varTypeIsSIMD(tree->gtGetOp1());
7377 case GT_LCL_VAR_ADDR:
7378 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7381 return isSIMDTypeLocal(tree);
7388 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7390 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7391 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7392 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7395 // Returns base type of a TYP_SIMD local.
7396 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7397 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7399 if (isSIMDTypeLocal(tree))
7401 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7407 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7409 return info.compCompHnd->isInSIMDModule(clsHnd);
7412 bool isSIMDClass(typeInfo* pTypeInfo)
7414 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7417 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7418 // if it is not a SIMD type or is an unsupported base type.
7419 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7421 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7423 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7426 // Get SIMD Intrinsic info given the method handle.
7427 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7428 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7429 CORINFO_METHOD_HANDLE methodHnd,
7430 CORINFO_SIG_INFO* sig,
7433 var_types* baseType,
7434 unsigned* sizeBytes);
7436 // Pops and returns GenTree node from importers type stack.
7437 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7438 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7440 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7441 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7443 // Creates a GT_SIMD tree for Select operation
7444 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7446 unsigned simdVectorSize,
7451 // Creates a GT_SIMD tree for Min/Max operation
7452 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7453 CORINFO_CLASS_HANDLE typeHnd,
7455 unsigned simdVectorSize,
7459 // Transforms operands and returns the SIMD intrinsic to be applied on
7460 // transformed operands to obtain given relop result.
7461 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7462 CORINFO_CLASS_HANDLE typeHnd,
7463 unsigned simdVectorSize,
7464 var_types* baseType,
7468 // Creates a GT_SIMD tree for Abs intrinsic.
7469 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7471 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7472 // Transforms operands and returns the SIMD intrinsic to be applied on
7473 // transformed operands to obtain == comparison result.
7474 SIMDIntrinsicID impSIMDLongRelOpEqual(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 impSIMDLongRelOpGreaterThan(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.
7488 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7489 unsigned simdVectorSize,
7493 // Transforms operands and returns the SIMD intrinsic to be applied on
7494 // transformed operands to obtain >= comparison result in case of int32
7495 // and small int base type vectors.
7496 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7497 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7498 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7500 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7501 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7502 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7503 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7504 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7506 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7507 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7508 GenTreePtr newobjThis,
7509 CORINFO_CLASS_HANDLE clsHnd,
7510 CORINFO_METHOD_HANDLE method,
7511 CORINFO_SIG_INFO* sig,
7514 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7516 // Whether SIMD vector occupies part of SIMD register.
7517 // SSE2: vector2f/3f are considered sub register SIMD types.
7518 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7519 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7521 unsigned sizeBytes = 0;
7522 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7523 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7526 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7528 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7531 // Get the type for the hardware SIMD vector.
7532 // This is the maximum SIMD type supported for this target.
7533 var_types getSIMDVectorType()
7535 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7542 assert(canUseSSE2());
7546 assert(!"getSIMDVectorType() unimplemented on target arch");
7551 // Get the size of the SIMD type in bytes
7552 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7554 unsigned sizeBytes = 0;
7555 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7559 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7560 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7562 // Get the the number of elements of basetype of SIMD vector given by its type handle
7563 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7565 // Get preferred alignment of SIMD type.
7566 int getSIMDTypeAlignment(var_types simdType);
7568 // Get the number of bytes in a SIMD Vector for the current compilation.
7569 unsigned getSIMDVectorRegisterByteLength()
7571 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7574 return YMM_REGSIZE_BYTES;
7578 assert(canUseSSE2());
7579 return XMM_REGSIZE_BYTES;
7582 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7587 // The minimum and maximum possible number of bytes in a SIMD vector.
7588 unsigned int maxSIMDStructBytes()
7590 return getSIMDVectorRegisterByteLength();
7592 unsigned int minSIMDStructBytes()
7594 return emitTypeSize(TYP_SIMD8);
7597 #ifdef FEATURE_AVX_SUPPORT
7598 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7599 static const unsigned maxPossibleSIMDStructBytes = 32;
7600 #else // !FEATURE_AVX_SUPPORT
7601 static const unsigned maxPossibleSIMDStructBytes = 16;
7602 #endif // !FEATURE_AVX_SUPPORT
7604 // Returns the codegen type for a given SIMD size.
7605 var_types getSIMDTypeForSize(unsigned size)
7607 var_types simdType = TYP_UNDEF;
7610 simdType = TYP_SIMD8;
7612 else if (size == 12)
7614 simdType = TYP_SIMD12;
7616 else if (size == 16)
7618 simdType = TYP_SIMD16;
7620 #ifdef FEATURE_AVX_SUPPORT
7621 else if (size == 32)
7623 simdType = TYP_SIMD32;
7625 #endif // FEATURE_AVX_SUPPORT
7628 noway_assert(!"Unexpected size for SIMD type");
7633 unsigned getSIMDInitTempVarNum()
7635 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7637 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7638 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7640 return lvaSIMDInitTempVarNum;
7643 #endif // FEATURE_SIMD
7646 //------------------------------------------------------------------------
7647 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7649 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7650 // candidate for enregistration.
7652 unsigned largestEnregisterableStructSize()
7655 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7656 if (vectorRegSize > TARGET_POINTER_SIZE)
7658 return vectorRegSize;
7661 #endif // FEATURE_SIMD
7663 return TARGET_POINTER_SIZE;
7668 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7669 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7670 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7672 // Is this var is of type simd struct?
7673 bool lclVarIsSIMDType(unsigned varNum)
7675 LclVarDsc* varDsc = lvaTable + varNum;
7676 return varDsc->lvIsSIMDType();
7679 // Is this Local node a SIMD local?
7680 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7682 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7685 // Returns true if the TYP_SIMD locals on stack are aligned at their
7686 // preferred byte boundary specified by getSIMDTypeAlignment().
7688 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7689 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7690 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7691 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7692 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7693 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7694 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7697 bool isSIMDTypeLocalAligned(unsigned varNum)
7699 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7700 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7703 int off = lvaFrameAddress(varNum, &ebpBased);
7704 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7705 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7706 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7709 #endif // FEATURE_SIMD
7714 // Whether SSE2 is available
7715 bool canUseSSE2() const
7717 #ifdef _TARGET_XARCH_
7718 return opts.compCanUseSSE2;
7724 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7725 bool CanUseSSE3_4() const
7727 #ifdef _TARGET_XARCH_
7728 return opts.compCanUseSSE3_4;
7734 bool canUseAVX() const
7736 #ifdef FEATURE_AVX_SUPPORT
7737 return opts.compCanUseAVX;
7744 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7745 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7749 XX Generic info about the compilation and the method being compiled. XX
7750 XX It is responsible for driving the other phases. XX
7751 XX It is also responsible for all the memory management. XX
7753 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7754 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7758 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7760 InlineResult* compInlineResult; // The result of importing the inlinee method.
7762 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7763 bool compJmpOpUsed; // Does the method do a JMP
7764 bool compLongUsed; // Does the method use TYP_LONG
7765 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7766 bool compTailCallUsed; // Does the method do a tailcall
7767 bool compLocallocUsed; // Does the method use localloc.
7768 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7769 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7770 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7772 // NOTE: These values are only reliable after
7773 // the importing is completely finished.
7775 #ifdef LEGACY_BACKEND
7776 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7777 // we can iterate over these efficiently.
7780 #if CPU_USES_BLOCK_MOVE
7781 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7785 // State information - which phases have completed?
