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 /*****************************************************************************/
29 #include "jithashtable.h"
38 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "jitexpandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
48 #include "namedintrinsiclist.h"
53 #include "codegeninterface.h"
55 #include "jitgcinfo.h"
57 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
65 // This is only used locally in the JIT to indicate that
66 // a verification block should be inserted
67 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
69 /*****************************************************************************
70 * Forward declarations
73 struct InfoHdr; // defined in GCInfo.h
74 struct escapeMapping_t; // defined in flowgraph.cpp
75 class emitter; // defined in emit.h
76 struct ShadowParamVarInfo; // defined in GSChecks.cpp
77 struct InitVarDscInfo; // defined in register_arg_convention.h
78 class FgStack; // defined in flowgraph.cpp
79 #if FEATURE_STACK_FP_X87
80 struct FlatFPStateX87; // defined in fp.h
83 class CSE_DataFlow; // defined in OptCSE.cpp
89 #ifndef LEGACY_BACKEND
90 class Lowering; // defined in lower.h
93 // The following are defined in this file, Compiler.h
97 /*****************************************************************************
103 /*****************************************************************************/
106 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
109 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
110 // caused these to be ambiguous with the global placement new operators.
111 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
113 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
115 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
116 #include "loopcloning.h"
118 /*****************************************************************************/
120 /* This is included here and not earlier as it needs the definition of "CSE"
121 * which is defined in the section above */
123 /*****************************************************************************/
125 unsigned genLog2(unsigned value);
126 unsigned genLog2(unsigned __int64 value);
128 var_types genActualType(var_types type);
129 var_types genUnsignedType(var_types type);
130 var_types genSignedType(var_types type);
132 unsigned ReinterpretHexAsDecimal(unsigned);
134 /*****************************************************************************/
136 #if defined(FEATURE_SIMD)
137 #if defined(_TARGET_XARCH_)
138 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
139 #elif defined(_TARGET_ARM64_)
140 const unsigned TEMP_MAX_SIZE = FP_REGSIZE_BYTES;
141 #endif // defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
142 #else // !FEATURE_SIMD
143 const unsigned TEMP_MAX_SIZE = sizeof(double);
144 #endif // !FEATURE_SIMD
145 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
147 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
150 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
153 // The following holds the Local var info (scope information)
154 typedef const char* VarName; // Actual ASCII string
157 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
158 IL_OFFSET vsdLifeEnd; // instr offset of end of life
159 unsigned vsdVarNum; // (remapped) LclVarDsc number
162 VarName vsdName; // name of the var
165 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
166 // Also, it is the index of this entry in the info.compVarScopes array,
167 // which is useful since the array is also accessed via the
168 // compEnterScopeList and compExitScopeList sorted arrays.
171 /*****************************************************************************
173 * The following holds the local variable counts and the descriptor table.
176 // This is the location of a definition.
182 DefLoc() : m_blk(nullptr), m_tree(nullptr)
187 // This class encapsulates all info about a local variable that may vary for different SSA names
192 ValueNumPair m_vnPair;
200 typedef JitExpandArray<LclSsaVarDsc> PerSsaArray;
205 // The constructor. Most things can just be zero'ed.
206 LclVarDsc(Compiler* comp);
208 // note this only packs because var_types is a typedef of unsigned char
209 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
211 unsigned char lvIsParam : 1; // is this a parameter?
212 unsigned char lvIsRegArg : 1; // is this a register argument?
213 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
215 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
216 // is to help determine whether to use block init in the prolog.
217 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
218 #ifdef LEGACY_BACKEND
219 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
221 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
222 // variable is in the same register for the entire function.
223 unsigned char lvTracked : 1; // is this a tracked variable?
224 bool lvTrackedNonStruct()
226 return lvTracked && lvType != TYP_STRUCT;
228 unsigned char lvPinned : 1; // is this a pinned variable?
230 unsigned char lvMustInit : 1; // must be initialized
231 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
232 // global location, etc.
233 // We cannot reason reliably about the value of the variable.
234 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
235 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
239 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
241 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
242 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
243 // read/write its value.
244 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
245 // the variable to be
246 // in the stack (at least at those boundaries.)
247 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
248 // particular byte from an int).
249 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
250 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
252 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
253 #ifdef LEGACY_BACKEND
254 unsigned char lvRefAssign : 1; // involved in pointer assignment
256 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
257 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
260 unsigned char lvHasILStoreOp : 1; // there is at least one STLOC or STARG on this local
261 unsigned char lvHasMultipleILStoreOp : 1; // there is more than one STLOC on this local
263 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp (if lvIsParam is false), or implicit byref parameter
264 // (if lvIsParam is true)
266 unsigned char lvIsBoolean : 1; // set if variable is boolean
269 unsigned char lvSingleDef : 1; // variable has a single def
270 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
271 unsigned char lvVolatileHint : 1; // hint for AssertionProp
274 #ifdef LEGACY_BACKEND
275 unsigned char lvSpilled : 1; // enregistered variable was spilled
277 #ifndef _TARGET_64BIT_
278 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
279 #endif // !_TARGET_64BIT_
280 #ifdef _TARGET_64BIT_
281 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
284 unsigned char lvKeepType : 1; // Don't change the type of this variable
285 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
287 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
289 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
290 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
291 // 32-bit target. For implicit byref parameters, this gets hijacked between
292 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to indicate whether
293 // references to the arg are being rewritten as references to a promoted shadow local.
294 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
295 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
296 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
297 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
299 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
300 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
303 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
304 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
305 // with (lvIsRegArg && lvIsHfa())
306 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
307 #endif // FEATURE_HFA
310 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
311 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
312 // Consider cleaning this up so this workaround is not required.
313 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
314 // I.e. there is no longer any reference to the struct directly.
315 // In this case we can simply remove this struct local.
317 #ifndef LEGACY_BACKEND
318 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
319 #endif // !LEGACY_BACKEND
322 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
323 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
324 unsigned char lvSIMDType : 1; // This is a SIMD struct
325 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
326 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
327 #endif // FEATURE_SIMD
328 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
330 unsigned char lvClassIsExact : 1; // lvClassHandle is the exact type
333 unsigned char lvClassInfoUpdated : 1; // true if this var has updated class handle or exactness
337 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
338 // local. For implicit byref parameters, this gets hijacked between
339 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to point to the
340 // struct local created to model the parameter's struct promotion, if any.
341 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
342 // Valid on promoted struct local fields.
345 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
346 unsigned char lvFldOffset;
347 unsigned char lvFldOrdinal;
349 #if FEATURE_MULTIREG_ARGS
350 regNumber lvRegNumForSlot(unsigned slotNum)
356 else if (slotNum == 1)
358 return lvOtherArgReg;
362 assert(false && "Invalid slotNum!");
367 #endif // FEATURE_MULTIREG_ARGS
385 bool lvIsHfaRegArg() const
388 return _lvIsHfaRegArg;
394 void lvSetIsHfaRegArg(bool value = true)
397 _lvIsHfaRegArg = value;
401 bool lvHfaTypeIsFloat() const
404 return _lvHfaTypeIsFloat;
410 void lvSetHfaTypeIsFloat(bool value)
413 _lvHfaTypeIsFloat = value;
417 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
418 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
420 unsigned lvHfaSlots() const
423 assert(varTypeIsStruct(lvType));
425 return lvExactSize / sizeof(float);
426 #else // _TARGET_ARM64_
427 if (lvHfaTypeIsFloat())
429 return lvExactSize / sizeof(float);
433 return lvExactSize / sizeof(double);
435 #endif // _TARGET_ARM64_
438 // lvIsMultiRegArgOrRet()
439 // returns true if this is a multireg LclVar struct used in an argument context
440 // or if this is a multireg LclVar struct assigned from a multireg call
441 bool lvIsMultiRegArgOrRet()
443 return lvIsMultiRegArg || lvIsMultiRegRet;
447 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
448 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
449 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
450 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
451 // to non-zero if the variable gets the same register assignment for its entire
453 #if !defined(_TARGET_64BIT_)
454 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
455 #endif // !defined(_TARGET_64BIT_)
457 regNumberSmall _lvArgReg; // The register in which this argument is passed.
459 #if FEATURE_MULTIREG_ARGS
460 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
461 // Note this is defined but not used by ARM32
462 #endif // FEATURE_MULTIREG_ARGS
464 #ifndef LEGACY_BACKEND
466 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
467 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
469 #endif // !LEGACY_BACKEND
472 // The register number is stored in a small format (8 bits), but the getters return and the setters take
473 // a full-size (unsigned) format, to localize the casts here.
475 /////////////////////
477 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
479 regNumber GetRegNum() const
481 return (regNumber)_lvRegNum;
484 void SetRegNum(regNumber reg)
486 _lvRegNum = (regNumberSmall)reg;
487 assert(_lvRegNum == reg);
490 /////////////////////
492 #if defined(_TARGET_64BIT_)
493 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
495 regNumber GetOtherReg() const
497 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
498 // "unreachable code" warnings
502 void SetOtherReg(regNumber reg)
504 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
505 // "unreachable code" warnings
507 #else // !_TARGET_64BIT_
508 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
510 regNumber GetOtherReg() const
512 return (regNumber)_lvOtherReg;
515 void SetOtherReg(regNumber reg)
517 _lvOtherReg = (regNumberSmall)reg;
518 assert(_lvOtherReg == reg);
520 #endif // !_TARGET_64BIT_
522 /////////////////////
524 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
526 regNumber GetArgReg() const
528 return (regNumber)_lvArgReg;
531 void SetArgReg(regNumber reg)
533 _lvArgReg = (regNumberSmall)reg;
534 assert(_lvArgReg == reg);
537 #if FEATURE_MULTIREG_ARGS
538 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
540 regNumber GetOtherArgReg() const
542 return (regNumber)_lvOtherArgReg;
545 void SetOtherArgReg(regNumber reg)
547 _lvOtherArgReg = (regNumberSmall)reg;
548 assert(_lvOtherArgReg == reg);
550 #endif // FEATURE_MULTIREG_ARGS
553 // Is this is a SIMD struct?
554 bool lvIsSIMDType() const
559 // Is this is a SIMD struct which is used for SIMD intrinsic?
560 bool lvIsUsedInSIMDIntrinsic() const
562 return lvUsedInSIMDIntrinsic;
565 // If feature_simd not enabled, return false
566 bool lvIsSIMDType() const
570 bool lvIsUsedInSIMDIntrinsic() const
576 /////////////////////
578 #ifndef LEGACY_BACKEND
579 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
581 regNumber GetArgInitReg() const
583 return (regNumber)_lvArgInitReg;
586 void SetArgInitReg(regNumber reg)
588 _lvArgInitReg = (regNumberSmall)reg;
589 assert(_lvArgInitReg == reg);
592 /////////////////////
594 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
596 regPairNo GetArgInitRegPair() const
598 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
599 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
603 void SetArgInitRegPair(regPairNo regPair)
605 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
606 _lvArgInitRegPair = (regPairNoSmall)regPair;
607 assert(_lvArgInitRegPair == regPair);
610 /////////////////////
612 bool lvIsRegCandidate() const
614 return lvLRACandidate != 0;
617 bool lvIsInReg() const
619 return lvIsRegCandidate() && (lvRegNum != REG_STK);
622 #else // LEGACY_BACKEND
624 bool lvIsRegCandidate() const
626 return lvTracked != 0;
629 bool lvIsInReg() const
631 return lvRegister != 0;
634 #endif // LEGACY_BACKEND
636 regMaskTP lvRegMask() const
638 regMaskTP regMask = RBM_NONE;
639 if (varTypeIsFloating(TypeGet()))
641 if (lvRegNum != REG_STK)
643 regMask = genRegMaskFloat(lvRegNum, TypeGet());
648 if (lvRegNum != REG_STK)
650 regMask = genRegMask(lvRegNum);
653 // For longs we may have two regs
654 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
656 regMask |= genRegMask(lvOtherReg);
662 regMaskSmall lvPrefReg; // set of regs it prefers to live in
664 unsigned short lvVarIndex; // variable tracking index
665 unsigned short lvRefCnt; // unweighted (real) reference count. For implicit by reference
666 // parameters, this gets hijacked from fgMarkImplicitByRefArgs
667 // through fgMarkDemotedImplicitByRefArgs, to provide a static
668 // appearance count (computed during address-exposed analysis)
669 // that fgMakeOutgoingStructArgCopy consults during global morph
670 // to determine if eliding its copy is legal.
671 unsigned lvRefCntWtd; // weighted reference count
672 int lvStkOffs; // stack offset of home
673 unsigned lvExactSize; // (exact) size of the type in bytes
675 // Is this a promoted struct?
676 // This method returns true only for structs (including SIMD structs), not for
677 // locals that are split on a 32-bit target.
678 // It is only necessary to use this:
679 // 1) if only structs are wanted, and
680 // 2) if Lowering has already been done.
681 // Otherwise lvPromoted is valid.
682 bool lvPromotedStruct()
684 #if !defined(_TARGET_64BIT_)
685 return (lvPromoted && !varTypeIsLong(lvType));
686 #else // defined(_TARGET_64BIT_)
688 #endif // defined(_TARGET_64BIT_)
691 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
693 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
694 // where the struct itself is no longer used because all access is via its member fields.
695 // When that happens, the struct is marked as unused and its type has been changed to
696 // TYP_INT (to keep the GC tracking code from looking at it).
697 // See Compiler::raAssignVars() for details. For example:
698 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
699 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
700 // float V03.f1 (offs=0x00) -> V12 tmp7
701 // f8 (last use) (last use) $345
702 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
703 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
705 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
707 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
708 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
709 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
710 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
711 if ((lvType == TYP_SIMD12) && !lvIsParam)
713 assert(lvExactSize == 12);
716 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
718 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
721 const size_t lvArgStackSize() const;
723 unsigned lvSlotNum; // original slot # (if remapped)
725 typeInfo lvVerTypeInfo; // type info needed for verification
727 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
729 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
731 BYTE* lvGcLayout; // GC layout info for structs
734 BlockSet lvRefBlks; // Set of blocks that contain refs
735 GenTree* lvDefStmt; // Pointer to the statement with the single definition
736 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
738 var_types TypeGet() const
740 return (var_types)lvType;
742 bool lvStackAligned() const
744 assert(lvIsStructField);
745 return ((lvFldOffset % TARGET_POINTER_SIZE) == 0);
747 bool lvNormalizeOnLoad() const
749 return varTypeIsSmall(TypeGet()) &&
750 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
751 (lvIsParam || lvAddrExposed || lvIsStructField);
754 bool lvNormalizeOnStore()
756 return varTypeIsSmall(TypeGet()) &&
757 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
758 !(lvIsParam || lvAddrExposed || lvIsStructField);
761 void lvaResetSortAgainFlag(Compiler* pComp);
762 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
763 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
764 void setPrefReg(regNumber regNum, Compiler* pComp);
765 void addPrefReg(regMaskTP regMask, Compiler* pComp);
766 bool IsFloatRegType() const
768 return isFloatRegType(lvType) || lvIsHfaRegArg();
770 var_types GetHfaType() const
772 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
774 void SetHfaType(var_types type)
776 assert(varTypeIsFloating(type));
777 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
780 #ifndef LEGACY_BACKEND
781 var_types lvaArgType();
784 PerSsaArray lvPerSsaData;
787 // Keep track of the # of SsaNames, for a bounds check.
788 unsigned lvNumSsaNames;
791 // Returns the address of the per-Ssa data for the given ssaNum (which is required
792 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
793 // not an SSA variable).
794 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
796 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
797 assert(SsaConfig::RESERVED_SSA_NUM == 0);
798 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
799 assert(zeroBased < lvNumSsaNames);
800 return &lvPerSsaData.GetRef(zeroBased);
805 void PrintVarReg() const
807 if (isRegPairType(TypeGet()))
809 printf("%s:%s", getRegName(lvOtherReg), // hi32
810 getRegName(lvRegNum)); // lo32
814 printf("%s", getRegName(lvRegNum));
819 }; // class LclVarDsc
822 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
823 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
827 XX The temporary lclVars allocated by the compiler for code generation XX
829 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
830 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
833 /*****************************************************************************
835 * The following keeps track of temporaries allocated in the stack frame
836 * during code-generation (after register allocation). These spill-temps are
837 * only used if we run out of registers while evaluating a tree.
839 * These are different from the more common temps allocated by lvaGrabTemp().
850 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
858 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
862 0); // temps must have a negative number (so they have a different number from all local variables)
863 tdOffs = BAD_TEMP_OFFSET;
867 IMPL_LIMITATION("too many spill temps");
872 bool tdLegalOffset() const
874 return tdOffs != BAD_TEMP_OFFSET;
878 int tdTempOffs() const
880 assert(tdLegalOffset());
883 void tdSetTempOffs(int offs)
886 assert(tdLegalOffset());
888 void tdAdjustTempOffs(int offs)
891 assert(tdLegalOffset());
894 int tdTempNum() const
899 unsigned tdTempSize() const
903 var_types tdTempType() const
909 // interface to hide linearscan implementation from rest of compiler
910 class LinearScanInterface
913 virtual void doLinearScan() = 0;
914 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
915 virtual bool willEnregisterLocalVars() const = 0;
918 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
920 // Information about arrays: their element type and size, and the offset of the first element.
921 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
922 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
923 // for example, in value numbering of array index expressions.
926 var_types m_elemType;
927 CORINFO_CLASS_HANDLE m_elemStructType;
929 unsigned m_elemOffset;
931 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
935 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
936 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
941 // This enumeration names the phases into which we divide compilation. The phases should completely
942 // partition a compilation.
945 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
946 #include "compphases.h"
950 extern const char* PhaseNames[];
951 extern const char* PhaseEnums[];
952 extern const LPCWSTR PhaseShortNames[];
954 // The following enum provides a simple 1:1 mapping to CLR API's
955 enum API_ICorJitInfo_Names
957 #define DEF_CLR_API(name) API_##name,
958 #include "ICorJitInfo_API_names.h"
962 //---------------------------------------------------------------
966 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
967 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
968 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
969 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
970 // by "m_timerFailure" being true.
971 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
974 #ifdef FEATURE_JIT_METHOD_PERF
975 // The string names of the phases.
976 static const char* PhaseNames[];
978 static bool PhaseHasChildren[];
979 static int PhaseParent[];
980 static bool PhaseReportsIRSize[];
982 unsigned m_byteCodeBytes;
983 unsigned __int64 m_totalCycles;
984 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
985 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
986 #if MEASURE_CLRAPI_CALLS
987 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
988 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
991 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
993 // For better documentation, we call EndPhase on
994 // non-leaf phases. We should also call EndPhase on the
995 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
996 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
997 // We add all such "redundant end phase" intervals to this variable below; we print
998 // it out in a report, so we can verify that it is, indeed, very small. If it ever
999 // isn't, this means that we're doing something significant between the end of the last
1000 // declared subphase and the end of its parent.
1001 unsigned __int64 m_parentPhaseEndSlop;
1002 bool m_timerFailure;
1004 #if MEASURE_CLRAPI_CALLS
1005 // The following measures the time spent inside each individual CLR API call.
1006 unsigned m_allClrAPIcalls;
1007 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1008 unsigned __int64 m_allClrAPIcycles;
1009 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1010 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1011 #endif // MEASURE_CLRAPI_CALLS
1013 CompTimeInfo(unsigned byteCodeBytes);
1017 #ifdef FEATURE_JIT_METHOD_PERF
1019 #if MEASURE_CLRAPI_CALLS
1020 struct WrapICorJitInfo;
1023 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1024 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1025 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1026 // threads, so it is protected by a lock.
1027 // This class is intended to be used as a singleton type, with only a single instance.
1028 class CompTimeSummaryInfo
1030 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1031 static CritSecObject s_compTimeSummaryLock;
1035 CompTimeInfo m_total;
1036 CompTimeInfo m_maximum;
1038 int m_numFilteredMethods;
1039 CompTimeInfo m_filtered;
1041 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1042 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1043 // If any OS interaction fails, returns 0.0.
1044 double CyclesPerSecond();
1046 // This can use what ever data you want to determine if the value to be added
1047 // belongs in the filtered section (it's always included in the unfiltered section)
1048 bool IncludedInFilteredData(CompTimeInfo& info);
1051 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1052 static CompTimeSummaryInfo s_compTimeSummary;
1054 CompTimeSummaryInfo()
1055 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1059 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1060 // This is thread safe.
1061 void AddInfo(CompTimeInfo& info, bool includePhases);
1063 // Print the summary information to "f".
1064 // This is not thread-safe; assumed to be called by only one thread.
1065 void Print(FILE* f);
1068 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1069 // and when the current phase started. This is intended to be part of a Compilation object. This is
1070 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1074 unsigned __int64 m_start; // Start of the compilation.
1075 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1076 #if MEASURE_CLRAPI_CALLS
1077 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1078 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1079 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1080 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1081 static double s_cyclesPerSec; // Cached for speedier measurements
1084 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1086 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1088 static CritSecObject s_csvLock; // Lock to protect the time log file.
1089 void PrintCsvMethodStats(Compiler* comp);
1092 void* operator new(size_t);
1093 void* operator new[](size_t);
1094 void operator delete(void*);
1095 void operator delete[](void*);
1098 // Initialized the timer instance
1099 JitTimer(unsigned byteCodeSize);
1101 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1103 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1106 static void PrintCsvHeader();
1108 // Ends the current phase (argument is for a redundant check).
1109 void EndPhase(Compiler* compiler, Phases phase);
1111 #if MEASURE_CLRAPI_CALLS
1112 // Start and end a timed CLR API call.
1113 void CLRApiCallEnter(unsigned apix);
1114 void CLRApiCallLeave(unsigned apix);
1115 #endif // MEASURE_CLRAPI_CALLS
1117 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1118 // and adds it to "sum".
1119 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1121 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1122 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1123 // "m_info" to true.
1124 bool GetThreadCycles(unsigned __int64* cycles)
1126 bool res = CycleTimer::GetThreadCyclesS(cycles);
1129 m_info.m_timerFailure = true;
1134 #endif // FEATURE_JIT_METHOD_PERF
1136 //------------------- Function/Funclet info -------------------------------
1137 enum FuncKind : BYTE
1139 FUNC_ROOT, // The main/root function (always id==0)
1140 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1141 FUNC_FILTER, // a funclet associated with an EH filter
1150 BYTE funFlags; // Currently unused, just here for padding
1151 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1152 // funclet. It is only valid if funKind field indicates this is a
1153 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1155 #if defined(_TARGET_AMD64_)
1157 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1158 emitLocation* startLoc;
1159 emitLocation* endLoc;
1160 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1161 emitLocation* coldEndLoc;
1162 UNWIND_INFO unwindHeader;
1163 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1164 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1165 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1166 unsigned unwindCodeSlot;
1168 #elif defined(_TARGET_X86_)
1170 #if defined(_TARGET_UNIX_)
1171 emitLocation* startLoc;
1172 emitLocation* endLoc;
1173 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1174 emitLocation* coldEndLoc;
1175 #endif // _TARGET_UNIX_
1177 #elif defined(_TARGET_ARMARCH_)
1179 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1180 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1181 // Note: we only have a pointer here instead of the actual object,
1182 // to save memory in the JIT case (compared to the NGEN case),
1183 // where we don't have any cold section.
1184 // Note 2: we currently don't support hot/cold splitting in functions
1185 // with EH, so uwiCold will be NULL for all funclets.
1187 #if defined(_TARGET_UNIX_)
1188 emitLocation* startLoc;
1189 emitLocation* endLoc;
1190 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1191 emitLocation* coldEndLoc;
1192 #endif // _TARGET_UNIX_
1194 #endif // _TARGET_ARMARCH_
1196 #if defined(_TARGET_UNIX_)
1197 jitstd::vector<CFI_CODE>* cfiCodes;
1198 #endif // _TARGET_UNIX_
1200 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1201 // that isn't shared between the main function body and funclets.
1204 struct fgArgTabEntry
1207 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1210 otherRegNum = REG_NA;
1211 isStruct = false; // is this a struct arg
1213 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1215 GenTree* node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1217 // it will point at the actual argument in the gtCallLateArgs list.
1218 GenTree* parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1220 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1222 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1224 unsigned numRegs; // Count of number of registers that this argument uses
1226 // A slot is a pointer sized region in the OutArg area.
1227 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1228 unsigned numSlots; // Count of number of slots that this argument uses
1230 unsigned alignment; // 1 or 2 (slots/registers)
1231 unsigned lateArgInx; // index into gtCallLateArgs list
1232 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1234 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1235 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1236 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1237 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1238 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1239 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1240 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1241 // previous arguments.
1242 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1243 // to be on the stack despite its arg list position.
1245 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1246 bool isStruct : 1; // True if this is a struct arg
1248 regNumber otherRegNum; // The (second) register to use when passing this argument.
1250 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1251 #elif !defined(_TARGET_64BIT_)
1252 __declspec(property(get = getIsStruct)) bool isStruct;
1255 return varTypeIsStruct(node);
1257 #endif // !_TARGET_64BIT_
1260 void SetIsHfaRegArg(bool hfaRegArg)
1262 isHfaRegArg = hfaRegArg;
1265 void SetIsBackFilled(bool backFilled)
1267 isBackFilled = backFilled;
1270 bool IsBackFilled() const
1272 return isBackFilled;
1274 #else // !_TARGET_ARM_
1275 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1277 void SetIsHfaRegArg(bool hfaRegArg)
1281 void SetIsBackFilled(bool backFilled)
1285 bool IsBackFilled() const
1289 #endif // !_TARGET_ARM_
1296 //-------------------------------------------------------------------------
1298 // The class fgArgInfo is used to handle the arguments
1299 // when morphing a GT_CALL node.
1304 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1305 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1306 unsigned argCount; // Updatable arg count value
1307 unsigned nextSlotNum; // Updatable slot count value
1308 unsigned stkLevel; // Stack depth when we make this call (for x86)
1310 #if defined(UNIX_X86_ABI)
1311 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1312 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1313 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1314 // Computed dynamically during codegen, based on stkSizeBytes and the current
1315 // stack level (genStackLevel) when the first stack adjustment is made for
1319 #if FEATURE_FIXED_OUT_ARGS
1320 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1323 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1324 bool hasRegArgs; // true if we have one or more register arguments
1325 bool hasStackArgs; // true if we have one or more stack arguments
1326 bool argsComplete; // marker for state
1327 bool argsSorted; // marker for state
1328 fgArgTabEntry** argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1331 void AddArg(fgArgTabEntry* curArgTabEntry);
1334 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1335 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1337 fgArgTabEntry* AddRegArg(
1338 unsigned argNum, GenTree* node, GenTree* parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1340 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1341 fgArgTabEntry* AddRegArg(unsigned argNum,
1347 const bool isStruct,
1348 const regNumber otherRegNum = REG_NA,
1349 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1350 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1352 fgArgTabEntry* AddStkArg(unsigned argNum,
1356 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1358 void RemorphReset();
1359 fgArgTabEntry* RemorphRegArg(
1360 unsigned argNum, GenTree* node, GenTree* parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1362 void RemorphStkArg(unsigned argNum, GenTree* node, GenTree* parent, unsigned numSlots, unsigned alignment);
1364 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1366 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTree* newNode);
1368 void ArgsComplete();
1372 void EvalArgsToTemps();
1374 void RecordStkLevel(unsigned stkLvl);
1375 unsigned RetrieveStkLevel();
1381 fgArgTabEntry** ArgTable()
1385 unsigned GetNextSlotNum()
1395 return hasStackArgs;
1397 bool AreArgsComplete() const
1399 return argsComplete;
1401 #if FEATURE_FIXED_OUT_ARGS
1402 unsigned GetOutArgSize() const
1406 void SetOutArgSize(unsigned newVal)
1408 outArgSize = newVal;
1410 #endif // FEATURE_FIXED_OUT_ARGS
1412 #if defined(UNIX_X86_ABI)
1413 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1415 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1418 unsigned GetStkAlign()
1423 void SetStkSizeBytes(unsigned newStkSizeBytes)
1425 stkSizeBytes = newStkSizeBytes;
1428 unsigned GetStkSizeBytes() const
1430 return stkSizeBytes;
1433 bool IsStkAlignmentDone() const
1435 return alignmentDone;
1438 void SetStkAlignmentDone()
1440 alignmentDone = true;
1442 #endif // defined(UNIX_X86_ABI)
1444 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1445 GenTree* GetLateArg(unsigned argIndex);
1447 void Dump(Compiler* compiler);
1451 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1452 // We have the ability to mark source expressions with "Test Labels."
1453 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1454 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1456 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1459 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1460 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1461 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1462 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1463 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1466 struct TestLabelAndNum
1471 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1476 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, TestLabelAndNum> NodeToTestDataMap;
1478 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1482 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1483 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1485 XX The big guy. The sections are currently organized as : XX
1487 XX o GenTree and BasicBlock XX
1499 XX o PrologScopeInfo XX
1500 XX o CodeGenerator XX
1505 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1506 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1509 struct HWIntrinsicInfo;
1513 friend class emitter;
1514 friend class UnwindInfo;
1515 friend class UnwindFragmentInfo;
1516 friend class UnwindEpilogInfo;
1517 friend class JitTimer;
1518 friend class LinearScan;
1519 friend class fgArgInfo;
1520 friend class Rationalizer;
1522 friend class Lowering;
1523 friend class CSE_DataFlow;
1524 friend class CSE_Heuristic;
1525 friend class CodeGenInterface;
1526 friend class CodeGen;
1527 friend class LclVarDsc;
1528 friend class TempDsc;
1530 friend class ObjectAllocator;
1531 friend struct GenTree;
1533 #ifndef _TARGET_64BIT_
1534 friend class DecomposeLongs;
1535 #endif // !_TARGET_64BIT_
1538 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1539 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1541 XX Misc structs definitions XX
1543 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1544 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1548 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1567 bool dumpIRDataflow;
1568 bool dumpIRBlockHeaders;
1570 LPCWSTR dumpIRPhase;
1571 LPCWSTR dumpIRFormat;
1573 bool shouldUseVerboseTrees();
1574 bool asciiTrees; // If true, dump trees using only ASCII characters
1575 bool shouldDumpASCIITrees();
1576 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1577 bool shouldUseVerboseSsa();
1578 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1579 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1581 const char* VarNameToStr(VarName name)
1586 DWORD expensiveDebugCheckLevel;
1589 #if FEATURE_MULTIREG_RET
1590 GenTree* impAssignMultiRegTypeToVar(GenTree* op, CORINFO_CLASS_HANDLE hClass);
1591 #endif // FEATURE_MULTIREG_RET
1594 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1595 #endif // ARM_SOFTFP
1597 //-------------------------------------------------------------------------
1598 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1599 // HFAs are one to four element structs where each element is the same
1600 // type, either all float or all double. They are treated specially
1601 // in the ARM Procedure Call Standard, specifically, they are passed in
1602 // floating-point registers instead of the general purpose registers.
1605 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1606 bool IsHfa(GenTree* tree);
1608 var_types GetHfaType(GenTree* tree);
1609 unsigned GetHfaCount(GenTree* tree);
1611 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1612 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1614 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1615 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1617 //-------------------------------------------------------------------------
1618 // The following is used for validating format of EH table
1622 typedef struct EHNodeDsc* pEHNodeDsc;
1624 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1625 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1638 EHBlockType ehnBlockType; // kind of EH block
1639 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1640 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1641 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1643 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1644 pEHNodeDsc ehnChild; // leftmost nested block
1646 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1647 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1649 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1650 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1652 inline void ehnSetTryNodeType()
1654 ehnBlockType = TryNode;
1656 inline void ehnSetFilterNodeType()
1658 ehnBlockType = FilterNode;
1660 inline void ehnSetHandlerNodeType()
1662 ehnBlockType = HandlerNode;
1664 inline void ehnSetFinallyNodeType()
1666 ehnBlockType = FinallyNode;
1668 inline void ehnSetFaultNodeType()
1670 ehnBlockType = FaultNode;
1673 inline BOOL ehnIsTryBlock()
1675 return ehnBlockType == TryNode;
1677 inline BOOL ehnIsFilterBlock()
1679 return ehnBlockType == FilterNode;
1681 inline BOOL ehnIsHandlerBlock()
1683 return ehnBlockType == HandlerNode;
1685 inline BOOL ehnIsFinallyBlock()
1687 return ehnBlockType == FinallyNode;
1689 inline BOOL ehnIsFaultBlock()
1691 return ehnBlockType == FaultNode;
1694 // returns true if there is any overlap between the two nodes
1695 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1697 if (node1->ehnStartOffset < node2->ehnStartOffset)
1699 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1703 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1707 // fails with BADCODE if inner is not completely nested inside outer
1708 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1710 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1714 //-------------------------------------------------------------------------
1715 // Exception handling functions
1718 #if !FEATURE_EH_FUNCLETS
1720 bool ehNeedsShadowSPslots()
1722 return (info.compXcptnsCount || opts.compDbgEnC);
1725 // 0 for methods with no EH
1726 // 1 for methods with non-nested EH, or where only the try blocks are nested
1727 // 2 for a method with a catch within a catch
1729 unsigned ehMaxHndNestingCount;
1731 #endif // !FEATURE_EH_FUNCLETS
1733 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1734 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1736 bool bbInCatchHandlerILRange(BasicBlock* blk);
1737 bool bbInFilterILRange(BasicBlock* blk);
1738 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1739 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1740 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1741 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1742 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1744 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1745 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1747 // Returns true if "block" is the start of a try region.
1748 bool bbIsTryBeg(BasicBlock* block);
1750 // Returns true if "block" is the start of a handler or filter region.
1751 bool bbIsHandlerBeg(BasicBlock* block);
1753 // Returns true iff "block" is where control flows if an exception is raised in the
1754 // try region, and sets "*regionIndex" to the index of the try for the handler.
1755 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1756 // block of the filter, but not for the filter's handler.
1757 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1759 bool ehHasCallableHandlers();
1761 // Return the EH descriptor for the given region index.
1762 EHblkDsc* ehGetDsc(unsigned regionIndex);
1764 // Return the EH index given a region descriptor.
1765 unsigned ehGetIndex(EHblkDsc* ehDsc);
1767 // Return the EH descriptor index of the enclosing try, for the given region index.
1768 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1770 // Return the EH descriptor index of the enclosing handler, for the given region index.
1771 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1773 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1774 // block is not in a 'try' region).
1775 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1777 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1778 // if this block is not in a filter or handler region).
1779 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1781 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1782 // nullptr if this block's exceptions propagate to caller).
1783 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1785 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1786 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1787 bool ehIsBlockEHLast(BasicBlock* block);
1789 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1791 // Return the region index of the most nested EH region this block is in.
1792 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1794 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1795 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1797 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1798 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1799 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1800 // (It can never be a filter.)
1801 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1803 // A block has been deleted. Update the EH table appropriately.
1804 void ehUpdateForDeletedBlock(BasicBlock* block);
1806 // Determine whether a block can be deleted while preserving the EH normalization rules.
1807 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1809 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1810 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1812 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1813 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1814 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1815 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1816 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1817 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1818 // lives in a filter.)
1819 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1821 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1822 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1823 // (nullptr if the last block is the last block in the program).
1824 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1825 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1828 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1829 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1830 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1833 #if FEATURE_EH_FUNCLETS
1834 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1835 // if there is a filter that protects a region with a nested EH clause (such as a
1836 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1837 // genFuncletProlog() for more details. However, the VM seems to use it for more
1838 // purposes, maybe including debugging. Until we are sure otherwise, always create
1839 // a PSPSym for functions with any EH.
1840 bool ehNeedsPSPSym() const
1844 #else // _TARGET_X86_
1845 return compHndBBtabCount > 0;
1846 #endif // _TARGET_X86_
1849 bool ehAnyFunclets(); // Are there any funclets in this function?
1850 unsigned ehFuncletCount(); // Return the count of funclets in the function
1852 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1853 #else // !FEATURE_EH_FUNCLETS
1854 bool ehAnyFunclets()
1858 unsigned ehFuncletCount()
1863 unsigned bbThrowIndex(BasicBlock* blk)
1865 return blk->bbTryIndex;
1866 } // Get the index to use as the cache key for sharing throw blocks
1867 #endif // !FEATURE_EH_FUNCLETS
1869 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1870 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1871 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1872 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1873 // convenient to also consider it a predecessor.)
1874 flowList* BlockPredsWithEH(BasicBlock* blk);
1876 // This table is useful for memoization of the method above.
1877 typedef JitHashTable<BasicBlock*, JitPtrKeyFuncs<BasicBlock>, flowList*> BlockToFlowListMap;
1878 BlockToFlowListMap* m_blockToEHPreds;
1879 BlockToFlowListMap* GetBlockToEHPreds()
1881 if (m_blockToEHPreds == nullptr)
1883 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1885 return m_blockToEHPreds;
1888 void* ehEmitCookie(BasicBlock* block);
1889 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1891 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1893 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1895 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1897 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1899 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1901 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1903 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1905 void fgAllocEHTable();
1907 void fgRemoveEHTableEntry(unsigned XTnum);
1909 #if FEATURE_EH_FUNCLETS
1911 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1913 #endif // FEATURE_EH_FUNCLETS
1917 #endif // !FEATURE_EH
1919 void fgSortEHTable();
1921 // Causes the EH table to obey some well-formedness conditions, by inserting
1922 // empty BB's when necessary:
1923 // * No block is both the first block of a handler and the first block of a try.
1924 // * No block is the first block of multiple 'try' regions.
1925 // * No block is the last block of multiple EH regions.
1926 void fgNormalizeEH();
1927 bool fgNormalizeEHCase1();
1928 bool fgNormalizeEHCase2();
1929 bool fgNormalizeEHCase3();
1932 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1933 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1934 void fgVerifyHandlerTab();
1935 void fgDispHandlerTab();
1938 bool fgNeedToSortEHTable;
1940 void verInitEHTree(unsigned numEHClauses);
1941 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1942 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1943 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1944 void verCheckNestingLevel(EHNodeDsc* initRoot);
1947 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1948 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1950 XX GenTree and BasicBlock XX
1952 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1954 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1955 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1958 // Functions to create nodes
1959 GenTreeStmt* gtNewStmt(GenTree* expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1962 GenTree* gtNewOperNode(genTreeOps oper, var_types type, GenTree* op1, bool doSimplifications = TRUE);
1964 // For binary opers.
1965 GenTree* gtNewOperNode(genTreeOps oper, var_types type, GenTree* op1, GenTree* op2);
1967 GenTree* gtNewQmarkNode(var_types type, GenTree* cond, GenTree* colon);
1969 GenTree* gtNewLargeOperNode(genTreeOps oper,
1970 var_types type = TYP_I_IMPL,
1971 GenTree* op1 = nullptr,
1972 GenTree* op2 = nullptr);
1974 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1976 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1978 GenTree* gtNewJmpTableNode();
1980 GenTree* gtNewIndOfIconHandleNode(var_types indType, size_t value, unsigned iconFlags, bool isInvariant);
1982 GenTree* gtNewIconHandleNode(size_t value, unsigned flags, FieldSeqNode* fields = nullptr);
1984 unsigned gtTokenToIconFlags(unsigned token);
1986 GenTree* gtNewIconEmbHndNode(void* value, void* pValue, unsigned flags, void* compileTimeHandle);
1988 GenTree* gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd);
1989 GenTree* gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd);
1990 GenTree* gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd);
1991 GenTree* gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd);
1993 GenTree* gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1995 GenTree* gtNewLconNode(__int64 value);
1997 GenTree* gtNewDconNode(double value);
1999 GenTree* gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2001 GenTree* gtNewZeroConNode(var_types type);
2003 GenTree* gtNewOneConNode(var_types type);
2006 GenTree* gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2007 GenTree* gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2010 GenTreeBlk* gtNewBlkOpNode(
2011 genTreeOps oper, GenTree* dst, GenTree* srcOrFillVal, GenTree* sizeOrClsTok, bool isVolatile);
2013 GenTree* gtNewBlkOpNode(GenTree* dst, GenTree* srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2015 GenTree* gtNewPutArgReg(var_types type, GenTree* arg, regNumber argReg);
2017 GenTree* gtNewBitCastNode(var_types type, GenTree* arg);
2020 void gtBlockOpInit(GenTree* result, GenTree* dst, GenTree* srcOrFillVal, bool isVolatile);
2023 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTree* addr);
2024 void gtSetObjGcInfo(GenTreeObj* objNode);
2025 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTree* addr);
2026 GenTree* gtNewBlockVal(GenTree* addr, unsigned size);
2028 GenTree* gtNewCpObjNode(GenTree* dst, GenTree* src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2030 GenTreeArgList* gtNewListNode(GenTree* op1, GenTreeArgList* op2);
2032 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2033 CORINFO_METHOD_HANDLE handle,
2035 GenTreeArgList* args,
2036 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2038 GenTreeCall* gtNewIndCallNode(GenTree* addr,
2040 GenTreeArgList* args,
2041 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2043 GenTreeCall* gtNewHelperCallNode(unsigned helper, var_types type, GenTreeArgList* args = nullptr);
2045 GenTree* gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2048 GenTreeSIMD* gtNewSIMDNode(
2049 var_types type, GenTree* op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2050 GenTreeSIMD* gtNewSIMDNode(
2051 var_types type, GenTree* op1, GenTree* op2, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2052 void SetOpLclRelatedToSIMDIntrinsic(GenTree* op);
2055 #ifdef FEATURE_HW_INTRINSICS
2056 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2057 NamedIntrinsic hwIntrinsicID,
2060 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(
2061 var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID, var_types baseType, unsigned size);
2062 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(
2063 var_types type, GenTree* op1, GenTree* op2, NamedIntrinsic hwIntrinsicID, var_types baseType, unsigned size);
2064 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2068 NamedIntrinsic hwIntrinsicID,
2071 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2076 NamedIntrinsic hwIntrinsicID,
2079 GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID);
2080 GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type,
2083 NamedIntrinsic hwIntrinsicID);
2084 GenTree* gtNewMustThrowException(unsigned helper, var_types type, CORINFO_CLASS_HANDLE clsHnd);
2085 CORINFO_CLASS_HANDLE gtGetStructHandleForHWSIMD(var_types simdType, var_types simdBaseType);
2086 #endif // FEATURE_HW_INTRINSICS
2088 GenTree* gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2089 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2090 GenTree* gtNewInlineCandidateReturnExpr(GenTree* inlineCandidate, var_types type);
2092 GenTree* gtNewCodeRef(BasicBlock* block);
2094 GenTree* gtNewFieldRef(
2095 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTree* obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2097 GenTree* gtNewIndexRef(var_types typ, GenTree* arrayOp, GenTree* indexOp);
2099 GenTreeArrLen* gtNewArrLen(var_types typ, GenTree* arrayOp, int lenOffset);
2101 GenTree* gtNewIndir(var_types typ, GenTree* addr);
2103 GenTreeArgList* gtNewArgList(GenTree* op);
2104 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2);
2105 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2, GenTree* op3);
2106 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2, GenTree* op3, GenTree* op4);
2108 static fgArgTabEntry* gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2109 static fgArgTabEntry* gtArgEntryByNode(GenTreeCall* call, GenTree* node);
2110 fgArgTabEntry* gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2111 bool gtArgIsThisPtr(fgArgTabEntry* argEntry);
2113 GenTree* gtNewAssignNode(GenTree* dst, GenTree* src);
2115 GenTree* gtNewTempAssign(unsigned tmp, GenTree* val);
2117 GenTree* gtNewRefCOMfield(GenTree* objPtr,
2118 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2119 CORINFO_ACCESS_FLAGS access,
2120 CORINFO_FIELD_INFO* pFieldInfo,
2122 CORINFO_CLASS_HANDLE structType,
2125 GenTree* gtNewNothingNode();
2127 GenTree* gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2129 GenTree* gtUnusedValNode(GenTree* expr);
2131 GenTreeCast* gtNewCastNode(var_types typ, GenTree* op1, bool fromUnsigned, var_types castType);
2133 GenTreeCast* gtNewCastNodeL(var_types typ, GenTree* op1, bool fromUnsigned, var_types castType);
2135 GenTree* gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTree* op1);
2137 GenTree* gtNewRuntimeLookup(CORINFO_GENERIC_HANDLE hnd, CorInfoGenericHandleType hndTyp, GenTree* lookupTree);
2139 //------------------------------------------------------------------------
2140 // Other GenTree functions
2142 GenTree* gtClone(GenTree* tree, bool complexOK = false);
2144 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2145 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2146 // IntCnses with value `deepVarVal`.
2147 GenTree* gtCloneExpr(
2148 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2150 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2151 // `varNum` to int constants with value `varVal`.
2152 GenTree* gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2154 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2157 GenTree* gtReplaceTree(GenTree* stmt, GenTree* tree, GenTree* replacementTree);
2159 void gtUpdateSideEffects(GenTree* stmt, GenTree* tree);
2161 void gtUpdateTreeAncestorsSideEffects(GenTree* tree);
2163 void gtUpdateStmtSideEffects(GenTree* stmt);
2165 void gtUpdateNodeSideEffects(GenTree* tree);
2167 void gtUpdateNodeOperSideEffects(GenTree* tree);
2169 // Returns "true" iff the complexity (not formally defined, but first interpretation
2170 // is #of nodes in subtree) of "tree" is greater than "limit".
2171 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2172 // before they have been set.)
2173 bool gtComplexityExceeds(GenTree** tree, unsigned limit);
2175 bool gtCompareTree(GenTree* op1, GenTree* op2);
2177 GenTree* gtReverseCond(GenTree* tree);
2179 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2181 bool gtHasLocalsWithAddrOp(GenTree* tree);
2183 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2185 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* base, bool constOnly);
2188 unsigned gtHashValue(GenTree* tree);
2190 GenTree* gtWalkOpEffectiveVal(GenTree* op);
2193 void gtPrepareCost(GenTree* tree);
2194 bool gtIsLikelyRegVar(GenTree* tree);
2196 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2198 // Returns true iff the secondNode can be swapped with firstNode.
2199 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2201 unsigned gtSetEvalOrder(GenTree* tree);
2203 #if FEATURE_STACK_FP_X87
2205 void gtComputeFPlvls(GenTree* tree);
2206 #endif // FEATURE_STACK_FP_X87
2208 void gtSetStmtInfo(GenTree* stmt);
2210 // Returns "true" iff "node" has any of the side effects in "flags".
2211 bool gtNodeHasSideEffects(GenTree* node, unsigned flags);
2213 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2214 bool gtTreeHasSideEffects(GenTree* tree, unsigned flags);
2216 // Appends 'expr' in front of 'list'
2217 // 'list' will typically start off as 'nullptr'
2218 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2219 GenTree* gtBuildCommaList(GenTree* list, GenTree* expr);
2221 void gtExtractSideEffList(GenTree* expr,
2223 unsigned flags = GTF_SIDE_EFFECT,
2224 bool ignoreRoot = false);
2226 GenTree* gtGetThisArg(GenTreeCall* call);
2228 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2229 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2230 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2231 // the given "fldHnd", is such an object pointer.