7786 // These are kept together for easy discoverability
7788 bool bRangeAllowStress;
7789 bool compCodeGenDone;
7790 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7791 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7792 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7793 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7796 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7797 bool fgLocalVarLivenessChanged;
7799 bool compStackProbePrologDone;
7801 #ifndef LEGACY_BACKEND
7803 #endif // !LEGACY_BACKEND
7804 bool compRationalIRForm;
7806 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7808 bool compGeneratingProlog;
7809 bool compGeneratingEpilog;
7810 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7811 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7812 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7813 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7814 bool getNeedsGSSecurityCookie() const
7816 return compNeedsGSSecurityCookie;
7818 void setNeedsGSSecurityCookie()
7820 compNeedsGSSecurityCookie = true;
7823 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7824 // frame layout calculations, this is the level we are currently
7827 //---------------------------- JITing options -----------------------------
7840 JitFlags* jitFlags; // all flags passed from the EE
7841 unsigned compFlags; // method attributes
7843 codeOptimize compCodeOpt; // what type of code optimizations
7847 #ifdef _TARGET_XARCH_
7848 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7849 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7851 #ifdef FEATURE_AVX_SUPPORT
7852 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7853 #endif // FEATURE_AVX_SUPPORT
7854 #endif // _TARGET_XARCH_
7856 // optimize maximally and/or favor speed over size?
7858 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7859 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7860 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7861 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7862 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7864 // Maximun number of locals before turning off the inlining
7865 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7868 unsigned instrCount;
7869 unsigned lvRefCount;
7870 bool compMinOptsIsSet;
7872 bool compMinOptsIsUsed;
7874 inline bool MinOpts()
7876 assert(compMinOptsIsSet);
7877 compMinOptsIsUsed = true;
7880 inline bool IsMinOptsSet()
7882 return compMinOptsIsSet;
7885 inline bool MinOpts()
7889 inline bool IsMinOptsSet()
7891 return compMinOptsIsSet;
7894 inline void SetMinOpts(bool val)
7896 assert(!compMinOptsIsUsed);
7897 assert(!compMinOptsIsSet || (compMinOpts == val));
7899 compMinOptsIsSet = true;
7902 // true if the CLFLG_* for an optimization is set.
7903 inline bool OptEnabled(unsigned optFlag)
7905 return !!(compFlags & optFlag);
7908 #ifdef FEATURE_READYTORUN_COMPILER
7909 inline bool IsReadyToRun()
7911 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7914 inline bool IsReadyToRun()
7920 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7921 // PInvoke transitions inline (e.g. when targeting CoreRT).
7922 inline bool ShouldUsePInvokeHelpers()
7924 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7927 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7929 inline bool IsReversePInvoke()
7931 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7934 // true if we must generate code compatible with JIT32 quirks
7935 inline bool IsJit32Compat()
7937 #if defined(_TARGET_X86_)
7938 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7944 // true if we must generate code compatible with Jit64 quirks
7945 inline bool IsJit64Compat()
7947 #if defined(_TARGET_AMD64_)
7948 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7949 #elif !defined(FEATURE_CORECLR)
7956 bool compScopeInfo; // Generate the LocalVar info ?
7957 bool compDbgCode; // Generate debugger-friendly code?
7958 bool compDbgInfo; // Gather debugging info?
7961 #ifdef PROFILING_SUPPORTED
7962 bool compNoPInvokeInlineCB;
7964 static const bool compNoPInvokeInlineCB;
7968 bool compGcChecks; // Check arguments and return values to ensure they are sane
7969 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7970 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7974 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7975 // to be allocated on the stack.
7976 // It will be set to true in the following cases:
7977 // 1. When the method being compiled has a declarative security
7978 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7979 // This is also the case when we inject a prolog and epilog in the method.
7981 // 2. When the method being compiled has imperative security (i.e. the method
7982 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7984 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7986 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7987 // which gets reported as a GC root to stackwalker.
7988 // (See also ICodeManager::GetAddrOfSecurityObject.)
7993 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7994 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7998 #ifdef UNIX_AMD64_ABI
7999 // This flag is indicating if there is a need to align the frame.
8000 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8001 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8002 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8003 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8004 // there are calls and making sure the frame alignment logic is executed.
8005 bool compNeedToAlignFrame;
8006 #endif // UNIX_AMD64_ABI
8008 bool compProcedureSplitting; // Separate cold code from hot code
8010 bool genFPorder; // Preserve FP order (operations are non-commutative)
8011 bool genFPopt; // Can we do frame-pointer-omission optimization?
8012 bool altJit; // True if we are an altjit and are compiling this method
8015 bool optRepeat; // Repeat optimizer phases k times
8019 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8020 bool dspCode; // Display native code generated
8021 bool dspEHTable; // Display the EH table reported to the VM
8022 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8023 bool dspEmit; // Display emitter output
8024 bool dspLines; // Display source-code lines intermixed with native code output
8025 bool dmpHex; // Display raw bytes in hex of native code output
8026 bool varNames; // Display variables names in native code output
8027 bool disAsm; // Display native code as it is generated
8028 bool disAsmSpilled; // Display native code when any register spilling occurs
8029 bool disDiffable; // Makes the Disassembly code 'diff-able'
8030 bool disAsm2; // Display native code after it is generated using external disassembler
8031 bool dspOrder; // Display names of each of the methods that we ngen/jit
8032 bool dspUnwind; // Display the unwind info output
8033 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8034 bool compLongAddress; // Force using large pseudo instructions for long address
8035 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8036 bool dspGCtbls; // Display the GC tables
8040 bool doLateDisasm; // Run the late disassembler
8041 #endif // LATE_DISASM
8043 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8044 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8045 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8046 static const bool dspGCtbls = true;
8049 // We need stack probes to guarantee that we won't trigger a stack overflow
8050 // when calling unmanaged code until they get a chance to set up a frame, because
8051 // the EE will have no idea where it is.
8053 // We will only be doing this currently for hosted environments. Unfortunately
8054 // we need to take care of stubs, so potentially, we will have to do the probes
8055 // for any call. We have a plan for not needing for stubs though
8056 bool compNeedStackProbes;
8058 #ifdef PROFILING_SUPPORTED
8059 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8060 // This option helps make the JIT behave as if it is running under a profiler.
8061 bool compJitELTHookEnabled;
8062 #endif // PROFILING_SUPPORTED
8064 #if FEATURE_TAILCALL_OPT
8065 // Whether opportunistic or implicit tail call optimization is enabled.
8066 bool compTailCallOpt;
8067 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8068 bool compTailCallLoopOpt;
8072 static const bool compUseSoftFP = true;
8073 #else // !ARM_SOFTFP
8074 static const bool compUseSoftFP = false;
8077 GCPollType compGCPollType;
8081 static bool s_pAltJitExcludeAssembliesListInitialized;
8082 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8087 template <typename T>
8090 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8093 template <typename T>
8096 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8099 static int dspTreeID(GenTree* tree)
8101 return tree->gtTreeID;
8103 static void printTreeID(GenTree* tree)
8105 if (tree == nullptr)
8111 printf("[%06d]", dspTreeID(tree));
8118 #define STRESS_MODES \
8122 /* "Variations" stress areas which we try to mix up with each other. */ \
8123 /* These should not be exhaustively used as they might */ \
8124 /* hide/trivialize other areas */ \
8127 STRESS_MODE(DBL_ALN) \
8128 STRESS_MODE(LCL_FLDS) \
8129 STRESS_MODE(UNROLL_LOOPS) \
8130 STRESS_MODE(MAKE_CSE) \
8131 STRESS_MODE(LEGACY_INLINE) \
8132 STRESS_MODE(CLONE_EXPR) \
8133 STRESS_MODE(USE_FCOMI) \
8134 STRESS_MODE(USE_CMOV) \
8136 STRESS_MODE(BB_PROFILE) \
8137 STRESS_MODE(OPT_BOOLS_GC) \
8138 STRESS_MODE(REMORPH_TREES) \
8139 STRESS_MODE(64RSLT_MUL) \
8140 STRESS_MODE(DO_WHILE_LOOPS) \
8141 STRESS_MODE(MIN_OPTS) \
8142 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8143 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8144 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8145 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8146 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8147 STRESS_MODE(NULL_OBJECT_CHECK) \
8148 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8149 STRESS_MODE(RANDOM_INLINE) \
8150 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8151 STRESS_MODE(GENERIC_VARN) \
8153 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8155 STRESS_MODE(COUNT_VARN) \
8157 /* "Check" stress areas that can be exhaustively used if we */ \
8158 /* dont care about performance at all */ \
8160 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8161 STRESS_MODE(CHK_FLOW_UPDATE) \
8162 STRESS_MODE(EMITTER) \
8163 STRESS_MODE(CHK_REIMPORT) \
8164 STRESS_MODE(FLATFP) \
8165 STRESS_MODE(GENERIC_CHECK) \
8170 #define STRESS_MODE(mode) STRESS_##mode,
8177 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8178 BYTE compActiveStressModes[STRESS_COUNT];
8181 #define MAX_STRESS_WEIGHT 100
8183 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8187 bool compInlineStress()
8189 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8192 bool compRandomInlineStress()
8194 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8199 bool compTailCallStress()
8202 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8208 codeOptimize compCodeOpt()
8211 // Switching between size & speed has measurable throughput impact
8212 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8213 // DEBUG, but should generate identical code between CHK & RET builds,
8214 // so that's not acceptable.
8215 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8216 // Investigate the cause of the throughput regression.
8218 return opts.compCodeOpt;
8220 return BLENDED_CODE;
8224 //--------------------- Info about the procedure --------------------------
8228 COMP_HANDLE compCompHnd;
8229 CORINFO_MODULE_HANDLE compScopeHnd;
8230 CORINFO_CLASS_HANDLE compClassHnd;
8231 CORINFO_METHOD_HANDLE compMethodHnd;
8232 CORINFO_METHOD_INFO* compMethodInfo;
8234 BOOL hasCircularClassConstraints;
8235 BOOL hasCircularMethodConstraints;
8237 #if defined(DEBUG) || defined(LATE_DISASM)
8238 const char* compMethodName;
8239 const char* compClassName;
8240 const char* compFullName;
8241 #endif // defined(DEBUG) || defined(LATE_DISASM)
8243 #if defined(DEBUG) || defined(INLINE_DATA)
8244 // Method hash is logcally const, but computed
8246 mutable unsigned compMethodHashPrivate;
8247 unsigned compMethodHash() const;
8248 #endif // defined(DEBUG) || defined(INLINE_DATA)
8250 #ifdef PSEUDORANDOM_NOP_INSERTION
8251 // things for pseudorandom nop insertion
8252 unsigned compChecksum;
8256 // The following holds the FLG_xxxx flags for the method we're compiling.
8259 // The following holds the class attributes for the method we're compiling.
8260 unsigned compClassAttr;
8262 const BYTE* compCode;
8263 IL_OFFSET compILCodeSize; // The IL code size
8264 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8265 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8266 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8267 // (2) the code is hot/cold split, and we issued less code than we expected
8268 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8270 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8271 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8272 bool compIsContextful : 1; // contextful method
8273 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8274 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8275 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8276 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8277 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8279 var_types compRetType; // Return type of the method as declared in IL
8280 var_types compRetNativeType; // Normalized return type as per target arch ABI
8281 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8282 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8283 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8284 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8285 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8286 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8287 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8288 unsigned compMaxStack;
8289 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8290 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8292 unsigned compCallUnmanaged; // count of unmanaged calls
8293 unsigned compLvFrameListRoot; // lclNum for the Frame root
8294 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8295 // You should generally use compHndBBtabCount instead: it is the
8296 // current number of EH clauses (after additions like synchronized
8297 // methods and funclets, and removals like unreachable code deletion).
8299 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8300 // and the VM expects that, or the JIT is a "self-host" compiler
8301 // (e.g., x86 hosted targeting x86) and the VM expects that.
8303 /* The following holds IL scope information about local variables.
8306 unsigned compVarScopesCount;
8307 VarScopeDsc* compVarScopes;
8309 /* The following holds information about instr offsets for
8310 * which we need to report IP-mappings
8313 IL_OFFSET* compStmtOffsets; // sorted
8314 unsigned compStmtOffsetsCount;
8315 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8317 #define CPU_X86 0x0100 // The generic X86 CPU
8318 #define CPU_X86_PENTIUM_4 0x0110
8320 #define CPU_X64 0x0200 // The generic x64 CPU
8321 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8322 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8324 #define CPU_ARM 0x0300 // The generic ARM CPU
8326 unsigned genCPU; // What CPU are we running on
8329 // Returns true if the method being compiled returns a non-void and non-struct value.
8330 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8331 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8332 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8333 // Methods returning such structs are considered to return non-struct return value and
8334 // this method returns true in that case.
8335 bool compMethodReturnsNativeScalarType()
8337 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8340 // Returns true if the method being compiled returns RetBuf addr as its return value
8341 bool compMethodReturnsRetBufAddr()
8343 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8344 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8346 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8347 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8348 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8349 // methods with hidden RetBufArg.
8351 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8352 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8353 // returning the address of RetBuf.
8355 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8356 // to be returned in RAX.
8357 CLANG_FORMAT_COMMENT_ANCHOR;
8359 #ifdef _TARGET_AMD64_
8360 return (info.compRetBuffArg != BAD_VAR_NUM);
8361 #else // !_TARGET_AMD64_
8362 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8363 #endif // !_TARGET_AMD64_
8366 // Returns true if the method returns a value in more than one return register
8367 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8368 // TODO-ARM64: Does this apply for ARM64 too?
8369 bool compMethodReturnsMultiRegRetType()
8371 #if FEATURE_MULTIREG_RET
8372 #if defined(_TARGET_X86_)
8373 // On x86 only 64-bit longs are returned in multiple registers
8374 return varTypeIsLong(info.compRetNativeType);
8375 #else // targets: X64-UNIX, ARM64 or ARM32
8376 // On all other targets that support multireg return values:
8377 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8378 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8379 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8380 #endif // TARGET_XXX
8382 #else // not FEATURE_MULTIREG_RET
8384 // For this architecture there are no multireg returns
8387 #endif // FEATURE_MULTIREG_RET
8390 #if FEATURE_MULTIREG_ARGS
8391 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8392 // return the gcPtr layout for the pointers sized fields
8393 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8394 #endif // FEATURE_MULTIREG_ARGS
8396 // Returns true if the method being compiled returns a value
8397 bool compMethodHasRetVal()
8399 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8400 compMethodReturnsMultiRegRetType();
8405 void compDispLocalVars();
8409 //-------------------------- Global Compiler Data ------------------------------------
8412 static unsigned s_compMethodsCount; // to produce unique label names
8413 unsigned compGenTreeID;
8416 BasicBlock* compCurBB; // the current basic block in process
8417 GenTreePtr compCurStmt; // the current statement in process
8419 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8422 // The following is used to create the 'method JIT info' block.
8423 size_t compInfoBlkSize;
8424 BYTE* compInfoBlkAddr;
8426 EHblkDsc* compHndBBtab; // array of EH data
8427 unsigned compHndBBtabCount; // element count of used elements in EH data array
8428 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8430 #if defined(_TARGET_X86_)
8432 //-------------------------------------------------------------------------
8433 // Tracking of region covered by the monitor in synchronized methods
8434 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8435 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8437 #endif // !_TARGET_X86_
8439 Phases previousCompletedPhase; // the most recently completed phase
8441 //-------------------------------------------------------------------------
8442 // The following keeps track of how many bytes of local frame space we've
8443 // grabbed so far in the current function, and how many argument bytes we
8444 // need to pop when we return.
8447 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8449 // Count of callee-saved regs we pushed in the prolog.
8450 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8451 // In case of Amd64 this doesn't include float regs saved on stack.
8452 unsigned compCalleeRegsPushed;
8454 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8455 // Mask of callee saved float regs on stack.
8456 regMaskTP compCalleeFPRegsSavedMask;
8458 #ifdef _TARGET_AMD64_
8459 // Quirk for VS debug-launch scenario to work:
8460 // Bytes of padding between save-reg area and locals.