2232 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2234 // Return true if call is a recursive call; return false otherwise.
2235 // Note when inlining, this looks for calls back to the root method.
2236 bool gtIsRecursiveCall(GenTreeCall* call)
2238 return gtIsRecursiveCall(call->gtCallMethHnd);
2241 bool gtIsRecursiveCall(CORINFO_METHOD_HANDLE callMethodHandle)
2243 return (callMethodHandle == impInlineRoot()->info.compMethodHnd);
2246 //-------------------------------------------------------------------------
2248 GenTree* gtFoldExpr(GenTree* tree);
2251 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2252 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2253 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2254 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2255 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2256 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2257 // optimizations for now.
2258 __attribute__((optnone))
2260 gtFoldExprConst(GenTree* tree);
2261 GenTree* gtFoldExprSpecial(GenTree* tree);
2262 GenTree* gtFoldExprCompare(GenTree* tree);
2263 GenTree* gtFoldExprCall(GenTreeCall* call);
2264 GenTree* gtFoldTypeCompare(GenTree* tree);
2265 GenTree* gtFoldTypeEqualityCall(CorInfoIntrinsics methodID, GenTree* op1, GenTree* op2);
2267 // Options to control behavior of gtTryRemoveBoxUpstreamEffects
2268 enum BoxRemovalOptions
2270 BR_REMOVE_AND_NARROW, // remove effects, minimize remaining work, return possibly narrowed source tree
2271 BR_REMOVE_AND_NARROW_WANT_TYPE_HANDLE, // remove effects and minimize remaining work, return type handle tree
2272 BR_REMOVE_BUT_NOT_NARROW, // remove effects, return original source tree
2273 BR_DONT_REMOVE, // check if removal is possible, return copy source tree
2274 BR_DONT_REMOVE_WANT_TYPE_HANDLE, // check if removal is possible, return type handle tree
2275 BR_MAKE_LOCAL_COPY // revise box to copy to temp local and return local's address
2278 GenTree* gtTryRemoveBoxUpstreamEffects(GenTree* tree, BoxRemovalOptions options = BR_REMOVE_AND_NARROW);
2279 GenTree* gtOptimizeEnumHasFlag(GenTree* thisOp, GenTree* flagOp);
2281 //-------------------------------------------------------------------------
2282 // Get the handle, if any.
2283 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTree* tree);
2284 // Get the handle, and assert if not found.
2285 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTree* tree);
2286 // Get the handle for a ref type.
2287 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTree* tree, bool* isExact, bool* isNonNull);
2289 //-------------------------------------------------------------------------
2290 // Functions to display the trees
2293 void gtDispNode(GenTree* tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2295 void gtDispVN(GenTree* tree);
2296 void gtDispConst(GenTree* tree);
2297 void gtDispLeaf(GenTree* tree, IndentStack* indentStack);
2298 void gtDispNodeName(GenTree* tree);
2299 void gtDispRegVal(GenTree* tree);
2311 void gtDispChild(GenTree* child,
2312 IndentStack* indentStack,
2314 __in_opt const char* msg = nullptr,
2315 bool topOnly = false);
2316 void gtDispTree(GenTree* tree,
2317 IndentStack* indentStack = nullptr,
2318 __in_opt const char* msg = nullptr,
2319 bool topOnly = false,
2320 bool isLIR = false);
2321 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2322 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2323 char* gtGetLclVarName(unsigned lclNum);
2324 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2325 void gtDispTreeList(GenTree* tree, IndentStack* indentStack = nullptr);
2326 void gtGetArgMsg(GenTreeCall* call, GenTree* arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2327 void gtGetLateArgMsg(GenTreeCall* call, GenTree* arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2328 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2329 void gtDispFieldSeq(FieldSeqNode* pfsn);
2331 void gtDispRange(LIR::ReadOnlyRange const& range);
2333 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2335 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2347 typedef fgWalkResult(fgWalkPreFn)(GenTree** pTree, fgWalkData* data);
2348 typedef fgWalkResult(fgWalkPostFn)(GenTree** pTree, fgWalkData* data);
2351 static fgWalkPreFn gtAssertColonCond;
2353 static fgWalkPreFn gtMarkColonCond;
2354 static fgWalkPreFn gtClearColonCond;
2356 GenTree** gtFindLink(GenTree* stmt, GenTree* node);
2357 bool gtHasCatchArg(GenTree* tree);
2358 bool gtHasUnmanagedCall(GenTree* tree);
2360 typedef ArrayStack<GenTree*> GenTreeStack;
2362 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2363 void gtCheckQuirkAddrExposedLclVar(GenTree* argTree, GenTreeStack* parentStack);
2365 //=========================================================================
2366 // BasicBlock functions
2368 // This is a debug flag we will use to assert when creating block during codegen
2369 // as this interferes with procedure splitting. If you know what you're doing, set
2370 // it to true before creating the block. (DEBUG only)
2371 bool fgSafeBasicBlockCreation;
2374 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2377 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2378 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2382 XX The variables to be used by the code generator. XX
2384 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2385 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2389 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2390 // be placed in the stack frame and it's fields must be laid out sequentially.
2392 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2393 // a local variable that can be enregistered or placed in the stack frame.
2394 // The fields do not need to be laid out sequentially
2396 enum lvaPromotionType
2398 PROMOTION_TYPE_NONE, // The struct local is not promoted
2399 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2400 // and its field locals are independent of its parent struct local.
2401 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2402 // but its field locals depend on its parent struct local.
2405 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2406 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2408 /*****************************************************************************/
2410 enum FrameLayoutState
2413 INITIAL_FRAME_LAYOUT,
2414 PRE_REGALLOC_FRAME_LAYOUT,
2415 REGALLOC_FRAME_LAYOUT,
2416 TENTATIVE_FRAME_LAYOUT,
2421 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2422 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2423 bool lvaSortAgain; // true: We need to sort the lvaTable
2424 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2425 unsigned lvaCount; // total number of locals
2427 unsigned lvaRefCount; // total number of references to locals
2428 LclVarDsc* lvaTable; // variable descriptor table
2429 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2431 LclVarDsc** lvaRefSorted; // table sorted by refcount
2433 unsigned short lvaTrackedCount; // actual # of locals being tracked
2434 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2436 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2437 // Only for AMD64 System V cache the first caller stack homed argument.
2438 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2439 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2442 VARSET_TP lvaTrackedVars; // set of tracked variables
2444 #ifndef _TARGET_64BIT_
2445 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2447 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2449 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2450 // It that changes, this changes. VarSets from different epochs
2451 // cannot be meaningfully combined.
2453 unsigned GetCurLVEpoch()
2458 // reverse map of tracked number to var number
2459 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2461 #ifdef LEGACY_BACKEND
2462 // variable interference graph
2463 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2465 // variable preference graph
2466 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2470 // # of procs compiled a with double-aligned stack
2471 static unsigned s_lvaDoubleAlignedProcsCount;
2475 // Getters and setters for address-exposed and do-not-enregister local var properties.
2476 bool lvaVarAddrExposed(unsigned varNum);
2477 void lvaSetVarAddrExposed(unsigned varNum);
2478 bool lvaVarDoNotEnregister(unsigned varNum);
2480 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2481 enum DoNotEnregisterReason
2486 DNER_VMNeedsStackAddr,
2487 DNER_LiveInOutOfHandler,
2488 DNER_LiveAcrossUnmanagedCall,
2489 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2490 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2491 DNER_DepField, // It is a field of a dependently promoted struct
2492 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2493 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2494 #if !defined(LEGACY_BACKEND) && !defined(_TARGET_64BIT_)
2495 DNER_LongParamField, // It is a decomposed field of a long parameter.
2497 #ifdef JIT32_GCENCODER
2502 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2504 unsigned lvaVarargsHandleArg;
2506 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2508 #endif // _TARGET_X86_
2510 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2511 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2512 #if FEATURE_FIXED_OUT_ARGS
2513 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2515 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2516 // that tracks whether the lock has been taken
2518 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2519 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2520 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2522 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2523 // in case there are multiple BBJ_RETURN blocks in the inlinee
2524 // or if the inlinee has GC ref locals.
2526 #if FEATURE_FIXED_OUT_ARGS
2527 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2528 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2529 #endif // FEATURE_FIXED_OUT_ARGS
2532 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2533 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2534 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2535 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2536 // this variable to be this scratch word whenever struct promotion occurs.
2537 unsigned lvaPromotedStructAssemblyScratchVar;
2538 #endif // _TARGET_ARM_
2541 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2542 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2545 unsigned lvaGenericsContextUseCount;
2547 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2548 // CORINFO_GENERICS_CTXT_FROM_THIS?
2549 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2551 //-------------------------------------------------------------------------
2552 // All these frame offsets are inter-related and must be kept in sync
2554 #if !FEATURE_EH_FUNCLETS
2555 // This is used for the callable handlers
2556 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2557 #endif // FEATURE_EH_FUNCLETS
2559 int lvaCachedGenericContextArgOffs;
2560 int lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2563 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2565 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2567 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2568 // after the reg predict we will use a computed maxTmpSize
2569 // which is based upon the number of spill temps predicted by reg predict
2570 // All this is necessary because if we under-estimate the size of the spill
2571 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2573 // Pre codegen max spill temp size.
2574 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2576 //-------------------------------------------------------------------------
2578 unsigned lvaGetMaxSpillTempSize();
2580 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2581 #endif // _TARGET_ARM_
2582 void lvaAssignFrameOffsets(FrameLayoutState curState);
2583 void lvaFixVirtualFrameOffsets();
2585 #ifndef LEGACY_BACKEND
2586 void lvaUpdateArgsWithInitialReg();
2587 #endif // !LEGACY_BACKEND
2589 void lvaAssignVirtualFrameOffsetsToArgs();
2590 #ifdef UNIX_AMD64_ABI
2591 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2592 #else // !UNIX_AMD64_ABI
2593 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2594 #endif // !UNIX_AMD64_ABI
2595 void lvaAssignVirtualFrameOffsetsToLocals();
2596 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2597 #ifdef _TARGET_AMD64_
2598 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2599 bool lvaIsCalleeSavedIntRegCountEven();
2601 void lvaAlignFrame();
2602 void lvaAssignFrameOffsetsToPromotedStructs();
2603 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2606 void lvaDumpRegLocation(unsigned lclNum);
2607 void lvaDumpFrameLocation(unsigned lclNum);
2608 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2609 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2610 // layout state defined by lvaDoneFrameLayout
2613 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2614 // to avoid bugs from borderline cases.
2615 #define MAX_FrameSize 0x3FFFFFFF
2616 void lvaIncrementFrameSize(unsigned size);
2618 unsigned lvaFrameSize(FrameLayoutState curState);
2620 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2621 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2623 // Returns the caller-SP-relative offset for the local variable "varNum."
2624 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2626 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2627 int lvaGetSPRelativeOffset(unsigned varNum);
2629 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2630 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2632 //------------------------ For splitting types ----------------------------
2634 void lvaInitTypeRef();
2636 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2637 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2638 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2639 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2640 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2641 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2643 void lvaInitVarDsc(LclVarDsc* varDsc,
2645 CorInfoType corInfoType,
2646 CORINFO_CLASS_HANDLE typeHnd,
2647 CORINFO_ARG_LIST_HANDLE varList,
2648 CORINFO_SIG_INFO* varSig);
2650 static unsigned lvaTypeRefMask(var_types type);
2652 var_types lvaGetActualType(unsigned lclNum);
2653 var_types lvaGetRealType(unsigned lclNum);
2655 //-------------------------------------------------------------------------
2659 unsigned lvaLclSize(unsigned varNum);
2660 unsigned lvaLclExactSize(unsigned varNum);
2662 bool lvaLclVarRefs(GenTree* tree, GenTree** findPtr, varRefKinds* refsPtr, void* result);
2664 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2665 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2666 // the return result.
2667 bool lvaLclVarRefsAccum(
2668 GenTree* tree, GenTree** findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2670 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2671 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2672 // and (destructively) unions "trkedVars" into "*result".
2673 void lvaLclVarRefsAccumIntoRes(GenTree** findPtr,
2675 ALLVARSET_VALARG_TP allVars,
2676 VARSET_VALARG_TP trkdVars);
2678 bool lvaHaveManyLocals() const;
2680 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2681 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2682 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2685 void lvaSortByRefCount();
2686 void lvaDumpRefCounts();
2688 void lvaMarkLocalVars(BasicBlock* block);
2690 void lvaMarkLocalVars(); // Local variable ref-counting
2692 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2694 VARSET_VALRET_TP lvaStmtLclMask(GenTree* stmt);
2696 void lvaIncRefCnts(GenTree* tree);
2697 void lvaDecRefCnts(GenTree* tree);
2699 void lvaDecRefCnts(BasicBlock* basicBlock, GenTree* tree);
2700 void lvaRecursiveDecRefCounts(GenTree* tree);
2701 void lvaRecursiveIncRefCounts(GenTree* tree);
2704 struct lvaStressLclFldArgs
2706 Compiler* m_pCompiler;
2710 static fgWalkPreFn lvaStressLclFldCB;
2711 void lvaStressLclFld();
2713 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2714 void lvaDispVarSet(VARSET_VALARG_TP set);
2719 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2721 int lvaFrameAddress(int varNum, bool* pFPbased);
2724 bool lvaIsParameter(unsigned varNum);
2725 bool lvaIsRegArgument(unsigned varNum);
2726 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2727 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2728 // that writes to arg0
2730 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2731 // (this is an overload of lvIsTemp because there are no temp parameters).
2732 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2733 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2734 bool lvaIsImplicitByRefLocal(unsigned varNum)
2736 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2737 LclVarDsc* varDsc = &(lvaTable[varNum]);
2738 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2740 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2743 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2747 // Returns true if this local var is a multireg struct
2748 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2750 // If the local is a TYP_STRUCT, get/set a class handle describing it
2751 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2752 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2754 // If the local is TYP_REF, set or update the associated class information.
2755 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2756 void lvaSetClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2757 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2758 void lvaUpdateClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2760 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2762 // Info about struct fields
2763 struct lvaStructFieldInfo
2765 CORINFO_FIELD_HANDLE fldHnd;
2766 unsigned char fldOffset;
2767 unsigned char fldOrdinal;
2770 CORINFO_CLASS_HANDLE fldTypeHnd;
2773 // Info about struct to be promoted.
2774 struct lvaStructPromotionInfo
2776 CORINFO_CLASS_HANDLE typeHnd;
2778 bool requiresScratchVar;
2781 unsigned char fieldCnt;
2782 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2784 lvaStructPromotionInfo()
2785 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2790 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2791 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2792 lvaStructPromotionInfo* StructPromotionInfo,
2794 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2795 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2796 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2797 #if !defined(_TARGET_64BIT_)
2798 void lvaPromoteLongVars();
2799 #endif // !defined(_TARGET_64BIT_)
2800 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2801 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2802 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2803 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2804 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2805 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2806 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2808 #if defined(FEATURE_SIMD)
2809 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2811 assert(varDsc->lvType == TYP_SIMD12);
2812 assert(varDsc->lvExactSize == 12);
2814 #if defined(_TARGET_64BIT_)
2815 assert(varDsc->lvSize() == 16);
2816 #endif // defined(_TARGET_64BIT_)
2818 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2819 // already does this calculation. However, we also need to prevent mapping types if the var is a
2820 // dependently promoted struct field, which must remain its exact size within its parent struct.
2821 // However, we don't know this until late, so we may have already pretended the field is bigger
2823 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2832 #endif // defined(FEATURE_SIMD)
2834 BYTE* lvaGetGcLayout(unsigned varNum);
2835 bool lvaTypeIsGC(unsigned varNum);
2836 unsigned lvaGSSecurityCookie; // LclVar number
2837 bool lvaTempsHaveLargerOffsetThanVars();
2839 unsigned lvaSecurityObject; // variable representing the security object on the stack
2840 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2842 #if FEATURE_EH_FUNCLETS
2843 unsigned lvaPSPSym; // variable representing the PSPSym
2846 InlineInfo* impInlineInfo;
2847 InlineStrategy* m_inlineStrategy;
2849 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2850 Compiler* impInlineRoot();
2852 #if defined(DEBUG) || defined(INLINE_DATA)
2853 unsigned __int64 getInlineCycleCount()
2855 return m_compCycles;
2857 #endif // defined(DEBUG) || defined(INLINE_DATA)
2859 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2860 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2862 //=========================================================================
2864 //=========================================================================
2867 //---------------- Local variable ref-counting ----------------------------
2870 BasicBlock* lvaMarkRefsCurBlock;
2871 GenTree* lvaMarkRefsCurStmt;
2873 BasicBlock::weight_t lvaMarkRefsWeight;
2875 void lvaMarkLclRefs(GenTree* tree);
2877 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2878 void SetVolatileHint(LclVarDsc* varDsc);
2880 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2881 PerSsaArray lvMemoryPerSsaData;
2882 unsigned lvMemoryNumSsaNames;
2885 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2886 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2887 // not an SSA variable).
2888 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2890 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2891 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2893 assert(ssaNum < lvMemoryNumSsaNames);
2894 return &lvMemoryPerSsaData.GetRef(ssaNum);
2898 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2899 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2903 XX Imports the given method and converts it to semantic trees XX
2905 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2906 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2912 void impImport(BasicBlock* method);
2914 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2915 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2916 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2917 CORINFO_CLASS_HANDLE impGetStringClass();
2918 CORINFO_CLASS_HANDLE impGetObjectClass();
2920 // Returns underlying type of handles returned by ldtoken instruction
2921 inline var_types GetRuntimeHandleUnderlyingType()
2923 // RuntimeTypeHandle is backed by raw pointer on CoreRT and by object reference on other runtimes
2924 return IsTargetAbi(CORINFO_CORERT_ABI) ? TYP_I_IMPL : TYP_REF;
2927 //=========================================================================
2929 //=========================================================================
2932 //-------------------- Stack manipulation ---------------------------------
2934 unsigned impStkSize; // Size of the full stack
2936 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2938 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2940 struct SavedStack // used to save/restore stack contents.