8461 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8462 unsigned compVSQuirkStackPaddingNeeded;
8463 bool compQuirkForPPPflag;
8466 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8468 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8469 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8470 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8472 //-------------------------------------------------------------------------
8474 static void compStartup(); // One-time initialization
8475 static void compShutdown(); // One-time finalization
8477 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8480 static void compDisplayStaticSizes(FILE* fout);
8482 //------------ Some utility functions --------------
8484 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8485 void** ppIndirection); /* OUT */
8487 // Several JIT/EE interface functions return a CorInfoType, and also return a
8488 // class handle as an out parameter if the type is a value class. Returns the
8489 // size of the type these describe.
8490 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8493 // Components used by the compiler may write unit test suites, and
8494 // have them run within this method. They will be run only once per process, and only
8495 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8496 // These should fail by asserting.
8497 void compDoComponentUnitTestsOnce();
8500 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8501 CORINFO_MODULE_HANDLE classPtr,
8502 COMP_HANDLE compHnd,
8503 CORINFO_METHOD_INFO* methodInfo,
8504 void** methodCodePtr,
8505 ULONG* methodCodeSize,
8506 JitFlags* compileFlags);
8507 void compCompileFinish();
8508 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8509 COMP_HANDLE compHnd,
8510 CORINFO_METHOD_INFO* methodInfo,
8511 void** methodCodePtr,
8512 ULONG* methodCodeSize,
8513 JitFlags* compileFlags,
8514 CorInfoInstantiationVerification instVerInfo);
8516 ArenaAllocator* compGetAllocator();
8518 #if MEASURE_MEM_ALLOC
8520 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8524 unsigned allocCnt; // # of allocs
8525 UINT64 allocSz; // total size of those alloc.
8526 UINT64 allocSzMax; // Maximum single allocation.
8527 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8528 UINT64 nraTotalSizeAlloc;
8529 UINT64 nraTotalSizeUsed;
8531 static const char* s_CompMemKindNames[]; // Names of the kinds.
8533 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8535 for (int i = 0; i < CMK_Count; i++)
8537 allocSzByKind[i] = 0;
8540 MemStats(const MemStats& ms)
8541 : allocCnt(ms.allocCnt)
8542 , allocSz(ms.allocSz)
8543 , allocSzMax(ms.allocSzMax)
8544 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8545 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8547 for (int i = 0; i < CMK_Count; i++)
8549 allocSzByKind[i] = ms.allocSzByKind[i];
8553 // Until we have ubiquitous constructors.
8556 this->MemStats::MemStats();
8559 void AddAlloc(size_t sz, CompMemKind cmk)
8563 if (sz > allocSzMax)
8567 allocSzByKind[cmk] += sz;
8570 void Print(FILE* f); // Print these stats to f.
8571 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8573 MemStats genMemStats;
8575 struct AggregateMemStats : public MemStats
8579 AggregateMemStats() : MemStats(), nMethods(0)
8583 void Add(const MemStats& ms)
8586 allocCnt += ms.allocCnt;
8587 allocSz += ms.allocSz;
8588 allocSzMax = max(allocSzMax, ms.allocSzMax);
8589 for (int i = 0; i < CMK_Count; i++)
8591 allocSzByKind[i] += ms.allocSzByKind[i];
8593 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8594 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8597 void Print(FILE* f); // Print these stats to jitstdout.
8600 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8601 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8602 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8604 #endif // MEASURE_MEM_ALLOC
8606 #if LOOP_HOIST_STATS
8607 unsigned m_loopsConsidered;
8608 bool m_curLoopHasHoistedExpression;
8609 unsigned m_loopsWithHoistedExpressions;
8610 unsigned m_totalHoistedExpressions;
8612 void AddLoopHoistStats();
8613 void PrintPerMethodLoopHoistStats();
8615 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8616 static unsigned s_loopsConsidered;
8617 static unsigned s_loopsWithHoistedExpressions;
8618 static unsigned s_totalHoistedExpressions;
8620 static void PrintAggregateLoopHoistStats(FILE* f);
8621 #endif // LOOP_HOIST_STATS
8623 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8624 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8625 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8626 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8627 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8628 void compFreeMem(void*);
8630 bool compIsForImportOnly();
8631 bool compIsForInlining();
8632 bool compDonotInline();
8635 const char* compLocalVarName(unsigned varNum, unsigned offs);
8636 VarName compVarName(regNumber reg, bool isFloatReg = false);
8637 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8638 const char* compRegPairName(regPairNo regPair);
8639 const char* compRegNameForSize(regNumber reg, size_t size);
8640 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8641 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8642 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8645 //-------------------------------------------------------------------------
8647 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8649 struct VarScopeMapInfo
8651 VarScopeListNode* head;
8652 VarScopeListNode* tail;
8653 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8655 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8662 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8663 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8665 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8666 VarNumToScopeDscMap;
8668 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8669 VarNumToScopeDscMap* compVarScopeMap;
8671 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8673 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8675 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8677 void compInitVarScopeMap();
8679 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8680 // enter scope, sorted by instr offset
8681 unsigned compNextEnterScope;
8683 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8684 // go out of scope, sorted by instr offset
8685 unsigned compNextExitScope;
8687 void compInitScopeLists();
8689 void compResetScopeLists();
8691 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8693 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8695 void compProcessScopesUntil(unsigned offset,
8697 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8698 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8701 void compDispScopeLists();
8704 bool compIsProfilerHookNeeded();
8706 //-------------------------------------------------------------------------
8707 /* Statistical Data Gathering */
8709 void compJitStats(); // call this function and enable
8710 // various ifdef's below for statistical data
8713 void compCallArgStats();
8714 static void compDispCallArgStats(FILE* fout);
8717 //-------------------------------------------------------------------------
8724 ArenaAllocator* compAllocator;
8727 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8728 // suitable for use by utilcode collection types.
8729 IAllocator* compAsIAllocator;
8731 #if MEASURE_MEM_ALLOC
8732 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8733 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8734 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8736 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8738 #endif // MEASURE_MEM_ALLOC
8740 void compFunctionTraceStart();
8741 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8744 size_t compMaxUncheckedOffsetForNullObject;
8746 void compInitOptions(JitFlags* compileFlags);
8748 void compSetProcessor();
8749 void compInitDebuggingInfo();
8750 void compSetOptimizationLevel();
8751 #ifdef _TARGET_ARMARCH_
8752 bool compRsvdRegCheck(FrameLayoutState curState);
8754 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8756 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8757 void ResetOptAnnotations();
8759 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8760 void RecomputeLoopInfo();
8762 #ifdef PROFILING_SUPPORTED
8763 // Data required for generating profiler Enter/Leave/TailCall hooks
8765 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8766 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8767 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8770 #ifdef _TARGET_AMD64_
8771 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8774 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8775 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8777 IAllocator* getAllocator()
8779 return compAsIAllocator;
8782 #if MEASURE_MEM_ALLOC
8783 IAllocator* getAllocatorBitset()
8785 return compAsIAllocatorBitset;
8787 IAllocator* getAllocatorGC()
8789 return compAsIAllocatorGC;
8791 IAllocator* getAllocatorLoopHoist()
8793 return compAsIAllocatorLoopHoist;
8795 #else // !MEASURE_MEM_ALLOC
8796 IAllocator* getAllocatorBitset()
8798 return compAsIAllocator;
8800 IAllocator* getAllocatorGC()
8802 return compAsIAllocator;
8804 IAllocator* getAllocatorLoopHoist()
8806 return compAsIAllocator;
8808 #endif // !MEASURE_MEM_ALLOC
8811 IAllocator* getAllocatorDebugOnly()
8813 #if MEASURE_MEM_ALLOC
8814 return compAsIAllocatorDebugOnly;
8815 #else // !MEASURE_MEM_ALLOC
8816 return compAsIAllocator;
8817 #endif // !MEASURE_MEM_ALLOC
8822 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8823 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8827 XX Checks for type compatibility and merges types XX
8829 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8830 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8834 // Set to TRUE if verification cannot be skipped for this method
8835 // If we detect unverifiable code, we will lazily check
8836 // canSkipMethodVerification() to see if verification is REALLY needed.
8837 BOOL tiVerificationNeeded;
8839 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8840 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8841 BOOL tiIsVerifiableCode;
8843 // Set to TRUE if runtime callout is needed for this method
8844 BOOL tiRuntimeCalloutNeeded;
8846 // Set to TRUE if security prolog/epilog callout is needed for this method
8847 // Note: This flag is different than compNeedSecurityCheck.
8848 // compNeedSecurityCheck means whether or not a security object needs
8849 // to be allocated on the stack, which is currently true for EnC as well.
8850 // tiSecurityCalloutNeeded means whether or not security callouts need
8851 // to be inserted in the jitted code.
8852 BOOL tiSecurityCalloutNeeded;
8854 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8855 // This support is necessary to suport attributes that are not described in
8856 // for example, signatures. For example, the permanent home byref (byref that
8857 // points to the gc heap), isn't a property of method signatures, therefore,
8858 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8859 // but when deciding if we need to reimport a block, we need to take these
8861 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8863 // Returns TRUE if child is equal to or a subtype of parent.
8864 // normalisedForStack indicates that both types are normalised for the stack
8865 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8867 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8868 // *pDest is modified to represent the merged type. Sets "*changed" to true
8869 // if this changes "*pDest".
8870 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8872 // Set pDest from the primitive value type.
8873 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8875 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8878 // <BUGNUM> VSW 471305
8879 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8880 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8881 // We use a "short" as we need to push/pop this scope.
8883 short compRegSetCheckLevel;
8887 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8888 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8890 XX IL verification stuff XX
8893 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8894 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8898 // The following is used to track liveness of local variables, initialization
8899 // of valueclass constructors, and type safe use of IL instructions.
8901 // dynamic state info needed for verification
8902 EntryState verCurrentState;
8904 // this ptr of object type .ctors are considered intited only after
8905 // the base class ctor is called, or an alternate ctor is called.
8906 // An uninited this ptr can be used to access fields, but cannot
8907 // be used to call a member function.
8908 BOOL verTrackObjCtorInitState;
8910 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8912 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8913 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8914 void verInitCurrentState();
8915 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8917 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8918 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8919 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8921 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8922 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8923 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8924 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8925 typeInfo verMakeTypeInfo(CorInfoType ciType,
8926 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8927 BOOL verIsSDArray(typeInfo ti);
8928 typeInfo verGetArrayElemType(typeInfo ti);
8930 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8931 BOOL verNeedsVerification();
8932 BOOL verIsByRefLike(const typeInfo& ti);
8933 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8935 // generic type variables range over types that satisfy IsBoxable
8936 BOOL verIsBoxable(const typeInfo& ti);
8938 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8939 DEBUGARG(unsigned line));
8940 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8941 DEBUGARG(unsigned line));
8942 bool verCheckTailCallConstraint(OPCODE opcode,
8943 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8944 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8945 // on a type parameter?
8946 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8947 // return false to the caller.
8948 // If false, it will throw.
8950 bool verIsBoxedValueType(typeInfo ti);
8952 void verVerifyCall(OPCODE opcode,
8953 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8954 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8956 bool readonlyCall, // is this a "readonly." call?
8957 const BYTE* delegateCreateStart,
8958 const BYTE* codeAddr,
8959 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8961 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8963 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8964 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8965 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8966 const CORINFO_FIELD_INFO& fieldInfo,
8967 const typeInfo* tiThis,
8969 BOOL allowPlainStructAsThis = FALSE);
8970 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8971 void verVerifyThisPtrInitialised();
8972 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8974 // Register allocator
8975 void raInitStackFP();
8976 void raEnregisterVarsPrePassStackFP();
8977 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8978 void raEnregisterVarsPostPassStackFP();
8979 void raGenerateFPRefCounts();
8980 void raEnregisterVarsStackFP();
8981 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8983 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8984 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8986 // returns true if enregistering v1 would save more mem accesses than v2
8987 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8990 void raDumpHeightsStackFP();
8991 void raDumpVariableRegIntfFloat();
8994 #if FEATURE_STACK_FP_X87
8996 // Currently, we use FP transition blocks in only 2 situations:
8998 // -conditional jump on longs where FP stack differs with target: it's not strictly
8999 // necessary, but its low frequency and the code would get complicated if we try to
9000 // inline the FP stack adjustment, as we have a lot of special casing going on to try
9001 // minimize the way we generate the jump code.
9002 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
9003 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
9005 // However, transition blocks have 2 problems
9007 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9008 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9009 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9010 // in the right place without preordering them), this causes us to have to generate the transition
9011 // blocks in the cold area if we want procedure splitting.
9014 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9015 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9016 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9017 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9018 // a big change in the exception.
9020 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9021 // optimizations. For these 2 cases:
9023 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9024 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9025 // a switch statement.
9027 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9028 // current procedure splitting and exception code have.
9029 bool compMayHaveTransitionBlocks;
9031 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9033 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9035 unsigned raCntStkStackFP;
9036 unsigned raCntWtdStkDblStackFP;
9037 unsigned raCntStkParamDblStackFP;
9039 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9040 // TODO: Do we want to put this in LclVarDsc?
9041 unsigned raPayloadStackFP[lclMAX_TRACKED];
9042 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9044 // Useful for debugging
9045 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9047 #endif // FEATURE_STACK_FP_X87
9050 // One line log function. Default level is 0. Increasing it gives you
9051 // more log information
9053 // levels are currently unused: #define JITDUMP(level,...) ();
9054 void JitLogEE(unsigned level, const char* fmt, ...);
9056 bool compDebugBreak;
9058 bool compJitHaltMethod();
9063 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9064 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9066 XX GS Security checks for unsafe buffers XX
9068 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9069 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9072 struct ShadowParamVarInfo
9074 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9075 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9077 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9079 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9080 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9081 // slots and update all trees to refer to shadow slots is done immediately after
9082 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9083 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9084 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9085 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9086 // creating a shadow slot even though this routine returns true.
9088 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9089 // required. There are two cases under which a reg arg could potentially be used from its
9091 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9092 // b) LSRA spills it
9094 // Possible solution to address case (a)
9095 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9096 // in this routine. Note that live out of exception handler is something we may not be
9097 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9098 // Therefore, for methods with exception handling and need GS cookie check we might have
9099 // to take conservative approach.
9101 // Possible solution to address case (b)
9102 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9103 // create a new spill temp if the method needs GS cookie check.
9104 return varDsc->lvIsParam;
9105 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9106 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9113 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9118 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9119 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9120 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9122 void gsGSChecksInitCookie(); // Grabs cookie variable
9123 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9124 bool gsFindVulnerableParams(); // Shadow param analysis code
9125 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9127 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9128 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9130 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9131 // This can be overwritten by setting complus_JITInlineSize env variable.
9133 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9136 #ifdef FEATURE_JIT_METHOD_PERF
9137 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9138 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9140 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9141 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9143 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9145 #if MEASURE_CLRAPI_CALLS
9146 // Thin wrappers that call into JitTimer (if present).
9147 inline void CLRApiCallEnter(unsigned apix);
9148 inline void CLRApiCallLeave(unsigned apix);
9151 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9152 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9157 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9158 // These variables are associated with maintaining SQM data about compile time.
9159 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9160 // in the current compilation.
9161 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9162 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9163 // the inlining phase in the current compilation.
9164 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9166 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9167 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9168 // type-loading and class initialization).
9169 void RecordStateAtEndOfInlining();
9170 // Assumes being called at the end of compilation. Update the SQM state.
9171 void RecordStateAtEndOfCompilation();
9173 #ifdef FEATURE_CLRSQM
9174 // Does anything SQM related necessary at process shutdown time.