2942 unsigned ssDepth; // number of values on stack
2943 StackEntry* ssTrees; // saved tree values
2946 bool impIsPrimitive(CorInfoType type);
2947 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2949 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2951 void impPushOnStack(GenTree* tree, typeInfo ti);
2952 void impPushNullObjRefOnStack();
2953 StackEntry impPopStack();
2954 StackEntry& impStackTop(unsigned n = 0);
2955 unsigned impStackHeight();
2957 void impSaveStackState(SavedStack* savePtr, bool copy);
2958 void impRestoreStackState(SavedStack* savePtr);
2960 GenTree* impImportLdvirtftn(GenTree* thisPtr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2962 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2964 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2966 bool impCanPInvokeInline();
2967 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2968 void impCheckForPInvokeCall(
2969 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2970 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2971 void impPopArgsForUnmanagedCall(GenTree* call, CORINFO_SIG_INFO* sig);
2973 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2974 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2975 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2977 var_types impImportCall(OPCODE opcode,
2978 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2979 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2981 GenTree* newobjThis,
2983 CORINFO_CALL_INFO* callInfo,
2984 IL_OFFSET rawILOffset);
2986 void impDevirtualizeCall(GenTreeCall* call,
2987 CORINFO_METHOD_HANDLE* method,
2988 unsigned* methodFlags,
2989 CORINFO_CONTEXT_HANDLE* contextHandle,
2990 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2992 CORINFO_CLASS_HANDLE impGetSpecialIntrinsicExactReturnType(CORINFO_METHOD_HANDLE specialIntrinsicHandle);
2994 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2996 GenTree* impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2998 GenTree* impFixupStructReturnType(GenTree* op, CORINFO_CLASS_HANDLE retClsHnd);
3001 var_types impImportJitTestLabelMark(int numArgs);
3004 GenTree* impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
3006 GenTree* impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
3008 GenTree* impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3009 CORINFO_ACCESS_FLAGS access,
3010 CORINFO_FIELD_INFO* pFieldInfo,
3013 static void impBashVarAddrsToI(GenTree* tree1, GenTree* tree2 = nullptr);
3015 GenTree* impImplicitIorI4Cast(GenTree* tree, var_types dstTyp);
3017 GenTree* impImplicitR4orR8Cast(GenTree* tree, var_types dstTyp);
3019 void impImportLeave(BasicBlock* block);
3020 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
3021 GenTree* impIntrinsic(GenTree* newobjThis,
3022 CORINFO_CLASS_HANDLE clsHnd,
3023 CORINFO_METHOD_HANDLE method,
3024 CORINFO_SIG_INFO* sig,
3025 unsigned methodFlags,
3029 CORINFO_RESOLVED_TOKEN* pContstrainedResolvedToken,
3030 CORINFO_THIS_TRANSFORM constraintCallThisTransform,
3031 CorInfoIntrinsics* pIntrinsicID,
3032 bool* isSpecialIntrinsic = nullptr);
3033 GenTree* impMathIntrinsic(CORINFO_METHOD_HANDLE method,
3034 CORINFO_SIG_INFO* sig,
3036 CorInfoIntrinsics intrinsicID,
3038 NamedIntrinsic lookupNamedIntrinsic(CORINFO_METHOD_HANDLE method);
3040 #ifdef FEATURE_HW_INTRINSICS
3041 GenTree* impHWIntrinsic(NamedIntrinsic intrinsic,
3042 CORINFO_METHOD_HANDLE method,
3043 CORINFO_SIG_INFO* sig,
3045 GenTree* impUnsupportedHWIntrinsic(unsigned helper,
3046 CORINFO_METHOD_HANDLE method,
3047 CORINFO_SIG_INFO* sig,
3049 #ifdef _TARGET_XARCH_
3050 static InstructionSet lookupHWIntrinsicISA(const char* className);
3051 static NamedIntrinsic lookupHWIntrinsic(const char* methodName, InstructionSet isa);
3052 static InstructionSet isaOfHWIntrinsic(NamedIntrinsic intrinsic);
3053 static bool isIntrinsicAnIsSupportedPropertyGetter(NamedIntrinsic intrinsic);
3054 static bool isFullyImplmentedISAClass(InstructionSet isa);
3055 GenTree* impSSEIntrinsic(NamedIntrinsic intrinsic,
3056 CORINFO_METHOD_HANDLE method,
3057 CORINFO_SIG_INFO* sig,
3059 GenTree* impSSE2Intrinsic(NamedIntrinsic intrinsic,
3060 CORINFO_METHOD_HANDLE method,
3061 CORINFO_SIG_INFO* sig,
3063 GenTree* impSSE42Intrinsic(NamedIntrinsic intrinsic,
3064 CORINFO_METHOD_HANDLE method,
3065 CORINFO_SIG_INFO* sig,
3067 GenTree* impAvxOrAvx2Intrinsic(NamedIntrinsic intrinsic,
3068 CORINFO_METHOD_HANDLE method,
3069 CORINFO_SIG_INFO* sig,
3071 GenTree* impAESIntrinsic(NamedIntrinsic intrinsic,
3072 CORINFO_METHOD_HANDLE method,
3073 CORINFO_SIG_INFO* sig,
3075 GenTree* impBMI1Intrinsic(NamedIntrinsic intrinsic,
3076 CORINFO_METHOD_HANDLE method,
3077 CORINFO_SIG_INFO* sig,
3079 GenTree* impBMI2Intrinsic(NamedIntrinsic intrinsic,
3080 CORINFO_METHOD_HANDLE method,
3081 CORINFO_SIG_INFO* sig,
3083 GenTree* impFMAIntrinsic(NamedIntrinsic intrinsic,
3084 CORINFO_METHOD_HANDLE method,
3085 CORINFO_SIG_INFO* sig,
3087 GenTree* impLZCNTIntrinsic(NamedIntrinsic intrinsic,
3088 CORINFO_METHOD_HANDLE method,
3089 CORINFO_SIG_INFO* sig,
3091 GenTree* impPCLMULQDQIntrinsic(NamedIntrinsic intrinsic,
3092 CORINFO_METHOD_HANDLE method,
3093 CORINFO_SIG_INFO* sig,
3095 GenTree* impPOPCNTIntrinsic(NamedIntrinsic intrinsic,
3096 CORINFO_METHOD_HANDLE method,
3097 CORINFO_SIG_INFO* sig,
3099 bool compSupportsHWIntrinsic(InstructionSet isa);
3100 bool isScalarISA(InstructionSet isa);
3101 static int ivalOfHWIntrinsic(NamedIntrinsic intrinsic);
3102 unsigned simdSizeOfHWIntrinsic(NamedIntrinsic intrinsic, CORINFO_SIG_INFO* sig);
3103 static GenTree* lastOpOfHWIntrinsic(GenTreeHWIntrinsic* node, int numArgs);
3104 static instruction insOfHWIntrinsic(NamedIntrinsic intrinsic, var_types type);
3107 static HWIntrinsicCategory categoryOfHWIntrinsic(NamedIntrinsic intrinsic);
3108 static int numArgsOfHWIntrinsic(GenTreeHWIntrinsic* node);
3111 static HWIntrinsicFlag flagsOfHWIntrinsic(NamedIntrinsic intrinsic);
3112 GenTree* getArgForHWIntrinsic(var_types argType, CORINFO_CLASS_HANDLE argClass);
3113 static int immUpperBoundOfHWIntrinsic(NamedIntrinsic intrinsic);
3114 GenTree* impNonConstFallback(NamedIntrinsic intrinsic, var_types simdType, var_types baseType);
3115 static bool isImmHWIntrinsic(NamedIntrinsic intrinsic, GenTree* lastOp);
3116 GenTree* addRangeCheckIfNeeded(NamedIntrinsic intrinsic, GenTree* lastOp, bool mustExpand);
3117 bool hwIntrinsicSignatureTypeSupported(var_types retType, CORINFO_SIG_INFO* sig, HWIntrinsicFlag flags);
3118 #endif // _TARGET_XARCH_
3119 #ifdef _TARGET_ARM64_
3120 InstructionSet lookupHWIntrinsicISA(const char* className);
3121 NamedIntrinsic lookupHWIntrinsic(const char* className, const char* methodName);
3122 bool impCheckImmediate(GenTree* immediateOp, unsigned int max);
3123 const HWIntrinsicInfo& getHWIntrinsicInfo(NamedIntrinsic);
3124 #endif // _TARGET_ARM64_
3125 #endif // FEATURE_HW_INTRINSICS
3126 GenTree* impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
3127 CORINFO_SIG_INFO* sig,
3130 CorInfoIntrinsics intrinsicID);
3131 GenTree* impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
3133 GenTree* impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
3135 GenTree* impTransformThis(GenTree* thisPtr,
3136 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
3137 CORINFO_THIS_TRANSFORM transform);
3139 //----------------- Manipulating the trees and stmts ----------------------
3141 GenTree* impTreeList; // Trees for the BB being imported
3142 GenTree* impTreeLast; // The last tree for the current BB
3147 CHECK_SPILL_ALL = -1,
3148 CHECK_SPILL_NONE = -2
3151 void impBeginTreeList();
3152 void impEndTreeList(BasicBlock* block, GenTree* firstStmt, GenTree* lastStmt);
3153 void impEndTreeList(BasicBlock* block);
3154 void impAppendStmtCheck(GenTree* stmt, unsigned chkLevel);
3155 void impAppendStmt(GenTree* stmt, unsigned chkLevel);
3156 void impInsertStmtBefore(GenTree* stmt, GenTree* stmtBefore);
3157 GenTree* impAppendTree(GenTree* tree, unsigned chkLevel, IL_OFFSETX offset);
3158 void impInsertTreeBefore(GenTree* tree, IL_OFFSETX offset, GenTree* stmtBefore);
3159 void impAssignTempGen(unsigned tmp,
3162 GenTree** pAfterStmt = nullptr,
3163 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3164 BasicBlock* block = nullptr);
3165 void impAssignTempGen(unsigned tmpNum,
3167 CORINFO_CLASS_HANDLE structHnd,
3169 GenTree** pAfterStmt = nullptr,
3170 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3171 BasicBlock* block = nullptr);
3172 GenTree* impCloneExpr(GenTree* tree,
3174 CORINFO_CLASS_HANDLE structHnd,
3176 GenTree** pAfterStmt DEBUGARG(const char* reason));
3177 GenTree* impAssignStruct(GenTree* dest,
3179 CORINFO_CLASS_HANDLE structHnd,
3181 GenTree** pAfterStmt = nullptr,
3182 BasicBlock* block = nullptr);
3183 GenTree* impAssignStructPtr(GenTree* dest,
3185 CORINFO_CLASS_HANDLE structHnd,
3187 GenTree** pAfterStmt = nullptr,
3188 BasicBlock* block = nullptr);
3190 GenTree* impGetStructAddr(GenTree* structVal, CORINFO_CLASS_HANDLE structHnd, unsigned curLevel, bool willDeref);
3192 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3193 BYTE* gcLayout = nullptr,
3194 unsigned* numGCVars = nullptr,
3195 var_types* simdBaseType = nullptr);
3197 GenTree* impNormStructVal(GenTree* structVal,
3198 CORINFO_CLASS_HANDLE structHnd,
3200 bool forceNormalization = false);
3202 GenTree* impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3203 BOOL* pRuntimeLookup = nullptr,
3204 BOOL mustRestoreHandle = FALSE,
3205 BOOL importParent = FALSE);
3207 GenTree* impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3208 BOOL* pRuntimeLookup = nullptr,
3209 BOOL mustRestoreHandle = FALSE)
3211 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3214 GenTree* impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3215 CORINFO_LOOKUP* pLookup,
3217 void* compileTimeHandle);
3219 GenTree* getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3221 GenTree* impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3222 CORINFO_LOOKUP* pLookup,
3223 void* compileTimeHandle);
3225 GenTree* impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3227 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3228 CorInfoHelpFunc helper,
3230 GenTreeArgList* arg = nullptr,
3231 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3233 GenTree* impCastClassOrIsInstToTree(GenTree* op1,
3235 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3238 GenTree* impOptimizeCastClassOrIsInst(GenTree* op1, CORINFO_RESOLVED_TOKEN* pResolvedToken, bool isCastClass);
3240 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3241 CORINFO_CLASS_HANDLE typeClass,
3245 bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3246 bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3247 bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3248 bool IsMathIntrinsic(GenTree* tree);
3251 //----------------- Importing the method ----------------------------------
3253 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3256 unsigned impCurOpcOffs;
3257 const char* impCurOpcName;
3258 bool impNestedStackSpill;
3260 // For displaying instrs with generated native code (-n:B)
3261 GenTree* impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3262 void impNoteLastILoffs();
3265 /* IL offset of the stmt currently being imported. It gets set to
3266 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3267 updated at IL offsets for which we have to report mapping info.
3268 It also includes flag bits, so use jitGetILoffs()
3269 to get the actual IL offset value.
3272 IL_OFFSETX impCurStmtOffs;
3273 void impCurStmtOffsSet(IL_OFFSET offs);
3275 void impNoteBranchOffs();
3277 unsigned impInitBlockLineInfo();
3279 GenTree* impCheckForNullPointer(GenTree* obj);
3280 bool impIsThis(GenTree* obj);
3281 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3282 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3283 bool impIsAnySTLOC(OPCODE opcode)
3285 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3286 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3289 GenTreeArgList* impPopList(unsigned count, CORINFO_SIG_INFO* sig, GenTreeArgList* prefixTree = nullptr);
3291 GenTreeArgList* impPopRevList(unsigned count, CORINFO_SIG_INFO* sig, unsigned skipReverseCount = 0);
3294 * Get current IL offset with stack-empty info incoporated
3296 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3298 //---------------- Spilling the importer stack ----------------------------
3300 // The maximum number of bytes of IL processed without clean stack state.
3301 // It allows to limit the maximum tree size and depth.
3302 static const unsigned MAX_TREE_SIZE = 200;
3303 bool impCanSpillNow(OPCODE prevOpcode);
3309 SavedStack pdSavedStack;
3310 ThisInitState pdThisPtrInit;
3313 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3314 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3316 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3317 JitExpandArray<BYTE> impPendingBlockMembers;
3319 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3320 // Operates on the map in the top-level ancestor.
3321 BYTE impGetPendingBlockMember(BasicBlock* blk)
3323 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3326 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3327 // Operates on the map in the top-level ancestor.
3328 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3330 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3333 bool impCanReimport;
3335 bool impSpillStackEntry(unsigned level,
3339 bool bAssertOnRecursion,
3344 void impSpillStackEnsure(bool spillLeaves = false);
3345 void impEvalSideEffects();
3346 void impSpillSpecialSideEff();
3347 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3348 void impSpillValueClasses();
3349 void impSpillEvalStack();
3350 static fgWalkPreFn impFindValueClasses;
3351 void impSpillLclRefs(ssize_t lclNum);
3353 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3355 void impImportBlockCode(BasicBlock* block);
3357 void impReimportMarkBlock(BasicBlock* block);
3358 void impReimportMarkSuccessors(BasicBlock* block);
3360 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3362 void impImportBlockPending(BasicBlock* block);
3364 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3365 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3366 // for the block, but instead, just re-uses the block's existing EntryState.
3367 void impReimportBlockPending(BasicBlock* block);
3369 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTree** pOp1, GenTree** pOp2);
3371 void impImportBlock(BasicBlock* block);
3373 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3374 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3375 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3376 // the variables that will be used -- and for all the predecessors of those successors, and the
3377 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3378 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3379 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3380 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3381 // of local variable numbers, so we represent them with the base local variable number), returns that.
3382 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3383 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3384 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3385 // on which kind of member of the clique the block is).
3386 unsigned impGetSpillTmpBase(BasicBlock* block);
3388 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3389 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3390 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3391 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3392 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3393 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3394 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3395 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3396 // successors receive a native int. Similarly float and double are unified to double.
3397 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3398 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3399 // predecessors, so they insert an upcast if needed).
3400 void impReimportSpillClique(BasicBlock* block);
3402 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3403 // block, and represent the predecessor and successor members of the clique currently being computed.
3404 // *** Access to these will need to be locked in a parallel compiler.
3405 JitExpandArray<BYTE> impSpillCliquePredMembers;
3406 JitExpandArray<BYTE> impSpillCliqueSuccMembers;
3414 // Abstract class for receiving a callback while walking a spill clique
3415 class SpillCliqueWalker
3418 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3421 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3422 class SetSpillTempsBase : public SpillCliqueWalker
3427 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3430 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3433 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3434 class ReimportSpillClique : public SpillCliqueWalker
3439 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3442 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3445 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3446 // predecessor or successor within the spill clique
3447 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3449 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3450 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3451 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3452 void impRetypeEntryStateTemps(BasicBlock* blk);
3454 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3455 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3457 void impPushVar(GenTree* op, typeInfo tiRetVal);
3458 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3459 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3461 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3463 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3464 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3465 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3468 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTree* op, CORINFO_CLASS_HANDLE hClass);
3471 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3472 struct BlockListNode
3475 BlockListNode* m_next;
3476 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3479 void* operator new(size_t sz, Compiler* comp);
3481 BlockListNode* impBlockListNodeFreeList;
3483 BlockListNode* AllocBlockListNode();
3484 void FreeBlockListNode(BlockListNode* node);
3486 bool impIsValueType(typeInfo* pTypeInfo);
3487 var_types mangleVarArgsType(var_types type);
3490 regNumber getCallArgIntRegister(regNumber floatReg);
3491 regNumber getCallArgFloatRegister(regNumber intReg);
3492 #endif // FEATURE_VARARG
3495 static unsigned jitTotalMethodCompiled;
3499 static LONG jitNestingLevel;
3502 static BOOL impIsAddressInLocal(GenTree* tree, GenTree** lclVarTreeOut);
3504 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3506 // STATIC inlining decision based on the IL code.
3507 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3508 CORINFO_METHOD_INFO* methInfo,
3510 InlineResult* inlineResult);
3512 void impCheckCanInline(GenTree* call,
3513 CORINFO_METHOD_HANDLE fncHandle,
3515 CORINFO_CONTEXT_HANDLE exactContextHnd,
3516 InlineCandidateInfo** ppInlineCandidateInfo,
3517 InlineResult* inlineResult);
3519 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3522 InlineResult* inlineResult);
3524 void impInlineInitVars(InlineInfo* pInlineInfo);
3526 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3528 GenTree* impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3530 BOOL impInlineIsThis(GenTree* tree, InlArgInfo* inlArgInfo);
3532 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTree* additionalTreesToBeEvaluatedBefore,
3533 GenTree* variableBeingDereferenced,
3534 InlArgInfo* inlArgInfo);
3536 void impMarkInlineCandidate(GenTree* call,
3537 CORINFO_CONTEXT_HANDLE exactContextHnd,
3538 bool exactContextNeedsRuntimeLookup,
3539 CORINFO_CALL_INFO* callInfo);
3541 bool impTailCallRetTypeCompatible(var_types callerRetType,
3542 CORINFO_CLASS_HANDLE callerRetTypeClass,
3543 var_types calleeRetType,
3544 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3546 bool impIsTailCallILPattern(bool tailPrefixed,
3548 const BYTE* codeAddrOfNextOpcode,
3549 const BYTE* codeEnd,
3551 bool* IsCallPopRet = nullptr);
3553 bool impIsImplicitTailCallCandidate(
3554 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3556 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3559 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3560 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3564 XX Info about the basic-blocks, their contents and the flow analysis XX
3566 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3567 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3571 BasicBlock* fgFirstBB; // Beginning of the basic block list
3572 BasicBlock* fgLastBB; // End of the basic block list
3573 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3574 #if FEATURE_EH_FUNCLETS
3575 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3577 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3579 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3580 unsigned fgEdgeCount; // # of control flow edges between the BBs
3581 unsigned fgBBcount; // # of BBs in the method
3583 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3585 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3586 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3587 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3588 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3590 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3591 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3592 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3593 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3594 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3595 // index). The arrays are of size fgBBNumMax + 1.
3596 unsigned* fgDomTreePreOrder;
3597 unsigned* fgDomTreePostOrder;
3599 bool fgBBVarSetsInited;
3601 // Allocate array like T* a = new T[fgBBNumMax + 1];
3602 // Using helper so we don't keep forgetting +1.
3603 template <typename T>
3604 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3606 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3609 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3610 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3611 // cannot be meaningfully combined. Note that new blocks can be created with higher
3612 // block numbers without changing the basic block epoch. These blocks *cannot*
3613 // participate in a block set until the blocks are all renumbered, causing the epoch
3614 // to change. This is useful if continuing to use previous block sets is valuable.
3615 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3616 unsigned fgCurBBEpoch;
3618 unsigned GetCurBasicBlockEpoch()
3620 return fgCurBBEpoch;
3623 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3624 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3625 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3626 unsigned fgCurBBEpochSize;
3628 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3629 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3630 unsigned fgBBSetCountInSizeTUnits;
3632 void NewBasicBlockEpoch()
3634 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3636 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3638 fgCurBBEpochSize = fgBBNumMax + 1;
3639 fgBBSetCountInSizeTUnits =
3640 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3643 // All BlockSet objects are now invalid!
3644 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3645 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3649 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3650 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3651 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3652 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3654 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3655 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3656 // array of size_t bitsets), then print that out.
3657 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3664 void EnsureBasicBlockEpoch()
3666 if (fgCurBBEpochSize != fgBBNumMax + 1)
3668 NewBasicBlockEpoch();
3672 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3673 void fgEnsureFirstBBisScratch();
3674 bool fgFirstBBisScratch();
3675 bool fgBBisScratch(BasicBlock* block);
3677 void fgExtendEHRegionBefore(BasicBlock* block);
3678 void fgExtendEHRegionAfter(BasicBlock* block);
3680 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3682 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3684 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3687 BasicBlock* nearBlk,
3688 bool putInFilter = false,
3689 bool runRarely = false,
3690 bool insertAtEnd = false);
3692 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3694 bool runRarely = false,
3695 bool insertAtEnd = false);
3697 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3699 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3700 BasicBlock* afterBlk,
3701 unsigned xcptnIndex,
3702 bool putInTryRegion);
3704 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3705 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3706 void fgUnlinkBlock(BasicBlock* block);
3708 unsigned fgMeasureIR();
3710 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3711 bool fgMultipleNots;
3714 bool fgModified; // True if the flow graph has been modified recently
3715 bool fgComputePredsDone; // Have we computed the bbPreds list
3716 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3717 bool fgDomsComputed; // Have we computed the dominator sets?
3718 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3720 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3721 bool fgHasPostfix; // any postfix ++/-- found?
3722 unsigned fgIncrCount; // number of increment nodes found
3724 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3728 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3729 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3732 bool fgRemoveRestOfBlock; // true if we know that we will throw
3733 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3735 // There are two modes for ordering of the trees.
3736 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3737 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3738 // by traversing the tree according to the order of the operands.
3739 // - In FGOrderLinear, the dominant ordering is the linear order.
3746 FlowGraphOrder fgOrder;
3748 // The following are boolean flags that keep track of the state of internal data structures
3750 bool fgStmtListThreaded; // true if the node list is now threaded
3751 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3752 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3753 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3754 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3755 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3756 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3757 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3758 // This is derived from the profile data
3759 // or is BB_UNITY_WEIGHT when we don't have profile data
3761 #if FEATURE_EH_FUNCLETS
3762 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3763 #endif // FEATURE_EH_FUNCLETS
3765 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3766 // since fgMorphTree can be called from several places
3768 bool impBoxTempInUse; // the temp below is valid and available
3769 unsigned impBoxTemp; // a temporary that is used for boxing
3772 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3773 // and we are trying to compile again in a "safer", minopts mode?
3777 unsigned impInlinedCodeSize;
3780 //-------------------------------------------------------------------------
3786 void fgTransformFatCalli();
3790 void fgRemoveEmptyTry();
3792 void fgRemoveEmptyFinally();
3794 void fgMergeFinallyChains();
3796 void fgCloneFinally();
3798 void fgCleanupContinuation(BasicBlock* continuation);
3800 void fgUpdateFinallyTargetFlags();
3802 void fgClearAllFinallyTargetBits();
3804 void fgAddFinallyTargetFlags();
3806 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3807 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3808 // when this is necessary.
3809 bool fgNeedToAddFinallyTargetBits;
3810 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3812 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3813 BasicBlock* handler,
3814 BlockToBlockMap& continuationMap);
3816 GenTree* fgGetCritSectOfStaticMethod();
3818 #if FEATURE_EH_FUNCLETS
3820 void fgAddSyncMethodEnterExit();
3822 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3824 void fgConvertSyncReturnToLeave(BasicBlock* block);
3826 #endif // FEATURE_EH_FUNCLETS
3828 void fgAddReversePInvokeEnterExit();
3830 bool fgMoreThanOneReturnBlock();
3832 // The number of separate return points in the method.
3833 unsigned fgReturnCount;
3835 void fgAddInternal();
3837 bool fgFoldConditional(BasicBlock* block);
3839 #ifdef LEGACY_BACKEND
3840 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3842 void fgMorphStmts(BasicBlock* block, bool* lnot, bool* loadw);
3844 void fgMorphBlocks();
3846 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3848 void fgCheckArgCnt();
3849 void fgSetOptions();
3852 static fgWalkPreFn fgAssertNoQmark;
3853 void fgPreExpandQmarkChecks(GenTree* expr);
3854 void fgPostExpandQmarkChecks();
3855 static void fgCheckQmarkAllowedForm(GenTree* tree);
3858 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3860 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3861 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3862 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3863 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3864 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3866 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, BasicBlock* block, IL_OFFSETX offs);
3867 GenTreeStmt* fgNewStmtFromTree(GenTree* tree);
3868 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, BasicBlock* block);
3869 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, IL_OFFSETX offs);
3871 GenTree* fgGetTopLevelQmark(GenTree* expr, GenTree** ppDst = nullptr);
3872 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTree* stmt);
3873 void fgExpandQmarkStmt(BasicBlock* block, GenTree* expr);
3874 void fgExpandQmarkNodes();
3878 // Do "simple lowering." This functionality is (conceptually) part of "general"
3879 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3880 void fgSimpleLowering();
3882 #ifdef LEGACY_BACKEND
3883 bool fgShouldCreateAssignOp(GenTree* tree, bool* bReverse);
3886 GenTree* fgInitThisClass();
3888 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3890 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3892 inline bool backendRequiresLocalVarLifetimes()
3894 #if defined(LEGACY_BACKEND)
3897 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3901 void fgLocalVarLiveness();
3903 void fgLocalVarLivenessInit();
3905 #ifdef LEGACY_BACKEND
3906 GenTree* fgLegacyPerStatementLocalVarLiveness(GenTree* startNode, GenTree* relopNode);
3908 void fgPerNodeLocalVarLiveness(GenTree* node);
3910 void fgPerBlockLocalVarLiveness();
3912 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3914 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3916 // This is used in the liveness computation, as a temporary. When we use the
3917 // arbitrary-length VarSet representation, it is better not to allocate a new one
3919 VARSET_TP fgMarkIntfUnionVS;
3921 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3923 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3925 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3927 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTree* clonedTree);
3929 void fgUpdateRefCntForExtract(GenTree* wholeTree, GenTree* keptTree);
3931 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3933 void fgComputeLifeTrackedLocalUse(VARSET_TP& life, LclVarDsc& varDsc, GenTreeLclVarCommon* node);
3934 bool fgComputeLifeTrackedLocalDef(VARSET_TP& life,
3935 VARSET_VALARG_TP keepAliveVars,
3937 GenTreeLclVarCommon* node);
3938 void fgComputeLifeUntrackedLocal(VARSET_TP& life,
3939 VARSET_VALARG_TP keepAliveVars,
3941 GenTreeLclVarCommon* lclVarNode,
3943 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3945 void fgComputeLife(VARSET_TP& life,
3948 VARSET_VALARG_TP volatileVars,
3949 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3951 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3953 bool fgRemoveDeadStore(GenTree** pTree,
3955 VARSET_VALARG_TP life,
3957 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3959 // For updating liveset during traversal AFTER fgComputeLife has completed
3960 VARSET_VALRET_TP fgGetVarBits(GenTree* tree);
3961 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTree* tree);
3963 // Returns the set of live variables after endTree,
3964 // assuming that liveSet is the set of live variables BEFORE tree.
3965 // Requires that fgComputeLife has completed, and that tree is in the same
3966 // statement as endTree, and that it comes before endTree in execution order
3968 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTree* tree, GenTree* endTree)
3970 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3971 while (tree != nullptr && tree != endTree->gtNext)
3973 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3974 tree = tree->gtNext;
3976 assert(tree == endTree->gtNext);
3980 void fgInterBlockLocalVarLiveness();
3982 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3983 // "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
3984 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3985 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3986 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, unsigned> NodeToUnsignedMap;
3987 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3988 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3990 if (m_opAsgnVarDefSsaNums == nullptr)
3992 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3994 return m_opAsgnVarDefSsaNums;
3997 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3998 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3999 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
4001 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTree* tree);
4003 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
4004 // Except: assumes that lcl is a def, and if it is
4005 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
4006 // rather than the "use" SSA number recorded in the tree "lcl".