9175 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9176 #endif // FEATURE_CLRSQM
9179 #if FUNC_INFO_LOGGING
9180 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9181 // filename to write it to.
9182 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9183 #endif // FUNC_INFO_LOGGING
9185 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9187 // Is the compilation in a full trust context?
9188 bool compIsFullTrust();
9191 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9192 #endif // MEASURE_NOWAY
9194 #ifndef FEATURE_TRACELOGGING
9195 // Should we actually fire the noway assert body and the exception handler?
9196 bool compShouldThrowOnNoway();
9197 #else // FEATURE_TRACELOGGING
9198 // Should we actually fire the noway assert body and the exception handler?
9199 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9201 // Telemetry instance to use per method compilation.
9202 JitTelemetry compJitTelemetry;
9204 // Get common parameters that have to be logged with most telemetry data.
9205 void compGetTelemetryDefaults(const char** assemblyName,
9206 const char** scopeName,
9207 const char** methodName,
9208 unsigned* methodHash);
9209 #endif // !FEATURE_TRACELOGGING
9213 NodeToTestDataMap* m_nodeTestData;
9215 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9216 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9217 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9218 // Current kept in this.
9220 NodeToTestDataMap* GetNodeTestData()
9222 Compiler* compRoot = impInlineRoot();
9223 if (compRoot->m_nodeTestData == nullptr)
9225 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9227 return compRoot->m_nodeTestData;
9230 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9232 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9233 // currently occur in the AST graph.
9234 NodeToIntMap* FindReachableNodesInNodeTestData();
9236 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9237 // test data, associate that data with "to".
9238 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9240 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9241 // have annotations, attach similar annotations to the corresponding nodes in "to".
9242 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9244 // These are the methods that test that the various conditions implied by the
9245 // test attributes are satisfied.
9246 void JitTestCheckSSA(); // SSA builder tests.
9247 void JitTestCheckVN(); // Value numbering tests.
9250 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9252 FieldSeqStore* m_fieldSeqStore;
9254 FieldSeqStore* GetFieldSeqStore()
9256 Compiler* compRoot = impInlineRoot();
9257 if (compRoot->m_fieldSeqStore == nullptr)
9259 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9260 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9261 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9263 return compRoot->m_fieldSeqStore;
9266 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9268 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9269 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9270 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9271 // attach the field sequence directly to the address node.
9272 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9274 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9276 // Don't need to worry about inlining here
9277 if (m_zeroOffsetFieldMap == nullptr)
9279 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9281 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9282 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9284 return m_zeroOffsetFieldMap;
9287 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9288 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9289 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9290 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9291 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9292 // record the the field sequence using the ZeroOffsetFieldMap described above.
9294 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9295 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9296 // CoreRT. Such case is handled same as the default case.
9297 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9299 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9301 NodeToArrayInfoMap* m_arrayInfoMap;
9303 NodeToArrayInfoMap* GetArrayInfoMap()
9305 Compiler* compRoot = impInlineRoot();
9306 if (compRoot->m_arrayInfoMap == nullptr)
9308 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9309 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9310 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9312 return compRoot->m_arrayInfoMap;
9315 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9317 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9318 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9319 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9320 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9322 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9324 // Use the same map for GCHeap and ByrefExposed when their states match.
9325 memoryKind = ByrefExposed;
9328 assert(memoryKind < MemoryKindCount);
9329 Compiler* compRoot = impInlineRoot();
9330 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9332 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9333 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9334 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9336 return compRoot->m_memorySsaMap[memoryKind];
9339 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9340 CORINFO_CLASS_HANDLE m_refAnyClass;
9341 CORINFO_FIELD_HANDLE GetRefanyDataField()
9343 if (m_refAnyClass == nullptr)
9345 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9347 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9349 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9351 if (m_refAnyClass == nullptr)
9353 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9355 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9359 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9361 #if ALLVARSET_COUNTOPS
9362 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9365 static HelperCallProperties s_helperCallProperties;
9367 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9368 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9369 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9371 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9374 unsigned __int8* offset0,
9375 unsigned __int8* offset1);
9376 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9377 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9379 void fgMorphMultiregStructArgs(GenTreeCall* call);
9380 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9382 }; // end of class Compiler
9384 // Inline methods of CompAllocator.
9385 void* CompAllocator::Alloc(size_t sz)
9387 #if MEASURE_MEM_ALLOC
9388 return m_comp->compGetMem(sz, m_cmk);
9390 return m_comp->compGetMem(sz);
9394 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9396 #if MEASURE_MEM_ALLOC
9397 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9399 return m_comp->compGetMemArray(elems, elemSize);
9403 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9404 inline LclVarDsc::LclVarDsc(Compiler* comp)
9405 : // Initialize the ArgRegs to REG_STK.
9406 // The morph will do the right thing to change
9407 // to the right register if passed in register.
9410 #if FEATURE_MULTIREG_ARGS
9411 _lvOtherArgReg(REG_STK)
9413 #endif // FEATURE_MULTIREG_ARGS
9415 lvRefBlks(BlockSetOps::UninitVal())
9417 #endif // ASSERTION_PROP
9418 lvPerSsaData(comp->getAllocator())
9422 //---------------------------------------------------------------------------------------------------------------------
9423 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9425 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9426 // shown in parentheses):
9428 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9429 // of a misnomer, as the first entry will always be the current node.
9431 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9432 // argument before visiting the node's operands.
9434 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9435 // argument after visiting the node's operands.
9437 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9438 // `DoPreOrder` must be true if this option is true.
9440 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9441 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9442 // visited before the first).
9444 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9446 // A simple pre-order visitor might look something like the following:
9448 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9453 // DoPreOrder = true
9456 // unsigned m_count;
9458 // CountingVisitor(Compiler* compiler)
9459 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9463 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9469 // This visitor would then be used like so:
9471 // CountingVisitor countingVisitor(compiler);
9472 // countingVisitor.WalkTree(root);
9474 template <typename TVisitor>
9475 class GenTreeVisitor
9478 typedef Compiler::fgWalkResult fgWalkResult;
9482 ComputeStack = false,
9484 DoPostOrder = false,
9485 DoLclVarsOnly = false,
9486 UseExecutionOrder = false,
9489 Compiler* m_compiler;
9490 ArrayStack<GenTree*> m_ancestors;
9492 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9494 assert(compiler != nullptr);
9496 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9497 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9500 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9502 return fgWalkResult::WALK_CONTINUE;
9505 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9507 return fgWalkResult::WALK_CONTINUE;
9511 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9513 assert(use != nullptr);
9515 GenTree* node = *use;
9517 if (TVisitor::ComputeStack)
9519 m_ancestors.Push(node);
9522 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9523 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9525 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9526 if (result == fgWalkResult::WALK_ABORT)