4007 inline unsigned GetSsaNumForLocalVarDef(GenTree* lcl);
4009 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
4010 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
4011 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
4012 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
4013 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
4015 // (byref addrS1 = &s1,
4016 // *(addrS1 * offsetof(f0)) = s2f0,
4018 // *(addrS1 * offsetof(fn)) = s2fn)
4020 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
4021 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
4022 // give it SSA names and value numbers?
4024 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
4025 // end with an instance of the structure below, whose fields are described in the declaration.
4026 struct IndirectAssignmentAnnotation
4028 unsigned m_lclNum; // The local num that is being indirectly assigned.
4029 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
4030 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
4031 // be the singleton field sequence "g". The individual assignments would
4032 // further append the fields of "s.g" to that.
4033 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
4034 // structure has a single field).
4035 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
4036 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
4039 IndirectAssignmentAnnotation(unsigned lclNum,
4040 FieldSeqNode* fldSeq,
4042 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
4043 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
4044 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
4048 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*> NodeToIndirAssignMap;
4049 NodeToIndirAssignMap* m_indirAssignMap;
4050 NodeToIndirAssignMap* GetIndirAssignMap()
4052 if (m_indirAssignMap == nullptr)
4054 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
4055 CompAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
4056 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
4058 return m_indirAssignMap;
4061 // Performs SSA conversion.
4064 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
4065 void fgResetForSsa();
4067 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
4069 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
4070 inline bool fgExcludeFromSsa(unsigned lclNum);
4072 // Returns "true" if a struct temp of the given type requires needs zero init in this block
4073 inline bool fgStructTempNeedsExplicitZeroInit(LclVarDsc* varDsc, BasicBlock* block);
4075 // The value numbers for this compilation.
4076 ValueNumStore* vnStore;
4079 ValueNumStore* GetValueNumStore()
4084 // Do value numbering (assign a value number to each
4086 void fgValueNumber();
4088 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
4089 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
4090 // The 'indType' is the indirection type of the lhs of the assignment and will typically
4091 // match the element type of the array or fldSeq. When this type doesn't match
4092 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
4094 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
4097 FieldSeqNode* fldSeq,
4101 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
4102 // has been parsed to yield the other input arguments. If evaluation of the address
4103 // can raise exceptions, those should be captured in the exception set "excVN."
4104 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
4105 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
4106 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
4107 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
4108 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
4110 ValueNum fgValueNumberArrIndexVal(GenTree* tree,
4111 CORINFO_CLASS_HANDLE elemTypeEq,
4115 FieldSeqNode* fldSeq);
4117 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
4118 // by evaluating the array index expression "tree". Returns the value number resulting from
4119 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
4120 // "GT_IND" that does the dereference, and it is given the returned value number.
4121 ValueNum fgValueNumberArrIndexVal(GenTree* tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
4123 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
4124 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
4126 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
4128 // Utility functions for fgValueNumber.
4130 // Perform value-numbering for the trees in "blk".
4131 void fgValueNumberBlock(BasicBlock* blk);
4133 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
4134 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
4135 // assumed for the memoryKind at the start "entryBlk".
4136 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
4138 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
4139 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
4140 void fgMutateGcHeap(GenTree* tree DEBUGARG(const char* msg));
4142 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
4144 void fgMutateAddressExposedLocal(GenTree* tree DEBUGARG(const char* msg));
4146 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
4147 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
4148 void recordGcHeapStore(GenTree* curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
4150 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
4151 void recordAddressExposedLocalStore(GenTree* curTree, ValueNum memoryVN DEBUGARG(const char* msg));
4153 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
4154 // value in that SSA #.
4155 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTree* tree);
4157 // The input 'tree' is a leaf node that is a constant
4158 // Assign the proper value number to the tree
4159 void fgValueNumberTreeConst(GenTree* tree);
4161 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
4162 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4164 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4166 void fgValueNumberTree(GenTree* tree, bool evalAsgLhsInd = false);
4168 // Does value-numbering for a block assignment.
4169 void fgValueNumberBlockAssignment(GenTree* tree, bool evalAsgLhsInd);
4171 // Does value-numbering for a cast tree.
4172 void fgValueNumberCastTree(GenTree* tree);
4174 // Does value-numbering for an intrinsic tree.
4175 void fgValueNumberIntrinsic(GenTree* tree);
4177 // Does value-numbering for a call. We interpret some helper calls.
4178 void fgValueNumberCall(GenTreeCall* call);
4180 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4181 void fgUpdateArgListVNs(GenTreeArgList* args);
4183 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4184 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4186 // Requires "helpCall" to be a helper call. Assigns it a value number;
4187 // we understand the semantics of some of the calls. Returns "true" if
4188 // the call may modify the heap (we assume arbitrary memory side effects if so).
4189 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4191 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4192 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4194 // These are the current value number for the memory implicit variables while
4195 // doing value numbering. These are the value numbers under the "liberal" interpretation
4196 // of memory values; the "conservative" interpretation needs no VN, since every access of
4197 // memory yields an unknown value.
4198 ValueNum fgCurMemoryVN[MemoryKindCount];
4200 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4201 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4202 // is 1, and the rest is an encoding of "elemTyp".
4203 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4205 if (elemStructType != nullptr)
4207 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4208 varTypeIsIntegral(elemTyp));
4209 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4210 return elemStructType;
4214 elemTyp = varTypeUnsignedToSigned(elemTyp);
4215 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4218 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4219 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4220 // the struct type of the element).
4221 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4223 size_t clsHndVal = size_t(clsHnd);
4224 if (clsHndVal & 0x1)
4226 return var_types(clsHndVal >> 1);
4234 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4235 var_types getJitGCType(BYTE gcType);
4237 enum structPassingKind
4239 SPK_Unknown, // Invalid value, never returned
4240 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4241 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4242 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4243 // parameters registers are used, then the stack will be used)
4244 // for X86 passed on the stack, for ARM32 passed in registers
4245 // or the stack or split between registers and the stack.
4246 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4248 }; // The struct is passed/returned by reference to a copy/buffer.
4250 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4251 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4252 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4253 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4255 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4257 // Get the type that is used to pass values of the given struct type.
4258 // If you have already retrieved the struct size then pass it as the optional third argument
4260 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4261 structPassingKind* wbPassStruct,
4262 unsigned structSize = 0);
4264 // Get the type that is used to return values of the given struct type.
4265 // If you have already retrieved the struct size then pass it as the optional third argument
4267 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4268 structPassingKind* wbPassStruct = nullptr,
4269 unsigned structSize = 0);
4272 // Print a representation of "vnp" or "vn" on standard output.
4273 // If "level" is non-zero, we also print out a partial expansion of the value.
4274 void vnpPrint(ValueNumPair vnp, unsigned level);
4275 void vnPrint(ValueNum vn, unsigned level);
4278 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4280 // Dominator computation member functions
4281 // Not exposed outside Compiler
4283 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4285 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4286 // flow graph. We first assume the fields bbIDom on each
4287 // basic block are invalid. This computation is needed later
4288 // by fgBuildDomTree to build the dominance tree structure.
4289 // Based on: A Simple, Fast Dominance Algorithm
4290 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4292 void fgCompDominatedByExceptionalEntryBlocks();
4294 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4295 // Note: this is relatively slow compared to calling fgDominate(),
4296 // especially if dealing with a single block versus block check.
4298 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4300 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4302 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4304 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4306 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4308 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4309 // processed in topological sort, this function takes care of that.
4311 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4313 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4314 // Returns this as a set.
4316 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4317 // root nodes. Returns this as a set.
4320 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4323 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4324 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4327 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4328 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4329 // && postOrder(A) >= postOrder(B) making the computation O(1).
4330 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4332 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4334 void fgUpdateChangedFlowGraph();
4337 // Compute the predecessors of the blocks in the control flow graph.
4338 void fgComputePreds();
4340 // Remove all predecessor information.
4341 void fgRemovePreds();
4343 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4344 // before the full predecessors lists are computed.
4345 void fgComputeCheapPreds();
4348 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4350 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4360 // Initialize the per-block variable sets (used for liveness analysis).
4361 void fgInitBlockVarSets();
4363 // true if we've gone through and created GC Poll calls.
4364 bool fgGCPollsCreated;
4365 void fgMarkGCPollBlocks();
4366 void fgCreateGCPolls();
4367 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4369 // Requires that "block" is a block that returns from
4370 // a finally. Returns the number of successors (jump targets of
4371 // of blocks in the covered "try" that did a "LEAVE".)
4372 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4374 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4375 // a finally. Returns its "i"th successor (jump targets of
4376 // of blocks in the covered "try" that did a "LEAVE".)
4377 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4378 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4381 // Factor out common portions of the impls of the methods above.
4382 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4385 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4386 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4387 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4388 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4389 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4390 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4391 // we leave the entry associated with the block, but it will no longer be accessed.)
4392 struct SwitchUniqueSuccSet
4394 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4395 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4398 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4399 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4400 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4401 void UpdateTarget(CompAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4404 typedef JitHashTable<BasicBlock*, JitPtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet> BlockToSwitchDescMap;
4407 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4408 // iteration over only the distinct successors.
4409 BlockToSwitchDescMap* m_switchDescMap;
4412 BlockToSwitchDescMap* GetSwitchDescMap(bool createIfNull = true)
4414 if ((m_switchDescMap == nullptr) && createIfNull)
4416 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4418 return m_switchDescMap;
4421 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4422 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4423 // we don't accidentally look up and return the wrong switch data.
4424 void InvalidateUniqueSwitchSuccMap()
4426 m_switchDescMap = nullptr;
4429 // Requires "switchBlock" to be a block that ends in a switch. Returns
4430 // the corresponding SwitchUniqueSuccSet.
4431 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4433 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4434 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4435 // remove it from "this", and ensure that "to" is a member.
4436 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4438 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4439 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4441 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4443 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4445 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4447 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4449 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4451 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4453 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4455 void fgRemoveBlockAsPred(BasicBlock* block);
4457 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4459 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4461 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4463 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4465 flowList* fgAddRefPred(BasicBlock* block,
4466 BasicBlock* blockPred,
4467 flowList* oldEdge = nullptr,
4468 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4471 void fgFindBasicBlocks();
4473 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4475 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4477 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4478 bool putInTryRegion,
4479 BasicBlock* startBlk,
4481 BasicBlock* nearBlk,
4482 BasicBlock* jumpBlk,
4485 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4487 void fgRemoveEmptyBlocks();
4489 void fgRemoveStmt(BasicBlock* block, GenTree* stmt, bool updateRefCnt = true);
4491 bool fgCheckRemoveStmt(BasicBlock* block, GenTree* stmt);
4493 void fgCreateLoopPreHeader(unsigned lnum);
4495 void fgUnreachableBlock(BasicBlock* block);
4497 void fgRemoveConditionalJump(BasicBlock* block);
4499 BasicBlock* fgLastBBInMainFunction();
4501 BasicBlock* fgEndBBAfterMainFunction();
4503 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4505 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4507 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4509 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4511 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4513 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4515 bool fgRenumberBlocks();
4517 bool fgExpandRarelyRunBlocks();
4519 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4521 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4523 enum FG_RELOCATE_TYPE
4525 FG_RELOCATE_TRY, // relocate the 'try' region
4526 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4528 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4530 #if FEATURE_EH_FUNCLETS
4531 #if defined(_TARGET_ARM_)
4532 void fgClearFinallyTargetBit(BasicBlock* block);
4533 #endif // defined(_TARGET_ARM_)
4534 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4535 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4536 void fgInsertFuncletPrologBlock(BasicBlock* block);
4537 void fgCreateFuncletPrologBlocks();
4538 void fgCreateFunclets();
4539 #else // !FEATURE_EH_FUNCLETS
4540 bool fgRelocateEHRegions();
4541 #endif // !FEATURE_EH_FUNCLETS
4543 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4545 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4547 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4549 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4551 bool fgOptimizeEmptyBlock(BasicBlock* block);
4553 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4555 bool fgOptimizeBranch(BasicBlock* bJump);
4557 bool fgOptimizeSwitchBranches(BasicBlock* block);
4559 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4561 bool fgOptimizeSwitchJumps();
4563 void fgPrintEdgeWeights();
4565 void fgComputeEdgeWeights();
4567 void fgReorderBlocks();
4569 void fgDetermineFirstColdBlock();
4571 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4573 bool fgUpdateFlowGraph(bool doTailDup = false);
4575 void fgFindOperOrder();
4577 // method that returns if you should split here
4578 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4580 void fgSetBlockOrder();
4582 void fgRemoveReturnBlock(BasicBlock* block);
4584 /* Helper code that has been factored out */
4585 inline void fgConvertBBToThrowBB(BasicBlock* block);
4587 bool fgCastNeeded(GenTree* tree, var_types toType);
4588 GenTree* fgDoNormalizeOnStore(GenTree* tree);
4589 GenTree* fgMakeTmpArgNode(
4590 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4592 // The following check for loops that don't execute calls
4593 bool fgLoopCallMarked;
4595 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4596 void fgLoopCallMark();
4598 void fgMarkLoopHead(BasicBlock* block);
4600 unsigned fgGetCodeEstimate(BasicBlock* block);
4603 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4604 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4605 bool fgDumpFlowGraph(Phases phase);
4607 #endif // DUMP_FLOWGRAPHS
4612 void fgDispBBLiveness(BasicBlock* block);
4613 void fgDispBBLiveness();
4614 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4615 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4616 void fgDispBasicBlocks(bool dumpTrees = false);
4617 void fgDumpStmtTree(GenTree* stmt, unsigned bbNum);
4618 void fgDumpBlock(BasicBlock* block);
4619 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4621 static fgWalkPreFn fgStress64RsltMulCB;
4622 void fgStress64RsltMul();
4623 void fgDebugCheckUpdate();
4624 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4625 void fgDebugCheckBlockLinks();
4626 void fgDebugCheckLinks(bool morphTrees = false);
4627 void fgDebugCheckStmtsList(BasicBlock* block, bool morphTrees);
4628 void fgDebugCheckNodeLinks(BasicBlock* block, GenTree* stmt);
4629 void fgDebugCheckNodesUniqueness();
4631 void fgDebugCheckFlags(GenTree* tree);
4632 void fgDebugCheckFlagsHelper(GenTree* tree, unsigned treeFlags, unsigned chkFlags);
4633 void fgDebugCheckTryFinallyExits();
4636 #ifdef LEGACY_BACKEND
4637 static void fgOrderBlockOps(GenTree* tree,
4641 GenTree** opsPtr, // OUT
4642 regMaskTP* regsPtr); // OUT
4643 #endif // LEGACY_BACKEND
4645 static GenTree* fgGetFirstNode(GenTree* tree);
4646 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4647 void fgTraverseRPO();
4649 //--------------------- Walking the trees in the IR -----------------------
4654 fgWalkPreFn* wtprVisitorFn;
4655 fgWalkPostFn* wtpoVisitorFn;
4656 void* pCallbackData; // user-provided data
4657 bool wtprLclsOnly; // whether to only visit lclvar nodes
4658 GenTree* parent; // parent of current node, provided to callback
4659 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4661 bool printModified; // callback can use this
4665 fgWalkResult fgWalkTreePre(GenTree** pTree,
4666 fgWalkPreFn* visitor,
4667 void* pCallBackData = nullptr,
4668 bool lclVarsOnly = false,
4669 bool computeStack = false);
4671 fgWalkResult fgWalkTree(GenTree** pTree,
4672 fgWalkPreFn* preVisitor,
4673 fgWalkPostFn* postVisitor,
4674 void* pCallBackData = nullptr);
4676 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4680 fgWalkResult fgWalkTreePost(GenTree** pTree,
4681 fgWalkPostFn* visitor,
4682 void* pCallBackData = nullptr,
4683 bool computeStack = false);
4685 // An fgWalkPreFn that looks for expressions that have inline throws in
4686 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4687 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4688 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4689 // properly propagated to parent trees). It returns WALK_CONTINUE
4691 static fgWalkResult fgChkThrowCB(GenTree** pTree, Compiler::fgWalkData* data);
4692 static fgWalkResult fgChkLocAllocCB(GenTree** pTree, Compiler::fgWalkData* data);
4693 static fgWalkResult fgChkQmarkCB(GenTree** pTree, Compiler::fgWalkData* data);
4695 /**************************************************************************
4697 *************************************************************************/
4700 friend class SsaBuilder;
4701 friend struct ValueNumberState;
4703 //--------------------- Detect the basic blocks ---------------------------
4705 BasicBlock** fgBBs; // Table of pointers to the BBs
4707 void fgInitBBLookup();
4708 BasicBlock* fgLookupBB(unsigned addr);
4710 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4712 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4714 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4716 void fgLinkBasicBlocks();
4718 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4720 void fgCheckBasicBlockControlFlow();
4722 void fgControlFlowPermitted(BasicBlock* blkSrc,
4723 BasicBlock* blkDest,
4724 BOOL IsLeave = false /* is the src a leave block */);
4726 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4728 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4730 void fgAdjustForAddressExposedOrWrittenThis();
4732 bool fgProfileData_ILSizeMismatch;
4733 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4734 ULONG fgProfileBufferCount;
4735 ULONG fgNumProfileRuns;
4737 unsigned fgStressBBProf()
4740 unsigned result = JitConfig.JitStressBBProf();
4743 if (compStressCompile(STRESS_BB_PROFILE, 15))
4754 bool fgHaveProfileData();
4755 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4756 void fgInstrumentMethod();
4759 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4760 // or if we have some fake profile data for the stress mode
4761 bool fgIsUsingProfileWeights()
4763 return (fgHaveProfileData() || fgStressBBProf());
4766 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4767 // or BB_UNITY_WEIGHT when we aren't using profile data.
4768 unsigned fgProfileRunsCount()
4770 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4773 //-------- Insert a statement at the start or end of a basic block --------
4777 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4781 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTree* node);
4783 public: // Used by linear scan register allocation
4784 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTree* node);
4787 GenTree* fgInsertStmtAtBeg(BasicBlock* block, GenTree* stmt);
4788 GenTree* fgInsertStmtAfter(BasicBlock* block, GenTree* insertionPoint, GenTree* stmt);
4790 public: // Used by linear scan register allocation
4791 GenTree* fgInsertStmtBefore(BasicBlock* block, GenTree* insertionPoint, GenTree* stmt);
4794 GenTree* fgInsertStmtListAfter(BasicBlock* block, GenTree* stmtAfter, GenTree* stmtList);
4796 GenTree* fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4798 // Create a new temporary variable to hold the result of *ppTree,
4799 // and transform the graph accordingly.
4800 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4801 GenTree* fgMakeMultiUse(GenTree** ppTree);
4804 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4805 GenTree* fgRecognizeAndMorphBitwiseRotation(GenTree* tree);
4806 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4808 //-------- Determine the order in which the trees will be evaluated -------
4810 unsigned fgTreeSeqNum;
4811 GenTree* fgTreeSeqLst;
4812 GenTree* fgTreeSeqBeg;
4814 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4815 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4816 void fgSetTreeSeqFinish(GenTree* tree, bool isLIR);
4817 void fgSetStmtSeq(GenTree* tree);
4818 void fgSetBlockOrder(BasicBlock* block);
4820 //------------------------- Morphing --------------------------------------
4822 unsigned fgPtrArgCntCur;
4823 unsigned fgPtrArgCntMax;
4826 //------------------------------------------------------------------------
4827 // fgGetPtrArgCntMax: Return the maximum number of pointer-sized stack arguments that calls inside this method
4828 // can push on the stack. This value is calculated during morph.
4831 // Returns fgPtrArgCntMax, that is a private field.
4833 unsigned fgGetPtrArgCntMax() const
4835 return fgPtrArgCntMax;
4838 //------------------------------------------------------------------------
4839 // fgSetPtrArgCntMax: Set the maximum number of pointer-sized stack arguments that calls inside this method
4840 // can push on the stack. This function is used during StackLevelSetter to fix incorrect morph calculations.
4842 void fgSetPtrArgCntMax(unsigned argCntMax)
4844 fgPtrArgCntMax = argCntMax;
4848 hashBv* fgOutgoingArgTemps;
4849 hashBv* fgCurrentlyInUseArgTemps;
4851 bool compCanEncodePtrArgCntMax();
4853 void fgSetRngChkTarget(GenTree* tree, bool delay = true);
4855 BasicBlock* fgSetRngChkTargetInner(SpecialCodeKind kind, bool delay, unsigned* stkDepth);
4858 void fgMoveOpsLeft(GenTree* tree);
4861 bool fgIsCommaThrow(GenTree* tree, bool forFolding = false);
4863 bool fgIsThrow(GenTree* tree);
4865 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4866 bool fgIsBlockCold(BasicBlock* block);
4868 GenTree* fgMorphCastIntoHelper(GenTree* tree, int helper, GenTree* oper);
4870 GenTree* fgMorphIntoHelperCall(GenTree* tree, int helper, GenTreeArgList* args);
4872 GenTree* fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4874 bool fgMorphRelopToQmark(GenTree* tree);
4876 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4877 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4878 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4879 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4880 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4881 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4882 // small; hence the other fields of MorphAddrContext.
4883 enum MorphAddrContextKind
4888 struct MorphAddrContext
4890 MorphAddrContextKind m_kind;
4891 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4892 // top-level indirection and here have been constants.
4893 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4894 // In that case, is the sum of those constant offsets.
4896 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4901 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4902 static MorphAddrContext s_CopyBlockMAC;
4905 GenTree* getSIMDStructFromField(GenTree* tree,
4906 var_types* baseTypeOut,
4908 unsigned* simdSizeOut,
4909 bool ignoreUsedInSIMDIntrinsic = false);
4910 GenTree* fgMorphFieldAssignToSIMDIntrinsicSet(GenTree* tree);
4911 GenTree* fgMorphFieldToSIMDIntrinsicGet(GenTree* tree);
4912 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTree* stmt);
4913 void impMarkContiguousSIMDFieldAssignments(GenTree* stmt);
4915 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4916 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4917 GenTree* fgPreviousCandidateSIMDFieldAsgStmt;
4919 #endif // FEATURE_SIMD
4920 GenTree* fgMorphArrayIndex(GenTree* tree);
4921 GenTree* fgMorphCast(GenTree* tree);
4922 GenTree* fgUnwrapProxy(GenTree* objRef);
4923 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4925 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4928 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4929 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4931 void fgFixupStructReturn(GenTree* call);
4932 GenTree* fgMorphLocalVar(GenTree* tree, bool forceRemorph);
4935 bool fgAddrCouldBeNull(GenTree* addr);
4938 GenTree* fgMorphField(GenTree* tree, MorphAddrContext* mac);
4939 bool fgCanFastTailCall(GenTreeCall* call);
4940 bool fgCheckStmtAfterTailCall();
4941 void fgMorphTailCall(GenTreeCall* call);
4942 GenTree* fgGetStubAddrArg(GenTreeCall* call);
4943 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4944 GenTree* fgAssignRecursiveCallArgToCallerParam(GenTree* arg,
4945 fgArgTabEntry* argTabEntry,
4947 IL_OFFSETX callILOffset,
4948 GenTree* tmpAssignmentInsertionPoint,
4949 GenTree* paramAssignmentInsertionPoint);
4950 static int fgEstimateCallStackSize(GenTreeCall* call);
4951 GenTree* fgMorphCall(GenTreeCall* call);
4952 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4953 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4955 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4956 static fgWalkPreFn fgFindNonInlineCandidate;
4958 GenTree* fgOptimizeDelegateConstructor(GenTreeCall* call,
4959 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4960 CORINFO_RESOLVED_TOKEN* ldftnToken);
4961 GenTree* fgMorphLeaf(GenTree* tree);
4962 void fgAssignSetVarDef(GenTree* tree);
4963 GenTree* fgMorphOneAsgBlockOp(GenTree* tree);
4964 GenTree* fgMorphInitBlock(GenTree* tree);
4965 GenTree* fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4966 GenTree* fgMorphGetStructAddr(GenTree** pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4967 GenTree* fgMorphBlkNode(GenTree* tree, bool isDest);
4968 GenTree* fgMorphBlockOperand(GenTree* tree, var_types asgType, unsigned blockWidth, bool isDest);
4969 void fgMorphUnsafeBlk(GenTreeObj* obj);
4970 GenTree* fgMorphCopyBlock(GenTree* tree);
4971 GenTree* fgMorphForRegisterFP(GenTree* tree);
4972 GenTree* fgMorphSmpOp(GenTree* tree, MorphAddrContext* mac = nullptr);
4973 GenTree* fgMorphSmpOpPre(GenTree* tree);
4974 GenTree* fgMorphModToSubMulDiv(GenTreeOp* tree);
4975 GenTree* fgMorphSmpOpOptional(GenTreeOp* tree);
4976 GenTree* fgMorphRecognizeBoxNullable(GenTree* compare);
4978 GenTree* fgMorphToEmulatedFP(GenTree* tree);
4979 GenTree* fgMorphConst(GenTree* tree);
4982 GenTree* fgMorphTree(GenTree* tree, MorphAddrContext* mac = nullptr);
4985 #if LOCAL_ASSERTION_PROP
4986 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTree* tree));
4987 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTree* tree));
4989 void fgMorphTreeDone(GenTree* tree, GenTree* oldTree = nullptr DEBUGARG(int morphNum = 0));
4991 GenTreeStmt* fgMorphStmt;
4993 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4994 // used when morphing big offset.
4996 //----------------------- Liveness analysis -------------------------------
4998 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4999 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
5001 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
5002 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
5003 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
5005 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
5007 void fgMarkUseDef(GenTreeLclVarCommon* tree);
5009 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
5010 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
5012 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
5013 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
5015 void fgExtendDbgScopes();
5016 void fgExtendDbgLifetimes();
5019 void fgDispDebugScopes();
5022 //-------------------------------------------------------------------------
5024 // The following keeps track of any code we've added for things like array
5025 // range checking or explicit calls to enable GC, and so on.
5030 AddCodeDsc* acdNext;
5031 BasicBlock* acdDstBlk; // block to which we jump
5033 SpecialCodeKind acdKind; // what kind of a special block is this?
5034 #if !FEATURE_FIXED_OUT_ARGS
5035 bool acdStkLvlInit; // has acdStkLvl value been already set?
5037 #endif // !FEATURE_FIXED_OUT_ARGS
5041 static unsigned acdHelper(SpecialCodeKind codeKind);
5043 AddCodeDsc* fgAddCodeList;
5045 bool fgRngChkThrowAdded;
5046 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
5048 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
5050 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
5053 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
5055 bool fgUseThrowHelperBlocks();
5057 AddCodeDsc* fgGetAdditionalCodeDescriptors()
5059 return fgAddCodeList;
5063 bool fgIsCodeAdded();
5065 bool fgIsThrowHlpBlk(BasicBlock* block);
5067 #if !FEATURE_FIXED_OUT_ARGS
5068 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
5069 #endif // !FEATURE_FIXED_OUT_ARGS
5071 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
5073 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
5074 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
5075 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
5076 GenTree* fgInlinePrependStatements(InlineInfo* inlineInfo);
5077 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTree* stmt);
5079 #if FEATURE_MULTIREG_RET
5080 GenTree* fgGetStructAsStructPtr(GenTree* tree);
5081 GenTree* fgAssignStructInlineeToVar(GenTree* child, CORINFO_CLASS_HANDLE retClsHnd);
5082 void fgAttachStructInlineeToAsg(GenTree* tree, GenTree* child, CORINFO_CLASS_HANDLE retClsHnd);
5083 #endif // FEATURE_MULTIREG_RET
5085 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
5088 static fgWalkPreFn fgDebugCheckInlineCandidates;
5090 void CheckNoFatPointerCandidatesLeft();
5091 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
5094 void fgPromoteStructs();
5095 fgWalkResult fgMorphStructField(GenTree* tree, fgWalkData* fgWalkPre);
5096 fgWalkResult fgMorphLocalField(GenTree* tree, fgWalkData* fgWalkPre);
5098 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
5099 void fgMarkImplicitByRefArgs();
5101 // Change implicit byrefs' types from struct to pointer, and for any that were
5102 // promoted, create new promoted struct temps.
5103 void fgRetypeImplicitByRefArgs();
5105 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
5106 bool fgMorphImplicitByRefArgs(GenTree* tree);
5107 GenTree* fgMorphImplicitByRefArgs(GenTree* tree, bool isAddr);
5109 // Clear up annotations for any struct promotion temps created for implicit byrefs.
5110 void fgMarkDemotedImplicitByRefArgs();
5112 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
5113 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
5114 void fgMarkAddressExposedLocals();
5115 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
5117 static fgWalkPreFn fgUpdateSideEffectsPre;
5118 static fgWalkPostFn fgUpdateSideEffectsPost;
5120 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
5122 bool fgFitsInOrNotLoc(GenTree* tree, unsigned width);
5124 // The given local variable, required to be a struct variable, is being assigned via
5125 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
5126 // the variable is not enregistered, and is therefore not promoted independently.
5127 void fgLclFldAssign(unsigned lclNum);
5129 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
5131 enum TypeProducerKind
5133 TPK_Unknown = 0, // May not be a RuntimeType
5134 TPK_Handle = 1, // RuntimeType via handle
5135 TPK_GetType = 2, // RuntimeType via Object.get_Type()
5136 TPK_Null = 3, // Tree value is null
5137 TPK_Other = 4 // RuntimeType via other means
5140 TypeProducerKind gtGetTypeProducerKind(GenTree* tree);
5141 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
5142 bool gtIsActiveCSE_Candidate(GenTree* tree);
5145 bool fgPrintInlinedMethods;
5148 bool fgIsBigOffset(size_t offset);
5150 #if defined(LEGACY_BACKEND)
5151 // The following are used when morphing special cases of integer div/mod operations and also by codegen
5152 bool fgIsSignedDivOptimizable(GenTree* divisor);
5153 bool fgIsUnsignedDivOptimizable(GenTree* divisor);
5154 bool fgIsSignedModOptimizable(GenTree* divisor);
5155 bool fgIsUnsignedModOptimizable(GenTree* divisor);
5156 #endif // LEGACY_BACKEND
5158 bool fgNeedReturnSpillTemp();
5161 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5162 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5166 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5167 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5174 LclVarDsc* optIsTrackedLocal(GenTree* tree);
5177 void optRemoveRangeCheck(GenTree* tree, GenTree* stmt);
5178 bool optIsRangeCheckRemovable(GenTree* tree);
5181 static fgWalkPreFn optValidRangeCheckIndex;
5182 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
5185 void optRemoveTree(GenTree* deadTree, GenTree* keepList);
5187 /**************************************************************************
5189 *************************************************************************/
5192 // Do hoisting for all loops.
5193 void optHoistLoopCode();
5195 // To represent sets of VN's that have already been hoisted in outer loops.
5196 typedef JitHashTable<ValueNum, JitSmallPrimitiveKeyFuncs<ValueNum>, bool> VNToBoolMap;
5197 typedef VNToBoolMap VNSet;
5199 struct LoopHoistContext
5202 // The set of variables hoisted in the current loop (or nullptr if there are none).
5203 VNSet* m_pHoistedInCurLoop;
5206 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
5207 VNSet m_hoistedInParentLoops;
5208 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
5209 // Previous decisions on loop-invariance of value numbers in the current loop.
5210 VNToBoolMap m_curLoopVnInvariantCache;
5212 VNSet* GetHoistedInCurLoop(Compiler* comp)
5214 if (m_pHoistedInCurLoop == nullptr)
5216 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
5218 return m_pHoistedInCurLoop;
5221 VNSet* ExtractHoistedInCurLoop()
5223 VNSet* res = m_pHoistedInCurLoop;
5224 m_pHoistedInCurLoop = nullptr;
5228 LoopHoistContext(Compiler* comp)
5229 : m_pHoistedInCurLoop(nullptr)
5230 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5231 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5236 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5237 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5238 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5239 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5241 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5242 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5243 // "m_hoistedInParentLoops".
5245 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5247 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5248 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5249 // expressions to "hoistInLoop".
5250 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5252 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5253 bool optIsProfitableToHoistableTree(GenTree* tree, unsigned lnum);
5255 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5256 // that are invariant in loop "lnum" (an index into the optLoopTable)
5257 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5258 // expressions to "hoistInLoop".
5259 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5260 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5261 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5262 bool optHoistLoopExprsForTree(GenTree* tree,
5264 LoopHoistContext* hoistCtxt,
5265 bool* firstBlockAndBeforeSideEffect,
5267 bool* pCctorDependent);
5269 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5270 void optHoistCandidate(GenTree* tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5272 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5273 // Constants and init values are always loop invariant.
5274 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5275 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5277 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5278 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5279 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5280 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5281 bool optTreeIsValidAtLoopHead(GenTree* tree, unsigned lnum);
5283 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5284 // in the loop table.
5285 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5287 // Records the set of "side effects" of all loops: fields (object instance and static)
5288 // written to, and SZ-array element type equivalence classes updated.
5289 void optComputeLoopSideEffects();
5292 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5293 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5294 // static) written to, and SZ-array element type equivalence classes updated.
5295 void optComputeLoopNestSideEffects(unsigned lnum);
5297 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5298 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5300 // Hoist the expression "expr" out of loop "lnum".
5301 void optPerformHoistExpr(GenTree* expr, unsigned lnum);
5304 void optOptimizeBools();
5307 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5309 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5312 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5314 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5315 // the loop into a "do-while" loop
5316 // Also finds all natural loops and records them in the loop table
5318 // Optionally clone loops in the loop table.
5319 void optCloneLoops();
5321 // Clone loop "loopInd" in the loop table.
5322 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5324 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5325 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5326 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5328 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5330 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5333 // This enumeration describes what is killed by a call.
5337 CALLINT_NONE, // no interference (most helpers)
5338 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5339 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5340 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5341 CALLINT_ALL, // kills everything (normal method call)
5345 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5346 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5347 // in bbNext order; we use comparisons on the bbNum to decide order.)
5348 // The blocks that define the body are
5349 // first <= top <= entry <= bottom .
5350 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5351 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5352 // Compiler::optFindNaturalLoops().
5355 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5356 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5357 // loop, but not the outer loop.)
5358 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5360 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5361 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5362 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5364 callInterf lpAsgCall; // "callInterf" for calls in the loop
5365 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5366 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5368 unsigned short lpFlags; // Mask of the LPFLG_* constants
5370 unsigned char lpExitCnt; // number of exits from the loop
5372 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5373 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5374 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5375 // (Actually, an "immediately" nested loop --
5376 // no other child of this loop is a parent of lpChild.)
5377 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5378 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5379 // by following "lpChild" then "lpSibling" links.
5381 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5382 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5384 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5385 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5386 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5388 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5389 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5391 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5392 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5393 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5394 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5396 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5397 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5398 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5400 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5401 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5402 // type are assigned to.
5404 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5405 // memory side effects. If this is set, the fields below
5406 // may not be accurate (since they become irrelevant.)
5407 bool lpContainsCall; // True if executing the loop body *may* execute a call
5409 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5410 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5412 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5414 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5415 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5417 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5419 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5420 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5422 typedef JitHashTable<CORINFO_FIELD_HANDLE, JitPtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>, bool> FieldHandleSet;
5423 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5424 // instance fields modified
5427 typedef JitHashTable<CORINFO_CLASS_HANDLE, JitPtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>, bool> ClassHandleSet;
5428 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5429 // arrays of that type are modified
5432 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5433 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5435 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5436 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5437 // (shifted left, with a low-order bit set to distinguish.)
5438 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5439 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5441 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5443 GenTree* lpIterTree; // The "i <op>= const" tree
5444 unsigned lpIterVar(); // iterator variable #
5445 int lpIterConst(); // the constant with which the iterator is incremented
5446 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5447 void VERIFY_lpIterTree();
5449 var_types lpIterOperType(); // For overflow instructions
5452 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5453 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5457 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5459 GenTree* lpTestTree; // pointer to the node containing the loop test
5460 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5461 void VERIFY_lpTestTree();
5463 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5464 GenTree* lpIterator(); // the iterator node in the loop test
5465 GenTree* lpLimit(); // the limit node in the loop test
5467 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5468 // LPFLG_CONST_LIMIT
5469 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5471 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5472 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5473 // LPFLG_ARRLEN_LIMIT
5475 // Returns "true" iff "*this" contains the blk.
5476 bool lpContains(BasicBlock* blk)
5478 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5480 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5481 // to be equal, but requiring bottoms to be different.)
5482 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5484 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5487 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5488 // bottoms to be different.)
5489 bool lpContains(const LoopDsc& lp2)
5491 return lpContains(lp2.lpFirst, lp2.lpBottom);
5494 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5495 // (allowing firsts to be equal, but requiring bottoms to be different.)
5496 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5498 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5501 // Returns "true" iff "*this" is (properly) contained by "lp2"
5502 // (allowing firsts to be equal, but requiring bottoms to be different.)
5503 bool lpContainedBy(const LoopDsc& lp2)
5505 return lpContains(lp2.lpFirst, lp2.lpBottom);
5508 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5509 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5511 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5513 // Returns "true" iff "*this" is disjoint from "lp2".
5514 bool lpDisjoint(const LoopDsc& lp2)
5516 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5518 // Returns "true" iff the loop is well-formed (see code for defn).
5521 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5522 lpEntry->bbNum <= lpBottom->bbNum &&
5523 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5528 bool fgMightHaveLoop(); // returns true if there are any backedges
5529 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5532 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5533 unsigned char optLoopCount; // number of tracked loops
5535 bool optRecordLoop(BasicBlock* head,
5541 unsigned char exitCnt);
5544 unsigned optCallCount; // number of calls made in the method
5545 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5546 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5547 unsigned optLoopsCloned; // number of loops cloned in the current method.
5550 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5551 void optPrintLoopInfo(unsigned loopNum,
5553 BasicBlock* lpFirst,
5555 BasicBlock* lpEntry,
5556 BasicBlock* lpBottom,
5557 unsigned char lpExitCnt,
5559 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5560 void optPrintLoopInfo(unsigned lnum);
5561 void optPrintLoopRecording(unsigned lnum);
5563 void optCheckPreds();
5566 void optSetBlockWeights();
5568 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5570 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5572 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5574 bool optIsLoopTestEvalIntoTemp(GenTree* test, GenTree** newTest);
5575 unsigned optIsLoopIncrTree(GenTree* incr);
5576 bool optCheckIterInLoopTest(unsigned loopInd, GenTree* test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5577 bool optComputeIterInfo(GenTree* incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5578 bool optPopulateInitInfo(unsigned loopInd, GenTree* init, unsigned iterVar);
5579 bool optExtractInitTestIncr(
5580 BasicBlock* head, BasicBlock* bottom, BasicBlock* exit, GenTree** ppInit, GenTree** ppTest, GenTree** ppIncr);
5582 void optFindNaturalLoops();
5584 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5585 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5586 bool optCanonicalizeLoopNest(unsigned char loopInd);
5588 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5589 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5590 bool optCanonicalizeLoop(unsigned char loopInd);
5592 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5593 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5594 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5595 bool optLoopContains(unsigned l1, unsigned l2);
5597 // Requires "loopInd" to be a valid index into the loop table.
5598 // Updates the loop table by changing loop "loopInd", whose head is required
5599 // to be "from", to be "to". Also performs this transformation for any
5600 // loop nested in "loopInd" that shares the same head as "loopInd".
5601 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5603 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5604 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5605 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5607 // Marks the containsCall information to "lnum" and any parent loops.
5608 void AddContainsCallAllContainingLoops(unsigned lnum);
5609 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5610 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5611 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5612 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5613 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5614 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5616 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5617 // of "from".) Copies the jump destination from "from" to "to".
5618 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5620 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5621 unsigned optLoopDepth(unsigned lnum)
5623 unsigned par = optLoopTable[lnum].lpParent;
5624 if (par == BasicBlock::NOT_IN_LOOP)
5630 return 1 + optLoopDepth(par);
5634 void fgOptWhileLoop(BasicBlock* block);
5636 bool optComputeLoopRep(int constInit,
5639 genTreeOps iterOper,
5641 genTreeOps testOper,
5644 unsigned* iterCount);
5645 #if FEATURE_STACK_FP_X87
5648 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5649 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5650 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5651 #endif // FEATURE_STACK_FP_X87
5654 static fgWalkPreFn optIsVarAssgCB;
5657 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTree* skip, unsigned var);
5659 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5661 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5663 bool optNarrowTree(GenTree* tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5665 /**************************************************************************
5666 * Optimization conditions
5667 *************************************************************************/
5669 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5670 bool optPentium4(void);
5671 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5672 bool optAvoidIntMult(void);
5677 // The following is the upper limit on how many expressions we'll keep track
5678 // of for the CSE analysis.
5680 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5682 static const int MIN_CSE_COST = 2;
5684 // Keeps tracked cse indices
5685 BitVecTraits* cseTraits;
5688 /* Generic list of nodes - used by the CSE logic */
5698 treeStmtLst* tslNext;
5699 GenTree* tslTree; // tree node
5700 GenTree* tslStmt; // statement containing the tree
5701 BasicBlock* tslBlock; // block containing the statement
5704 // The following logic keeps track of expressions via a simple hash table.
5708 CSEdsc* csdNextInBucket; // used by the hash table
5710 unsigned csdHashValue; // the orginal hashkey
5712 unsigned csdIndex; // 1..optCSECandidateCount
5713 char csdLiveAcrossCall; // 0 or 1
5715 unsigned short csdDefCount; // definition count
5716 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5718 unsigned csdDefWtCnt; // weighted def count
5719 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5721 GenTree* csdTree; // treenode containing the 1st occurance
5722 GenTree* csdStmt; // stmt containing the 1st occurance
5723 BasicBlock* csdBlock; // block containing the 1st occurance
5725 treeStmtLst* csdTreeList; // list of matching tree nodes: head
5726 treeStmtLst* csdTreeLast; // list of matching tree nodes: tail
5728 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5729 // number, this will reflect it; otherwise, NoVN.
5732 static const size_t s_optCSEhashSize;
5733 CSEdsc** optCSEhash;
5736 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, GenTree*> NodeToNodeMap;
5738 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5739 // re-numbered with the bound to improve range check elimination
5741 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5742 void optCseUpdateCheckedBoundMap(GenTree* compare);
5746 CSEdsc* optCSEfindDsc(unsigned index);
5747 bool optUnmarkCSE(GenTree* tree);
5749 // user defined callback data for the tree walk function optCSE_MaskHelper()
5750 struct optCSE_MaskData
5752 EXPSET_TP CSE_defMask;
5753 EXPSET_TP CSE_useMask;
5756 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5757 static fgWalkPreFn optCSE_MaskHelper;
5759 // This function walks all the node for an given tree
5760 // and return the mask of CSE definitions and uses for the tree
5762 void optCSE_GetMaskData(GenTree* tree, optCSE_MaskData* pMaskData);
5764 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5765 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5766 bool optCSE_canSwap(GenTree* tree);
5768 static fgWalkPostFn optPropagateNonCSE;
5769 static fgWalkPreFn optHasNonCSEChild;
5771 static fgWalkPreFn optUnmarkCSEs;
5772 static fgWalkPreFn optHasCSEdefWithSideeffect;
5774 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5775 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5777 void optCleanupCSEs();
5780 void optEnsureClearCSEInfo();
5783 #endif // FEATURE_ANYCSE
5785 #if FEATURE_VALNUM_CSE
5786 /**************************************************************************
5787 * Value Number based CSEs
5788 *************************************************************************/
5791 void optOptimizeValnumCSEs();
5794 void optValnumCSE_Init();
5795 unsigned optValnumCSE_Index(GenTree* tree, GenTree* stmt);
5796 unsigned optValnumCSE_Locate();
5797 void optValnumCSE_InitDataFlow();
5798 void optValnumCSE_DataFlow();
5799 void optValnumCSE_Availablity();
5800 void optValnumCSE_Heuristic();
5801 bool optValnumCSE_UnmarkCSEs(GenTree* deadTree, GenTree** wbKeepList);
5803 #endif // FEATURE_VALNUM_CSE
5806 bool optDoCSE; // True when we have found a duplicate CSE tree
5807 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5808 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5809 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5810 unsigned optCSEstart; // The first local variable number that is a CSE
5811 unsigned optCSEcount; // The total count of CSE's introduced.
5812 unsigned optCSEweight; // The weight of the current block when we are
5813 // scanning for CSE expressions
5815 bool optIsCSEcandidate(GenTree* tree);
5817 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5819 bool lclNumIsTrueCSE(unsigned lclNum) const
5821 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5824 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5826 bool lclNumIsCSE(unsigned lclNum) const
5828 return lvaTable[lclNum].lvIsCSE;
5832 bool optConfigDisableCSE();
5833 bool optConfigDisableCSE2();
5835 void optOptimizeCSEs();
5837 #endif // FEATURE_ANYCSE
5845 unsigned ivaVar; // Variable we are interested in, or -1
5846 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5847 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5848 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5849 callInterf ivaMaskCall; // What kind of calls are there?
5852 static callInterf optCallInterf(GenTreeCall* call);
5855 // VN based copy propagation.
5856 typedef ArrayStack<GenTree*> GenTreePtrStack;
5857 typedef JitHashTable<unsigned, JitSmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*> LclNumToGenTreePtrStack;
5859 // Kill set to track variables with intervening definitions.
5860 VARSET_TP optCopyPropKillSet;
5862 // Copy propagation functions.
5863 void optCopyProp(BasicBlock* block, GenTree* stmt, GenTree* tree, LclNumToGenTreePtrStack* curSsaName);
5864 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5865 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5866 bool optIsSsaLocal(GenTree* tree);
5867 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5868 void optVnCopyProp();
5869 INDEBUG(void optDumpCopyPropStack(LclNumToGenTreePtrStack* curSsaName));
5871 /**************************************************************************
5872 * Early value propagation
5873 *************************************************************************/
5879 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5883 static unsigned GetHashCode(SSAName ssaNm)
5885 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5888 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5890 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5894 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5895 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5896 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5897 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5898 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5899 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5901 bool doesMethodHaveFatPointer()
5903 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5906 void setMethodHasFatPointer()
5908 optMethodFlags |= OMF_HAS_FATPOINTER;
5911 void clearMethodHasFatPointer()
5913 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5916 void addFatPointerCandidate(GenTreeCall* call);
5918 unsigned optMethodFlags;
5920 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5921 // No throughput diff was found with backward walk bound between 3-8.