9532 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9538 switch (node->OperGet())
9543 case GT_LCL_VAR_ADDR:
9544 case GT_LCL_FLD_ADDR:
9545 if (TVisitor::DoLclVarsOnly)
9547 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9548 if (result == fgWalkResult::WALK_ABORT)
9564 case GT_MEMORYBARRIER:
9569 case GT_START_NONGC:
9571 #if !FEATURE_EH_FUNCLETS
9573 #endif // !FEATURE_EH_FUNCLETS
9575 #ifndef LEGACY_BACKEND
9577 #endif // LEGACY_BACKEND
9580 case GT_CLS_VAR_ADDR:
9584 case GT_PINVOKE_PROLOG:
9585 case GT_PINVOKE_EPILOG:
9589 // Lclvar unary operators
9590 case GT_STORE_LCL_VAR:
9591 case GT_STORE_LCL_FLD:
9592 if (TVisitor::DoLclVarsOnly)
9594 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9595 if (result == fgWalkResult::WALK_ABORT)
9602 // Standard unary operators
9630 GenTreeUnOp* const unOp = node->AsUnOp();
9631 if (unOp->gtOp1 != nullptr)
9633 result = WalkTree(&unOp->gtOp1, unOp);
9634 if (result == fgWalkResult::WALK_ABORT)
9645 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9647 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9648 if (result == fgWalkResult::WALK_ABORT)
9652 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9653 if (result == fgWalkResult::WALK_ABORT)
9657 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9658 if (result == fgWalkResult::WALK_ABORT)
9665 case GT_ARR_BOUNDS_CHECK:
9668 #endif // FEATURE_SIMD
9670 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9672 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9673 if (result == fgWalkResult::WALK_ABORT)
9677 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9678 if (result == fgWalkResult::WALK_ABORT)
9687 GenTreeField* const field = node->AsField();
9689 if (field->gtFldObj != nullptr)
9691 result = WalkTree(&field->gtFldObj, field);
9692 if (result == fgWalkResult::WALK_ABORT)
9702 GenTreeArrElem* const arrElem = node->AsArrElem();
9704 result = WalkTree(&arrElem->gtArrObj, arrElem);
9705 if (result == fgWalkResult::WALK_ABORT)
9710 const unsigned rank = arrElem->gtArrRank;
9711 for (unsigned dim = 0; dim < rank; dim++)
9713 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9714 if (result == fgWalkResult::WALK_ABORT)
9724 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9726 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9727 if (result == fgWalkResult::WALK_ABORT)
9731 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9732 if (result == fgWalkResult::WALK_ABORT)
9736 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9737 if (result == fgWalkResult::WALK_ABORT)
9746 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9748 GenTree** op1Use = &dynBlock->gtOp1;
9749 GenTree** op2Use = &dynBlock->gtDynamicSize;
9751 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9753 std::swap(op1Use, op2Use);
9756 result = WalkTree(op1Use, dynBlock);
9757 if (result == fgWalkResult::WALK_ABORT)
9761 result = WalkTree(op2Use, dynBlock);
9762 if (result == fgWalkResult::WALK_ABORT)
9769 case GT_STORE_DYN_BLK:
9771 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9773 GenTree** op1Use = &dynBlock->gtOp1;
9774 GenTree** op2Use = &dynBlock->gtOp2;
9775 GenTree** op3Use = &dynBlock->gtDynamicSize;
9777 if (TVisitor::UseExecutionOrder)
9779 if (dynBlock->IsReverseOp())
9781 std::swap(op1Use, op2Use);
9783 if (dynBlock->gtEvalSizeFirst)
9785 std::swap(op3Use, op2Use);
9786 std::swap(op2Use, op1Use);
9790 result = WalkTree(op1Use, dynBlock);
9791 if (result == fgWalkResult::WALK_ABORT)
9795 result = WalkTree(op2Use, dynBlock);
9796 if (result == fgWalkResult::WALK_ABORT)
9800 result = WalkTree(op3Use, dynBlock);
9801 if (result == fgWalkResult::WALK_ABORT)
9810 GenTreeCall* const call = node->AsCall();
9812 if (call->gtCallObjp != nullptr)
9814 result = WalkTree(&call->gtCallObjp, call);
9815 if (result == fgWalkResult::WALK_ABORT)
9821 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9823 result = WalkTree(args->pCurrent(), call);
9824 if (result == fgWalkResult::WALK_ABORT)
9830 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9832 result = WalkTree(args->pCurrent(), call);
9833 if (result == fgWalkResult::WALK_ABORT)
9839 if (call->gtCallType == CT_INDIRECT)
9841 if (call->gtCallCookie != nullptr)
9843 result = WalkTree(&call->gtCallCookie, call);
9844 if (result == fgWalkResult::WALK_ABORT)
9850 result = WalkTree(&call->gtCallAddr, call);
9851 if (result == fgWalkResult::WALK_ABORT)
9857 if (call->gtControlExpr != nullptr)
9859 result = WalkTree(&call->gtControlExpr, call);
9860 if (result == fgWalkResult::WALK_ABORT)
9872 assert(node->OperIsBinary());
9874 GenTreeOp* const op = node->AsOp();
9876 GenTree** op1Use = &op->gtOp1;
9877 GenTree** op2Use = &op->gtOp2;
9879 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9881 std::swap(op1Use, op2Use);
9884 if (*op1Use != nullptr)
9886 result = WalkTree(op1Use, op);
9887 if (result == fgWalkResult::WALK_ABORT)
9893 if (*op2Use != nullptr)
9895 result = WalkTree(op2Use, op);
9896 if (result == fgWalkResult::WALK_ABORT)
9906 // Finally, visit the current node
9907 if (TVisitor::DoPostOrder)
9909 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9912 if (TVisitor::ComputeStack)
9921 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9922 class GenericTreeWalker final
9923 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
9928 ComputeStack = computeStack,
9929 DoPreOrder = doPreOrder,
9930 DoPostOrder = doPostOrder,
9931 DoLclVarsOnly = doLclVarsOnly,
9932 UseExecutionOrder = useExecutionOrder,
9936 Compiler::fgWalkData* m_walkData;
9939 GenericTreeWalker(Compiler::fgWalkData* walkData)
9940 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
9942 , m_walkData(walkData)
9944 assert(walkData != nullptr);
9948 walkData->parentStack = &this->m_ancestors;
9952 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9954 m_walkData->parent = user;
9955 return m_walkData->wtprVisitorFn(use, m_walkData);
9958 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9960 m_walkData->parent = user;
9961 return m_walkData->wtpoVisitorFn(use, m_walkData);
9965 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
9971 DoLclVarsOnly = true
9974 IncLclVarRefCountsVisitor(Compiler* compiler);
9975 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9977 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9980 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
9986 DoLclVarsOnly = true
9989 DecLclVarRefCountsVisitor(Compiler* compiler);
9990 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9992 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9996 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9997 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9999 XX Miscellaneous Compiler stuff XX
10001 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10002 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10005 // Values used to mark the types a stack slot is used for
10007 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10008 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10009 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10010 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10011 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10012 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10013 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10014 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10016 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10018 /*****************************************************************************
10020 * Variables to keep track of total code amounts.
10025 extern size_t grossVMsize;
10026 extern size_t grossNCsize;
10027 extern size_t totalNCsize;
10029 extern unsigned genMethodICnt;
10030 extern unsigned genMethodNCnt;
10031 extern size_t gcHeaderISize;
10032 extern size_t gcPtrMapISize;
10033 extern size_t gcHeaderNSize;
10034 extern size_t gcPtrMapNSize;
10036 #endif // DISPLAY_SIZES
10038 /*****************************************************************************
10040 * Variables to keep track of basic block counts (more data on 1 BB methods)
10043 #if COUNT_BASIC_BLOCKS
10044 extern Histogram bbCntTable;
10045 extern Histogram bbOneBBSizeTable;
10048 /*****************************************************************************
10050 * Used by optFindNaturalLoops to gather statistical information such as
10051 * - total number of natural loops
10052 * - number of loops with 1, 2, ... exit conditions
10053 * - number of loops that have an iterator (for like)
10054 * - number of loops that have a constant iterator
10059 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10060 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10061 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10062 extern unsigned totalLoopCount; // counts the total number of natural loops
10063 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10064 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10065 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10066 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10068 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10069 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10070 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10071 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10072 extern Histogram loopCountTable; // Histogram of loop counts
10073 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10075 #endif // COUNT_LOOPS
10077 /*****************************************************************************
10078 * variables to keep track of how many iterations we go in a dataflow pass
10083 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10084 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10086 #endif // DATAFLOW_ITER
10088 #if MEASURE_BLOCK_SIZE
10089 extern size_t genFlowNodeSize;
10090 extern size_t genFlowNodeCnt;
10091 #endif // MEASURE_BLOCK_SIZE
10093 #if MEASURE_NODE_SIZE
10094 struct NodeSizeStats
10098 genTreeNodeCnt = 0;
10099 genTreeNodeSize = 0;
10100 genTreeNodeActualSize = 0;
10103 // Count of tree nodes allocated.