5922 static const int optEarlyPropRecurBound = 5;
5924 enum class optPropKind
5932 bool gtIsVtableRef(GenTree* tree);
5933 GenTree* getArrayLengthFromAllocation(GenTree* tree);
5934 GenTree* getObjectHandleNodeFromAllocation(GenTree* tree);
5935 GenTree* optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5936 GenTree* optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5937 GenTree* optEarlyPropRewriteTree(GenTree* tree);
5938 bool optDoEarlyPropForBlock(BasicBlock* block);
5939 bool optDoEarlyPropForFunc();
5940 void optEarlyProp();
5941 void optFoldNullCheck(GenTree* tree);
5942 bool optCanMoveNullCheckPastTree(GenTree* tree, bool isInsideTry);
5945 /**************************************************************************
5946 * Value/Assertion propagation
5947 *************************************************************************/
5949 // Data structures for assertion prop
5950 BitVecTraits* apTraits;
5953 enum optAssertionKind
5970 O1K_CONSTANT_LOOP_BND,
5991 optAssertionKind assertionKind;
5994 unsigned lclNum; // assigned to or property of this local var number
6002 struct AssertionDscOp1
6004 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
6011 struct AssertionDscOp2
6013 optOp2Kind kind; // a const or copy assignment
6017 ssize_t iconVal; // integer
6018 unsigned iconFlags; // gtFlags
6020 struct Range // integer subrange
6034 bool IsCheckedBoundArithBound()
6036 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
6038 bool IsCheckedBoundBound()
6040 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
6042 bool IsConstantBound()
6044 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
6045 op1.kind == O1K_CONSTANT_LOOP_BND);
6047 bool IsBoundsCheckNoThrow()
6049 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
6052 bool IsCopyAssertion()
6054 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
6057 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
6059 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
6060 a1->op2.kind == a2->op2.kind;
6063 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
6065 if (kind == OAK_EQUAL)
6067 return kind2 == OAK_NOT_EQUAL;
6069 else if (kind == OAK_NOT_EQUAL)
6071 return kind2 == OAK_EQUAL;
6076 static ssize_t GetLowerBoundForIntegralType(var_types type)
6095 static ssize_t GetUpperBoundForIntegralType(var_types type)
6118 bool HasSameOp1(AssertionDsc* that, bool vnBased)
6120 if (op1.kind != that->op1.kind)
6124 else if (op1.kind == O1K_ARR_BND)
6127 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
6131 return ((vnBased && (op1.vn == that->op1.vn)) ||
6132 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
6136 bool HasSameOp2(AssertionDsc* that, bool vnBased)
6138 if (op2.kind != that->op2.kind)
6144 case O2K_IND_CNS_INT:
6146 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
6148 case O2K_CONST_LONG:
6149 return (op2.lconVal == that->op2.lconVal);
6151 case O2K_CONST_DOUBLE:
6152 // exact match because of positive and negative zero.
6153 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
6155 case O2K_LCLVAR_COPY:
6157 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
6158 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
6161 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
6164 // we will return false
6168 assert(!"Unexpected value for op2.kind in AssertionDsc.");
6174 bool Complementary(AssertionDsc* that, bool vnBased)
6176 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
6177 HasSameOp2(that, vnBased);
6180 bool Equals(AssertionDsc* that, bool vnBased)
6182 if (assertionKind != that->assertionKind)
6186 else if (assertionKind == OAK_NO_THROW)
6188 assert(op2.kind == O2K_INVALID);
6189 return HasSameOp1(that, vnBased);
6193 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
6199 static fgWalkPreFn optAddCopiesCallback;
6200 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
6201 unsigned optAddCopyLclNum;
6202 GenTree* optAddCopyAsgnNode;
6204 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6205 bool optAssertionPropagated; // set to true if we modified the trees
6206 bool optAssertionPropagatedCurrentStmt;
6208 GenTree* optAssertionPropCurrentTree;
6210 AssertionIndex* optComplementaryAssertionMap;
6211 JitExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6212 // using the value of a local var) for each local var
6213 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6214 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6215 AssertionIndex optMaxAssertionCount;
6218 void optVnNonNullPropCurStmt(BasicBlock* block, GenTree* stmt, GenTree* tree);
6219 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTree* stmt, GenTree* tree);
6220 GenTree* optVNConstantPropOnRelOp(GenTree* tree);
6221 GenTree* optVNConstantPropOnJTrue(BasicBlock* block, GenTree* stmt, GenTree* test);
6222 GenTree* optVNConstantPropOnTree(BasicBlock* block, GenTree* stmt, GenTree* tree);
6223 GenTree* optPrepareTreeForReplacement(GenTree* extractTree, GenTree* replaceTree);
6225 AssertionIndex GetAssertionCount()
6227 return optAssertionCount;
6229 ASSERT_TP* bbJtrueAssertionOut;
6230 typedef JitHashTable<ValueNum, JitSmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP> ValueNumToAssertsMap;
6231 ValueNumToAssertsMap* optValueNumToAsserts;
6233 // Assertion prop helpers.
6234 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6235 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6236 void optAssertionInit(bool isLocalProp);
6237 void optAssertionTraitsInit(AssertionIndex assertionCount);
6238 #if LOCAL_ASSERTION_PROP
6239 void optAssertionReset(AssertionIndex limit);
6240 void optAssertionRemove(AssertionIndex index);
6243 // Assertion prop data flow functions.
6244 void optAssertionPropMain();
6245 GenTree* optVNAssertionPropCurStmt(BasicBlock* block, GenTree* stmt);
6246 bool optIsTreeKnownIntValue(bool vnBased, GenTree* tree, ssize_t* pConstant, unsigned* pIconFlags);
6247 ASSERT_TP* optInitAssertionDataflowFlags();
6248 ASSERT_TP* optComputeAssertionGen();
6250 // Assertion Gen functions.
6251 void optAssertionGen(GenTree* tree);
6252 AssertionIndex optAssertionGenPhiDefn(GenTree* tree);
6253 AssertionInfo optCreateJTrueBoundsAssertion(GenTree* tree);
6254 AssertionInfo optAssertionGenJtrue(GenTree* tree);
6255 AssertionIndex optCreateJtrueAssertions(GenTree* op1, GenTree* op2, Compiler::optAssertionKind assertionKind);
6256 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6257 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6259 // Assertion creation functions.
6260 AssertionIndex optCreateAssertion(GenTree* op1, GenTree* op2, optAssertionKind assertionKind);
6261 AssertionIndex optCreateAssertion(GenTree* op1,
6263 optAssertionKind assertionKind,
6264 AssertionDsc* assertion);
6265 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTree* op1, GenTree* op2);
6267 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6268 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6269 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6271 void optPrintVnAssertionMapping();
6273 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6275 // Used for respective assertion propagations.
6276 AssertionIndex optAssertionIsSubrange(GenTree* tree, var_types toType, ASSERT_VALARG_TP assertions);
6277 AssertionIndex optAssertionIsSubtype(GenTree* tree, GenTree* methodTableArg, ASSERT_VALARG_TP assertions);
6278 AssertionIndex optAssertionIsNonNullInternal(GenTree* op, ASSERT_VALARG_TP assertions);
6279 bool optAssertionIsNonNull(GenTree* op,
6280 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6282 // Used for Relop propagation.
6283 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTree* op1, GenTree* op2);
6284 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6285 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6287 // Assertion prop for lcl var functions.
6288 bool optAssertionProp_LclVarTypeCheck(GenTree* tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6289 GenTree* optCopyAssertionProp(AssertionDsc* curAssertion,
6291 GenTree* stmt DEBUGARG(AssertionIndex index));
6292 GenTree* optConstantAssertionProp(AssertionDsc* curAssertion,
6294 GenTree* stmt DEBUGARG(AssertionIndex index));
6295 GenTree* optVnConstantAssertionProp(GenTree* tree, GenTree* stmt);
6297 // Assertion propagation functions.
6298 GenTree* optAssertionProp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6299 GenTree* optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6300 GenTree* optAssertionProp_Ind(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6301 GenTree* optAssertionProp_Cast(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6302 GenTree* optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, GenTree* stmt);
6303 GenTree* optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6304 GenTree* optAssertionProp_Comma(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6305 GenTree* optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6306 GenTree* optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6307 GenTree* optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6308 GenTree* optAssertionProp_Update(GenTree* newTree, GenTree* tree, GenTree* stmt);
6309 GenTree* optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, GenTree* stmt);
6311 // Implied assertion functions.
6312 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6313 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6314 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6315 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6318 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6319 void optDebugCheckAssertion(AssertionDsc* assertion);
6320 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6322 void optAddCopies();
6323 #endif // ASSERTION_PROP
6325 /**************************************************************************
6327 *************************************************************************/
6330 struct LoopCloneVisitorInfo
6332 LoopCloneContext* context;
6335 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTree* stmt)
6336 : context(context), loopNum(loopNum), stmt(nullptr)
6341 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6342 bool optExtractArrIndex(GenTree* tree, ArrIndex* result, unsigned lhsNum);
6343 bool optReconstructArrIndex(GenTree* tree, ArrIndex* result, unsigned lhsNum);
6344 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6345 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6346 fgWalkResult optCanOptimizeByLoopCloning(GenTree* tree, LoopCloneVisitorInfo* info);
6347 void optObtainLoopCloningOpts(LoopCloneContext* context);
6348 bool optIsLoopClonable(unsigned loopInd);
6350 bool optCanCloneLoops();
6353 void optDebugLogLoopCloning(BasicBlock* block, GenTree* insertBefore);
6355 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6356 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6357 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6358 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6362 void optInsertLoopCloningStress(BasicBlock* head);
6364 #if COUNT_RANGECHECKS
6365 static unsigned optRangeChkRmv;
6366 static unsigned optRangeChkAll;
6375 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6380 RngChkDsc* rcdNextInBucket; // used by the hash table
6382 unsigned short rcdHashValue; // to make matching faster
6383 unsigned short rcdIndex; // 0..optRngChkCount-1
6385 GenTree* rcdTree; // the array index tree
6388 unsigned optRngChkCount;
6389 static const size_t optRngChkHashSize;
6391 ssize_t optGetArrayRefScaleAndIndex(GenTree* mul, GenTree** pIndex DEBUGARG(bool bRngChk));
6392 GenTree* optFindLocalInit(BasicBlock* block, GenTree* local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6394 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6397 bool optLoopsMarked;
6400 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6401 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6405 XX Does the register allocation and puts the remaining lclVars on the stack XX
6407 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6408 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6412 #ifndef LEGACY_BACKEND
6417 #else // LEGACY_BACKEND
6422 #endif // LEGACY_BACKEND
6424 #ifdef LEGACY_BACKEND
6426 void raAssignVars(); // register allocation
6427 #endif // LEGACY_BACKEND
6429 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6431 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6433 void raMarkStkVars();
6436 // Some things are used by both LSRA and regpredict allocators.
6438 FrameType rpFrameType;
6439 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6441 #ifdef LEGACY_BACKEND
6442 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6444 #endif // LEGACY_BACKEND
6446 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6448 #if FEATURE_FP_REGALLOC
6449 enum enumConfigRegisterFP
6451 CONFIG_REGISTER_FP_NONE = 0x0,
6452 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6453 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6454 CONFIG_REGISTER_FP_FULL = 0x3,
6456 enumConfigRegisterFP raConfigRegisterFP();
6457 #endif // FEATURE_FP_REGALLOC
6460 regMaskTP raConfigRestrictMaskFP();
6463 #ifndef LEGACY_BACKEND
6464 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6465 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6466 #else // LEGACY_BACKEND
6467 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6468 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6469 bool raNewBlocks; // True is we added killing blocks for FPU registers
6470 unsigned rpPasses; // Number of passes made by the register predicter
6471 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6472 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6473 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6474 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6475 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6476 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6477 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6478 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6479 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6480 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6481 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6482 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6484 bool rpRegAllocDone; // Set to true after we have completed register allocation
6486 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6488 void raSetupArgMasks(RegState* r);
6490 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6492 void raDumpVarIntf(); // Dump the variable to variable interference graph
6493 void raDumpRegIntf(); // Dump the variable to register interference graph
6495 void raAdjustVarIntf();
6497 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6499 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6501 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6502 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6504 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6506 static fgWalkPreFn rpMarkRegIntf;
6508 regMaskTP rpPredictAddressMode(
6509 GenTree* tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTree* lenCSE);
6511 void rpPredictRefAssign(unsigned lclNum);
6513 regMaskTP rpPredictBlkAsgRegUse(GenTree* tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6515 regMaskTP rpPredictTreeRegUse(GenTree* tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6517 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6519 void rpPredictRegUse(); // Entry point
6521 unsigned raPredictTreeRegUse(GenTree* tree);
6522 unsigned raPredictListRegUse(GenTree* list);
6524 void raSetRegVarOrder(var_types regType,
6525 regNumber* customVarOrder,
6526 unsigned* customVarOrderSize,
6528 regMaskTP avoidReg);
6530 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6531 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6532 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6533 void raAddToStkPredict(unsigned val)
6535 unsigned newStkPredict = rpStkPredict + val;
6536 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6537 rpStkPredict = UINT_MAX - 1;
6539 rpStkPredict = newStkPredict;
6543 #if !FEATURE_FP_REGALLOC
6544 void raDispFPlifeInfo();
6548 regMaskTP genReturnRegForTree(GenTree* tree);
6549 #endif // LEGACY_BACKEND
6551 /* raIsVarargsStackArg is called by raMaskStkVars and by
6552 lvaSortByRefCount. It identifies the special case
6553 where a varargs function has a parameter passed on the
6554 stack, other than the special varargs handle. Such parameters
6555 require special treatment, because they cannot be tracked
6556 by the GC (their offsets in the stack are not known
6560 bool raIsVarargsStackArg(unsigned lclNum)
6564 LclVarDsc* varDsc = &lvaTable[lclNum];
6566 assert(varDsc->lvIsParam);
6568 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6570 #else // _TARGET_X86_
6574 #endif // _TARGET_X86_
6577 #ifdef LEGACY_BACKEND
6578 // Records the current prediction, if it's better than any previous recorded prediction.
6579 void rpRecordPrediction();
6580 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6581 void rpUseRecordedPredictionIfBetter();
6583 // Data members used in the methods above.
6584 unsigned rpBestRecordedStkPredict;
6585 struct VarRegPrediction
6587 bool m_isEnregistered;
6588 regNumberSmall m_regNum;
6589 regNumberSmall m_otherReg;
6591 VarRegPrediction* rpBestRecordedPrediction;
6592 #endif // LEGACY_BACKEND
6595 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6596 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6600 XX Get to the class and method info from the Execution Engine given XX
6601 XX tokens for the class and method XX
6603 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6604 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6608 /* These are the different addressing modes used to access a local var.
6609 * The JIT has to report the location of the locals back to the EE
6610 * for debugging purposes.
6616 VLT_REG_BYREF, // this type is currently only used for value types on X64
6619 VLT_STK_BYREF, // this type is currently only used for value types on X64
6633 siVarLocType vlType;
6636 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6638 // VLT_REG_BYREF -- the specified register contains the address of the variable
6646 // VLT_STK -- Any 32 bit value which is on the stack
6647 // eg. [ESP+0x20], or [EBP-0x28]
6648 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6649 // eg. mov EAX, [ESP+0x20]; [EAX]
6653 regNumber vlsBaseReg;
6654 NATIVE_OFFSET vlsOffset;
6657 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6666 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6667 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6675 regNumber vlrssBaseReg;
6676 NATIVE_OFFSET vlrssOffset;
6680 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6681 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6687 regNumber vlsrsBaseReg;
6688 NATIVE_OFFSET vlsrsOffset;
6694 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6695 // eg 2 DWords at [ESP+0x10]
6699 regNumber vls2BaseReg;
6700 NATIVE_OFFSET vls2Offset;
6703 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6704 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6711 // VLT_FIXED_VA -- fixed argument of a varargs function.
6712 // The argument location depends on the size of the variable
6713 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6714 // location of the first arg. This argument can then be accessed
6715 // relative to the position of the first arg
6719 unsigned vlfvOffset;
6726 void* rpValue; // pointer to the in-process
6727 // location of the value.
6733 bool vlIsInReg(regNumber reg);
6734 bool vlIsOnStk(regNumber reg, signed offset);
6737 /*************************************************************************/
6742 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6743 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6744 CORINFO_CALLINFO_FLAGS flags,
6745 CORINFO_CALL_INFO* pResult);
6746 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6748 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6749 CORINFO_ACCESS_FLAGS flags,
6750 CORINFO_FIELD_INFO* pResult);
6754 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6756 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6758 bool IsSuperPMIException(unsigned code)
6760 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6762 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6763 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6764 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6765 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6766 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6767 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6768 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6769 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6773 case EXCEPTIONCODE_DebugBreakorAV:
6774 case EXCEPTIONCODE_MC:
6775 case EXCEPTIONCODE_LWM:
6776 case EXCEPTIONCODE_SASM:
6777 case EXCEPTIONCODE_SSYM:
6778 case EXCEPTIONCODE_CALLUTILS:
6779 case EXCEPTIONCODE_TYPEUTILS:
6780 case EXCEPTIONCODE_ASSERT:
6787 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6788 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6790 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6791 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6794 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6795 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6796 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6798 // VOM info, method sigs
6800 void eeGetSig(unsigned sigTok,
6801 CORINFO_MODULE_HANDLE scope,
6802 CORINFO_CONTEXT_HANDLE context,
6803 CORINFO_SIG_INFO* retSig);
6805 void eeGetCallSiteSig(unsigned sigTok,
6806 CORINFO_MODULE_HANDLE scope,
6807 CORINFO_CONTEXT_HANDLE context,
6808 CORINFO_SIG_INFO* retSig);
6810 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6812 // Method entry-points, instrs
6814 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6816 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6818 CORINFO_EE_INFO eeInfo;
6819 bool eeInfoInitialized;
6821 CORINFO_EE_INFO* eeGetEEInfo();
6823 // Gets the offset of a SDArray's first element
6824 unsigned eeGetArrayDataOffset(var_types type);
6825 // Gets the offset of a MDArray's first element
6826 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6828 GenTree* eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6830 // Returns the page size for the target machine as reported by the EE.
6831 inline size_t eeGetPageSize()
6833 return eeGetEEInfo()->osPageSize;
6836 // Returns the frame size at which we will generate a loop to probe the stack.
6837 inline size_t getVeryLargeFrameSize()
6840 // The looping probe code is 40 bytes, whereas the straight-line probing for
6841 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6842 // or greater, to generate smaller code.
6843 return 2 * eeGetPageSize();
6845 return 3 * eeGetPageSize();
6849 //------------------------------------------------------------------------
6850 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6852 // It represents Abi and target specific registers for the parameter.
6854 class VirtualStubParamInfo
6857 VirtualStubParamInfo(bool isCoreRTABI)
6859 #if defined(_TARGET_X86_)
6862 #elif defined(_TARGET_AMD64_)
6873 #elif defined(_TARGET_ARM_)
6884 #elif defined(_TARGET_ARM64_)
6888 #error Unsupported or unset target architecture
6891 #ifdef LEGACY_BACKEND
6892 #if defined(_TARGET_X86_)
6893 predict = PREDICT_REG_EAX;
6894 #elif defined(_TARGET_ARM_)
6895 predict = PREDICT_REG_R4;
6897 #error Unsupported or unset target architecture
6899 #endif // LEGACY_BACKEND
6902 regNumber GetReg() const
6907 _regMask_enum GetRegMask() const
6912 #ifdef LEGACY_BACKEND
6913 rpPredictReg GetPredict() const
6921 _regMask_enum regMask;
6923 #ifdef LEGACY_BACKEND
6924 rpPredictReg predict;
6928 VirtualStubParamInfo* virtualStubParamInfo;
6930 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6932 return eeGetEEInfo()->targetAbi == abi;
6935 inline bool generateCFIUnwindCodes()
6937 #if defined(_TARGET_UNIX_)
6938 return IsTargetAbi(CORINFO_CORERT_ABI);
6946 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6948 // Debugging support - Line number info
6950 void eeGetStmtOffsets();
6952 unsigned eeBoundariesCount;
6954 struct boundariesDsc
6956 UNATIVE_OFFSET nativeIP;
6958 unsigned sourceReason;
6959 } * eeBoundaries; // Boundaries to report to EE
6960 void eeSetLIcount(unsigned count);
6961 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6965 static void eeDispILOffs(IL_OFFSET offs);
6966 static void eeDispLineInfo(const boundariesDsc* line);
6967 void eeDispLineInfos();
6970 // Debugging support - Local var info
6974 unsigned eeVarsCount;
6976 struct VarResultInfo
6978 UNATIVE_OFFSET startOffset;
6979 UNATIVE_OFFSET endOffset;
6983 void eeSetLVcount(unsigned count);
6984 void eeSetLVinfo(unsigned which,
6985 UNATIVE_OFFSET startOffs,
6986 UNATIVE_OFFSET length,
6991 const siVarLoc& loc);
6995 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6996 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6999 // ICorJitInfo wrappers
7001 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
7003 void eeAllocUnwindInfo(BYTE* pHotCode,
7009 CorJitFuncKind funcKind);
7011 void eeSetEHcount(unsigned cEH);
7013 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
7015 WORD eeGetRelocTypeHint(void* target);
7017 // ICorStaticInfo wrapper functions
7019 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
7021 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
7023 static void dumpSystemVClassificationType(SystemVClassificationType ct);
7026 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
7027 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
7028 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
7029 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
7031 template <typename ParamType>
7032 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
7034 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
7037 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
7039 // Utility functions
7041 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
7044 const wchar_t* eeGetCPString(size_t stringHandle);
7047 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
7049 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
7050 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
7052 static fgWalkPreFn CountSharedStaticHelper;
7053 static bool IsSharedStaticHelper(GenTree* tree);
7054 static bool IsTreeAlwaysHoistable(GenTree* tree);
7055 static bool IsGcSafePoint(GenTree* tree);
7057 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
7058 // returns true/false if 'field' is a Jit Data offset
7059 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
7060 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
7061 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
7063 /*****************************************************************************/
7068 enum TEMP_USAGE_TYPE
7074 static var_types tmpNormalizeType(var_types type);
7075 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
7076 void tmpRlsTemp(TempDsc* temp);
7077 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
7080 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
7081 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
7085 bool tmpAllFree() const;
7088 #ifndef LEGACY_BACKEND
7089 void tmpPreAllocateTemps(var_types type, unsigned count);
7090 #endif // !LEGACY_BACKEND
7093 #ifdef LEGACY_BACKEND
7094 unsigned tmpIntSpillMax; // number of int-sized spill temps
7095 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
7096 #endif // LEGACY_BACKEND
7098 unsigned tmpCount; // Number of temps
7099 unsigned tmpSize; // Size of all the temps
7102 // Used by RegSet::rsSpillChk()
7103 unsigned tmpGetCount; // Temps which haven't been released yet
7106 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
7108 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
7109 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
7112 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7113 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7117 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7118 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7122 CodeGenInterface* codeGen;
7124 // The following holds information about instr offsets in terms of generated code.
7128 IPmappingDsc* ipmdNext; // next line# record
7129 IL_OFFSETX ipmdILoffsx; // the instr offset
7130 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
7131 bool ipmdIsLabel; // Can this code be a branch label?
7134 // Record the instr offset mapping to the generated code
7136 IPmappingDsc* genIPmappingList;
7137 IPmappingDsc* genIPmappingLast;
7139 // Managed RetVal - A side hash table meant to record the mapping from a
7140 // GT_CALL node to its IL offset. This info is used to emit sequence points
7141 // that can be used by debugger to determine the native offset at which the
7142 // managed RetVal will be available.