10104 unsigned __int64 genTreeNodeCnt;
10106 // The size we allocate.
10107 unsigned __int64 genTreeNodeSize;
10109 // The actual size of the node. Note that the actual size will likely be smaller
10110 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10111 // a smaller node to a larger one. TODO-Cleanup: add stats on
10112 // SetOper()/ChangeOper() usage to quantify this.
10113 unsigned __int64 genTreeNodeActualSize;
10115 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10116 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10117 extern Histogram genTreeNcntHist;
10118 extern Histogram genTreeNsizHist;
10119 #endif // MEASURE_NODE_SIZE
10121 /*****************************************************************************
10122 * Count fatal errors (including noway_asserts).
10126 extern unsigned fatal_badCode;
10127 extern unsigned fatal_noWay;
10128 extern unsigned fatal_NOMEM;
10129 extern unsigned fatal_noWayAssertBody;
10131 extern unsigned fatal_noWayAssertBodyArgs;
10133 extern unsigned fatal_NYI;
10134 #endif // MEASURE_FATAL
10136 /*****************************************************************************
10140 #ifdef _TARGET_XARCH_
10142 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10143 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10144 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10146 const instruction INS_AND = INS_and;
10147 const instruction INS_OR = INS_or;
10148 const instruction INS_XOR = INS_xor;
10149 const instruction INS_NEG = INS_neg;
10150 const instruction INS_TEST = INS_test;
10151 const instruction INS_MUL = INS_imul;
10152 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10153 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10154 const instruction INS_BREAKPOINT = INS_int3;
10155 const instruction INS_ADDC = INS_adc;
10156 const instruction INS_SUBC = INS_sbb;
10157 const instruction INS_NOT = INS_not;
10161 #ifdef _TARGET_ARM_
10163 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10164 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10165 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10167 const instruction INS_AND = INS_and;
10168 const instruction INS_OR = INS_orr;
10169 const instruction INS_XOR = INS_eor;
10170 const instruction INS_NEG = INS_rsb;
10171 const instruction INS_TEST = INS_tst;
10172 const instruction INS_MUL = INS_mul;
10173 const instruction INS_MULADD = INS_mla;
10174 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10175 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10176 const instruction INS_BREAKPOINT = INS_bkpt;
10177 const instruction INS_ADDC = INS_adc;
10178 const instruction INS_SUBC = INS_sbc;
10179 const instruction INS_NOT = INS_mvn;
10181 const instruction INS_ABS = INS_vabs;
10182 const instruction INS_ROUND = INS_invalid;
10183 const instruction INS_SQRT = INS_vsqrt;
10187 #ifdef _TARGET_ARM64_
10189 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10190 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10191 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10193 const instruction INS_AND = INS_and;
10194 const instruction INS_OR = INS_orr;
10195 const instruction INS_XOR = INS_eor;
10196 const instruction INS_NEG = INS_neg;
10197 const instruction INS_TEST = INS_tst;
10198 const instruction INS_MUL = INS_mul;
10199 const instruction INS_MULADD = INS_madd;
10200 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10201 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10202 const instruction INS_BREAKPOINT = INS_bkpt;
10203 const instruction INS_ADDC = INS_adc;
10204 const instruction INS_SUBC = INS_sbc;
10205 const instruction INS_NOT = INS_mvn;
10207 const instruction INS_ABS = INS_fabs;
10208 const instruction INS_ROUND = INS_frintn;
10209 const instruction INS_SQRT = INS_fsqrt;
10213 /*****************************************************************************/
10215 extern const BYTE genTypeSizes[];
10216 extern const BYTE genTypeAlignments[];
10217 extern const BYTE genTypeStSzs[];
10218 extern const BYTE genActualTypes[];
10220 /*****************************************************************************/
10222 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10223 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10225 #ifdef _TARGET_ARM_
10226 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10227 #elif defined(_TARGET_ARM64_)
10228 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10231 /*****************************************************************************/
10233 #define REG_CORRUPT regNumber(REG_NA + 1)
10234 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10235 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10237 /*****************************************************************************/
10239 extern BasicBlock dummyBB;
10241 /*****************************************************************************/
10242 /*****************************************************************************/
10244 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10245 // nodes of a statement in execution order.
10246 // __stmt: a GT_STMT type GenTree*
10247 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10249 #define foreach_treenode_execution_order(__node, __stmt) \
10250 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10252 // foreach_block: An iterator over all blocks in the function.
10253 // __compiler: the Compiler* object
10254 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10256 #define foreach_block(__compiler, __block) \
10257 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10259 /*****************************************************************************/
10260 /*****************************************************************************/
10264 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10266 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10267 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10269 XX Debugging helpers XX
10271 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10272 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10275 /*****************************************************************************/
10276 /* The following functions are intended to be called from the debugger, to dump
10277 * various data structures. The can be used in the debugger Watch or Quick Watch
10278 * windows. They are designed to be short to type and take as few arguments as
10279 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10280 * See the function definition comment for more details.
10283 void cBlock(Compiler* comp, BasicBlock* block);
10284 void cBlocks(Compiler* comp);
10285 void cBlocksV(Compiler* comp);
10286 void cTree(Compiler* comp, GenTree* tree);
10287 void cTrees(Compiler* comp);
10288 void cEH(Compiler* comp);
10289 void cVar(Compiler* comp, unsigned lclNum);
10290 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10291 void cVars(Compiler* comp);
10292 void cVarsFinal(Compiler* comp);
10293 void cBlockPreds(Compiler* comp, BasicBlock* block);
10294 void cReach(Compiler* comp);
10295 void cDoms(Compiler* comp);
10296 void cLiveness(Compiler* comp);
10297 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10299 void cFuncIR(Compiler* comp);
10300 void cBlockIR(Compiler* comp, BasicBlock* block);
10301 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10302 void cTreeIR(Compiler* comp, GenTree* tree);
10303 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10304 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10305 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10306 int cOperandIR(Compiler* comp, GenTree* operand);
10307 int cLeafIR(Compiler* comp, GenTree* tree);
10308 int cIndirIR(Compiler* comp, GenTree* tree);
10309 int cListIR(Compiler* comp, GenTree* list);
10310 int cSsaNumIR(Compiler* comp, GenTree* tree);
10311 int cValNumIR(Compiler* comp, GenTree* tree);
10312 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10314 void dBlock(BasicBlock* block);
10317 void dTree(GenTree* tree);
10320 void dVar(unsigned lclNum);
10321 void dVarDsc(LclVarDsc* varDsc);
10324 void dBlockPreds(BasicBlock* block);
10328 void dCVarSet(VARSET_VALARG_TP vars);
10330 void dVarSet(VARSET_VALARG_TP vars);
10331 void dRegMask(regMaskTP mask);
10334 void dBlockIR(BasicBlock* block);
10335 void dTreeIR(GenTree* tree);
10336 void dLoopIR(Compiler::LoopDsc* loop);
10337 void dLoopNumIR(unsigned loopNum);
10338 int dTabStopIR(int curr, int tabstop);
10339 int dTreeTypeIR(GenTree* tree);
10340 int dTreeKindsIR(GenTree* tree);
10341 int dTreeFlagsIR(GenTree* tree);
10342 int dOperandIR(GenTree* operand);
10343 int dLeafIR(GenTree* tree);
10344 int dIndirIR(GenTree* tree);
10345 int dListIR(GenTree* list);
10346 int dSsaNumIR(GenTree* tree);
10347 int dValNumIR(GenTree* tree);
10348 int dDependsIR(GenTree* comma);
10351 GenTree* dFindTree(GenTree* tree, unsigned id);
10352 GenTree* dFindTree(unsigned id);
10353 GenTreeStmt* dFindStmt(unsigned id);
10354 BasicBlock* dFindBlock(unsigned bbNum);
10358 #include "compiler.hpp" // All the shared inline functions
10360 /*****************************************************************************/
10361 #endif //_COMPILER_H_
10362 /*****************************************************************************/