7144 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
7145 // favor of a side table for two reasons: 1) We need IL offset for only those
7146 // GT_CALL nodes (created during importation) that correspond to an IL call and
7147 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
7148 // structure and IL offset is needed only when generating debuggable code. Therefore
7149 // it is desirable to avoid memory size penalty in retail scenarios.
7150 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, IL_OFFSETX> CallSiteILOffsetTable;
7151 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
7153 unsigned genReturnLocal; // Local number for the return value when applicable.
7154 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
7156 // The following properties are part of CodeGenContext. Getters are provided here for
7157 // convenience and backward compatibility, but the properties can only be set by invoking
7158 // the setter on CodeGenContext directly.
7160 __declspec(property(get = getEmitter)) emitter* genEmitter;
7161 emitter* getEmitter()
7163 return codeGen->getEmitter();
7166 const bool isFramePointerUsed()
7168 return codeGen->isFramePointerUsed();
7171 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
7172 bool getInterruptible()
7174 return codeGen->genInterruptible;
7176 void setInterruptible(bool value)
7178 codeGen->setInterruptible(value);
7181 #ifdef _TARGET_ARMARCH_
7182 __declspec(property(get = getHasTailCalls, put = setHasTailCalls)) bool hasTailCalls;
7183 bool getHasTailCalls()
7185 return codeGen->hasTailCalls;
7187 void setHasTailCalls(bool value)
7189 codeGen->setHasTailCalls(value);
7191 #endif // _TARGET_ARMARCH_
7194 const bool genDoubleAlign()
7196 return codeGen->doDoubleAlign();
7198 DWORD getCanDoubleAlign();
7199 bool shouldDoubleAlign(unsigned refCntStk,
7201 unsigned refCntWtdReg,
7202 unsigned refCntStkParam,
7203 unsigned refCntWtdStkDbl);
7204 #endif // DOUBLE_ALIGN
7206 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
7207 bool getFullPtrRegMap()
7209 return codeGen->genFullPtrRegMap;
7211 void setFullPtrRegMap(bool value)
7213 codeGen->setFullPtrRegMap(value);
7216 // Things that MAY belong either in CodeGen or CodeGenContext
7218 #if FEATURE_EH_FUNCLETS
7219 FuncInfoDsc* compFuncInfos;
7220 unsigned short compCurrFuncIdx;
7221 unsigned short compFuncInfoCount;
7223 unsigned short compFuncCount()
7225 assert(fgFuncletsCreated);
7226 return compFuncInfoCount;
7229 #else // !FEATURE_EH_FUNCLETS
7231 // This is a no-op when there are no funclets!
7232 void genUpdateCurrentFunclet(BasicBlock* block)
7237 FuncInfoDsc compFuncInfoRoot;
7239 static const unsigned compCurrFuncIdx = 0;
7241 unsigned short compFuncCount()
7246 #endif // !FEATURE_EH_FUNCLETS
7248 FuncInfoDsc* funCurrentFunc();
7249 void funSetCurrentFunc(unsigned funcIdx);
7250 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7251 unsigned int funGetFuncIdx(BasicBlock* block);
7255 VARSET_TP compCurLife; // current live variables
7256 GenTree* compCurLifeTree; // node after which compCurLife has been computed
7258 template <bool ForCodeGen>
7259 void compChangeLife(VARSET_VALARG_TP newLife);
7261 void genChangeLife(VARSET_VALARG_TP newLife)
7263 compChangeLife</*ForCodeGen*/ true>(newLife);
7266 template <bool ForCodeGen>
7267 void compUpdateLife(GenTree* tree);
7269 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7270 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7271 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7272 template <bool ForCodeGen>
7273 void compUpdateLifeVar(GenTree* tree, VARSET_TP* pLastUseVars = nullptr);
7275 template <bool ForCodeGen>
7276 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7278 // Gets a register mask that represent the kill set for a helper call since
7279 // not all JIT Helper calls follow the standard ABI on the target architecture.
7280 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7282 // Gets a register mask that represent the kill set for a NoGC helper call.
7283 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7286 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7287 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7288 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7289 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7290 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7291 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7292 #endif // _TARGET_ARM_
7294 // 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
7296 static GenTree* fgIsIndirOfAddrOfLocal(GenTree* tree);
7298 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7299 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7300 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7301 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7302 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7303 // for the tracked var indices of the field vars, as in a live var set).
7304 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7306 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7308 if (m_promotedStructDeathVars == nullptr)
7310 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7312 return m_promotedStructDeathVars;
7316 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7317 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7321 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7322 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7325 #if !defined(__GNUC__)
7326 #pragma region Unwind information
7331 // Infrastructure functions: start/stop/reserve/emit.
7334 void unwindBegProlog();
7335 void unwindEndProlog();
7336 void unwindBegEpilog();
7337 void unwindEndEpilog();
7338 void unwindReserve();
7339 void unwindEmit(void* pHotCode, void* pColdCode);
7342 // Specific unwind information functions: called by code generation to indicate a particular
7343 // prolog or epilog unwindable instruction has been generated.
7346 void unwindPush(regNumber reg);
7347 void unwindAllocStack(unsigned size);
7348 void unwindSetFrameReg(regNumber reg, unsigned offset);
7349 void unwindSaveReg(regNumber reg, unsigned offset);
7351 #if defined(_TARGET_ARM_)
7352 void unwindPushMaskInt(regMaskTP mask);
7353 void unwindPushMaskFloat(regMaskTP mask);
7354 void unwindPopMaskInt(regMaskTP mask);
7355 void unwindPopMaskFloat(regMaskTP mask);
7356 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7357 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7358 // called via unwindPadding().
7359 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7360 // instruction and the current location.
7361 #endif // _TARGET_ARM_
7363 #if defined(_TARGET_ARM64_)
7365 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7366 // instruction and the current location.
7367 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7368 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7369 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7370 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7371 void unwindSaveNext(); // unwind code: save_next
7372 void unwindReturn(regNumber reg); // ret lr
7373 #endif // defined(_TARGET_ARM64_)
7376 // Private "helper" functions for the unwind implementation.
7380 #if FEATURE_EH_FUNCLETS
7381 void unwindGetFuncLocations(FuncInfoDsc* func,
7382 bool getHotSectionData,
7383 /* OUT */ emitLocation** ppStartLoc,
7384 /* OUT */ emitLocation** ppEndLoc);
7385 #endif // FEATURE_EH_FUNCLETS
7387 void unwindReserveFunc(FuncInfoDsc* func);
7388 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7390 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7392 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7393 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7395 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7397 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7399 #if defined(_TARGET_AMD64_)
7401 void unwindBegPrologWindows();
7402 void unwindPushWindows(regNumber reg);
7403 void unwindAllocStackWindows(unsigned size);
7404 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7405 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7407 #ifdef UNIX_AMD64_ABI
7408 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7409 #endif // UNIX_AMD64_ABI
7410 #elif defined(_TARGET_ARM_)
7412 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7413 void unwindPushPopMaskFloat(regMaskTP mask);
7414 void unwindSplit(FuncInfoDsc* func);
7416 #endif // _TARGET_ARM_
7418 #if defined(_TARGET_UNIX_)
7419 int mapRegNumToDwarfReg(regNumber reg);
7420 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7421 void unwindPushPopCFI(regNumber reg);
7422 void unwindBegPrologCFI();
7423 void unwindPushPopMaskCFI(regMaskTP regMask, bool isFloat);
7424 void unwindAllocStackCFI(unsigned size);
7425 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7426 void unwindEmitFuncCFI(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7428 void DumpCfiInfo(bool isHotCode,
7429 UNATIVE_OFFSET startOffset,
7430 UNATIVE_OFFSET endOffset,
7432 const CFI_CODE* const pCfiCode);
7435 #endif // _TARGET_UNIX_
7437 #if !defined(__GNUC__)
7438 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7442 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7443 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7447 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7448 XX that contains the distinguished, well-known SIMD type definitions). XX
7450 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7451 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7454 // Get highest available level for SIMD codegen
7455 SIMDLevel getSIMDSupportLevel()
7457 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7458 if (compSupports(InstructionSet_AVX2))
7460 return SIMD_AVX2_Supported;
7463 // SIMD_SSE4_Supported actually requires all of SSE3, SSSE3, SSE4.1, and SSE4.2
7464 // to be supported. We can only enable it if all four are enabled in the compiler
7465 if (compSupports(InstructionSet_SSE42) && compSupports(InstructionSet_SSE41) &&
7466 compSupports(InstructionSet_SSSE3) && compSupports(InstructionSet_SSE3))
7468 return SIMD_SSE4_Supported;
7472 assert(canUseSSE2());
7473 return SIMD_SSE2_Supported;
7475 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7477 return SIMD_Not_Supported;
7483 // Should we support SIMD intrinsics?
7486 // Have we identified any SIMD types?
7487 // This is currently used by struct promotion to avoid getting type information for a struct
7488 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7490 bool _usesSIMDTypes;
7491 bool usesSIMDTypes()
7493 return _usesSIMDTypes;
7495 void setUsesSIMDTypes(bool value)
7497 _usesSIMDTypes = value;
7500 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7501 // that require indexed access to the individual fields of the vector, which is not well supported
7502 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7503 unsigned lvaSIMDInitTempVarNum;
7506 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7507 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7508 CORINFO_CLASS_HANDLE SIMDIntHandle;
7509 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7510 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7511 CORINFO_CLASS_HANDLE SIMDShortHandle;
7512 CORINFO_CLASS_HANDLE SIMDByteHandle;
7513 CORINFO_CLASS_HANDLE SIMDLongHandle;
7514 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7515 CORINFO_CLASS_HANDLE SIMDULongHandle;
7516 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7517 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7518 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7519 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7521 #ifdef FEATURE_HW_INTRINSICS
7522 #if defined(_TARGET_ARM64_)
7523 CORINFO_CLASS_HANDLE Vector64FloatHandle;
7524 CORINFO_CLASS_HANDLE Vector64UIntHandle;
7525 CORINFO_CLASS_HANDLE Vector64UShortHandle;
7526 CORINFO_CLASS_HANDLE Vector64UByteHandle;
7527 CORINFO_CLASS_HANDLE Vector64ShortHandle;
7528 CORINFO_CLASS_HANDLE Vector64ByteHandle;
7529 CORINFO_CLASS_HANDLE Vector64IntHandle;
7530 #endif // defined(_TARGET_ARM64_)
7531 CORINFO_CLASS_HANDLE Vector128FloatHandle;
7532 CORINFO_CLASS_HANDLE Vector128DoubleHandle;
7533 CORINFO_CLASS_HANDLE Vector128IntHandle;
7534 CORINFO_CLASS_HANDLE Vector128UShortHandle;
7535 CORINFO_CLASS_HANDLE Vector128UByteHandle;
7536 CORINFO_CLASS_HANDLE Vector128ShortHandle;
7537 CORINFO_CLASS_HANDLE Vector128ByteHandle;
7538 CORINFO_CLASS_HANDLE Vector128LongHandle;
7539 CORINFO_CLASS_HANDLE Vector128UIntHandle;
7540 CORINFO_CLASS_HANDLE Vector128ULongHandle;
7541 #if defined(_TARGET_XARCH_)
7542 CORINFO_CLASS_HANDLE Vector256FloatHandle;
7543 CORINFO_CLASS_HANDLE Vector256DoubleHandle;
7544 CORINFO_CLASS_HANDLE Vector256IntHandle;
7545 CORINFO_CLASS_HANDLE Vector256UShortHandle;
7546 CORINFO_CLASS_HANDLE Vector256UByteHandle;
7547 CORINFO_CLASS_HANDLE Vector256ShortHandle;
7548 CORINFO_CLASS_HANDLE Vector256ByteHandle;
7549 CORINFO_CLASS_HANDLE Vector256LongHandle;
7550 CORINFO_CLASS_HANDLE Vector256UIntHandle;
7551 CORINFO_CLASS_HANDLE Vector256ULongHandle;
7552 #endif // defined(_TARGET_XARCH_)
7553 #endif // FEATURE_HW_INTRINSICS
7555 // Get the handle for a SIMD type.
7556 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7558 if (simdBaseType == TYP_FLOAT)
7563 return SIMDVector2Handle;
7565 return SIMDVector3Handle;
7567 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7569 return SIMDVector4Handle;
7578 assert(simdType == getSIMDVectorType());
7579 switch (simdBaseType)
7582 return SIMDFloatHandle;
7584 return SIMDDoubleHandle;
7586 return SIMDIntHandle;
7588 return SIMDUShortHandle;
7590 return SIMDUByteHandle;
7592 return SIMDShortHandle;
7594 return SIMDByteHandle;
7596 return SIMDLongHandle;
7598 return SIMDUIntHandle;
7600 return SIMDULongHandle;
7602 assert(!"Didn't find a class handle for simdType");
7604 return NO_CLASS_HANDLE;
7608 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7609 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7610 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7612 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7613 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7614 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7615 bool isSIMDTypeLocal(GenTree* tree)
7617 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7620 // Returns true if the type of the tree is a byref of TYP_SIMD
7621 bool isAddrOfSIMDType(GenTree* tree)
7623 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7625 switch (tree->OperGet())
7628 return varTypeIsSIMD(tree->gtGetOp1());
7630 case GT_LCL_VAR_ADDR:
7631 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7634 return isSIMDTypeLocal(tree);
7641 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7643 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7644 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7645 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7648 // Returns base type of a TYP_SIMD local.
7649 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7650 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7652 if (isSIMDTypeLocal(tree))
7654 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7660 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7662 return info.compCompHnd->isInSIMDModule(clsHnd);
7665 bool isIntrinsicType(CORINFO_CLASS_HANDLE clsHnd)
7667 return (info.compCompHnd->getClassAttribs(clsHnd) & CORINFO_FLG_INTRINSIC_TYPE) != 0;
7670 const char* getClassNameFromMetadata(CORINFO_CLASS_HANDLE cls, const char** namespaceName)
7672 return info.compCompHnd->getClassNameFromMetadata(cls, namespaceName);
7675 CORINFO_CLASS_HANDLE getTypeInstantiationArgument(CORINFO_CLASS_HANDLE cls, unsigned index)
7677 return info.compCompHnd->getTypeInstantiationArgument(cls, index);
7680 bool isSIMDClass(typeInfo* pTypeInfo)
7682 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7685 bool isHWSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7687 #ifdef FEATURE_HW_INTRINSICS
7688 if (isIntrinsicType(clsHnd))
7690 const char* namespaceName = nullptr;
7691 (void)getClassNameFromMetadata(clsHnd, &namespaceName);
7692 return strcmp(namespaceName, "System.Runtime.Intrinsics") == 0;
7694 #endif // FEATURE_HW_INTRINSICS
7698 bool isHWSIMDClass(typeInfo* pTypeInfo)
7700 #ifdef FEATURE_HW_INTRINSICS
7701 return pTypeInfo->IsStruct() && isHWSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7707 bool isSIMDorHWSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7709 return isSIMDClass(clsHnd) || isHWSIMDClass(clsHnd);
7712 bool isSIMDorHWSIMDClass(typeInfo* pTypeInfo)
7714 return isSIMDClass(pTypeInfo) || isHWSIMDClass(pTypeInfo);
7717 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7718 // if it is not a SIMD type or is an unsupported base type.
7719 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7721 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7723 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7726 // Get SIMD Intrinsic info given the method handle.
7727 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7728 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7729 CORINFO_METHOD_HANDLE methodHnd,
7730 CORINFO_SIG_INFO* sig,
7733 var_types* baseType,
7734 unsigned* sizeBytes);
7736 // Pops and returns GenTree node from importers type stack.
7737 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7738 GenTree* impSIMDPopStack(var_types type, bool expectAddr = false);
7740 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7741 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7743 // Creates a GT_SIMD tree for Select operation
7744 GenTree* impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7746 unsigned simdVectorSize,
7751 // Creates a GT_SIMD tree for Min/Max operation
7752 GenTree* impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7753 CORINFO_CLASS_HANDLE typeHnd,
7755 unsigned simdVectorSize,
7759 // Transforms operands and returns the SIMD intrinsic to be applied on
7760 // transformed operands to obtain given relop result.
7761 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7762 CORINFO_CLASS_HANDLE typeHnd,
7763 unsigned simdVectorSize,
7764 var_types* baseType,
7768 // Creates a GT_SIMD tree for Abs intrinsic.
7769 GenTree* impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7771 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7772 // Transforms operands and returns the SIMD intrinsic to be applied on
7773 // transformed operands to obtain == comparison result.
7774 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7775 unsigned simdVectorSize,
7779 // Transforms operands and returns the SIMD intrinsic to be applied on
7780 // transformed operands to obtain > comparison result.
7781 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7782 unsigned simdVectorSize,
7786 // Transforms operands and returns the SIMD intrinsic to be applied on
7787 // transformed operands to obtain >= comparison result.
7788 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7789 unsigned simdVectorSize,
7793 // Transforms operands and returns the SIMD intrinsic to be applied on
7794 // transformed operands to obtain >= comparison result in case of int32
7795 // and small int base type vectors.
7796 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7797 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7798 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7800 void setLclRelatedToSIMDIntrinsic(GenTree* tree);
7801 bool areFieldsContiguous(GenTree* op1, GenTree* op2);
7802 bool areArrayElementsContiguous(GenTree* op1, GenTree* op2);
7803 bool areArgumentsContiguous(GenTree* op1, GenTree* op2);
7804 GenTree* createAddressNodeForSIMDInit(GenTree* tree, unsigned simdSize);
7806 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7807 GenTree* impSIMDIntrinsic(OPCODE opcode,
7808 GenTree* newobjThis,
7809 CORINFO_CLASS_HANDLE clsHnd,
7810 CORINFO_METHOD_HANDLE method,
7811 CORINFO_SIG_INFO* sig,
7814 GenTree* getOp1ForConstructor(OPCODE opcode, GenTree* newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7816 // Whether SIMD vector occupies part of SIMD register.
7817 // SSE2: vector2f/3f are considered sub register SIMD types.
7818 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7819 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7821 unsigned sizeBytes = 0;
7822 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7823 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7826 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7828 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7831 // Get the type for the hardware SIMD vector.
7832 // This is the maximum SIMD type supported for this target.
7833 var_types getSIMDVectorType()
7835 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7836 if (getSIMDSupportLevel() == SIMD_AVX2_Supported)
7842 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7845 #elif defined(_TARGET_ARM64_)
7848 assert(!"getSIMDVectorType() unimplemented on target arch");
7853 // Get the size of the SIMD type in bytes
7854 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7856 unsigned sizeBytes = 0;
7857 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7861 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7862 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7864 // Get the the number of elements of basetype of SIMD vector given by its type handle
7865 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7867 // Get preferred alignment of SIMD type.
7868 int getSIMDTypeAlignment(var_types simdType);
7870 // Get the number of bytes in a System.Numeric.Vector<T> for the current compilation.
7871 // Note - cannot be used for System.Runtime.Intrinsic
7872 unsigned getSIMDVectorRegisterByteLength()
7874 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7875 if (getSIMDSupportLevel() == SIMD_AVX2_Supported)
7877 return YMM_REGSIZE_BYTES;
7881 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7882 return XMM_REGSIZE_BYTES;
7884 #elif defined(_TARGET_ARM64_)
7885 return FP_REGSIZE_BYTES;
7887 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7892 // The minimum and maximum possible number of bytes in a SIMD vector.
7894 // maxSIMDStructBytes
7895 // The minimum SIMD size supported by System.Numeric.Vectors or System.Runtime.Intrinsic
7896 // SSE: 16-byte Vector<T> and Vector128<T>
7897 // AVX: 32-byte Vector256<T> (Vector<T> is 16-byte)
7898 // AVX2: 32-byte Vector<T> and Vector256<T>
7899 unsigned int maxSIMDStructBytes()
7901 #if defined(FEATURE_HW_INTRINSICS) && defined(_TARGET_XARCH_)
7902 if (compSupports(InstructionSet_AVX))
7904 return YMM_REGSIZE_BYTES;
7908 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7909 return XMM_REGSIZE_BYTES;
7912 return getSIMDVectorRegisterByteLength();
7915 unsigned int minSIMDStructBytes()
7917 return emitTypeSize(TYP_SIMD8);
7920 // Returns the codegen type for a given SIMD size.
7921 var_types getSIMDTypeForSize(unsigned size)
7923 var_types simdType = TYP_UNDEF;
7926 simdType = TYP_SIMD8;
7928 else if (size == 12)
7930 simdType = TYP_SIMD12;
7932 else if (size == 16)
7934 simdType = TYP_SIMD16;
7936 else if (size == 32)
7938 simdType = TYP_SIMD32;
7942 noway_assert(!"Unexpected size for SIMD type");
7947 unsigned getSIMDInitTempVarNum()
7949 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7951 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7952 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7954 return lvaSIMDInitTempVarNum;
7957 #endif // FEATURE_SIMD
7960 //------------------------------------------------------------------------
7961 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7963 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7964 // candidate for enregistration.
7966 unsigned largestEnregisterableStructSize()
7969 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7970 if (vectorRegSize > TARGET_POINTER_SIZE)
7972 return vectorRegSize;
7975 #endif // FEATURE_SIMD
7977 return TARGET_POINTER_SIZE;
7982 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7983 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7984 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7986 // Is this var is of type simd struct?
7987 bool lclVarIsSIMDType(unsigned varNum)
7989 LclVarDsc* varDsc = lvaTable + varNum;
7990 return varDsc->lvIsSIMDType();
7993 // Is this Local node a SIMD local?
7994 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7996 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7999 // Returns true if the TYP_SIMD locals on stack are aligned at their
8000 // preferred byte boundary specified by getSIMDTypeAlignment().
8002 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
8003 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
8004 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
8005 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
8006 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
8007 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
8008 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
8011 bool isSIMDTypeLocalAligned(unsigned varNum)
8013 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
8014 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
8017 int off = lvaFrameAddress(varNum, &ebpBased);
8018 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
8019 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
8020 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
8023 #endif // FEATURE_SIMD
8028 // Whether SSE and SSE2 is available
8029 bool canUseSSE2() const
8031 #ifdef _TARGET_XARCH_
8032 return opts.compCanUseSSE2;
8038 bool compSupports(InstructionSet isa) const
8040 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
8041 return (opts.compSupportsISA & (1ULL << isa)) != 0;
8047 bool canUseVexEncoding() const
8049 #ifdef _TARGET_XARCH_
8050 return compSupports(InstructionSet_AVX);
8057 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8058 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8062 XX Generic info about the compilation and the method being compiled. XX
8063 XX It is responsible for driving the other phases. XX
8064 XX It is also responsible for all the memory management. XX
8066 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8067 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8071 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
8073 InlineResult* compInlineResult; // The result of importing the inlinee method.
8075 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
8076 bool compJmpOpUsed; // Does the method do a JMP
8077 bool compLongUsed; // Does the method use TYP_LONG
8078 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
8079 bool compTailCallUsed; // Does the method do a tailcall
8080 bool compLocallocUsed; // Does the method use localloc.
8081 bool compLocallocOptimized; // Does the method have an optimized localloc
8082 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
8083 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
8084 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
8086 // NOTE: These values are only reliable after
8087 // the importing is completely finished.
8089 #ifdef LEGACY_BACKEND
8090 JitExpandArrayStack<GenTree*>* compQMarks; // The set of QMark nodes created in the current compilation, so
8091 // we can iterate over these efficiently.
8094 #if CPU_USES_BLOCK_MOVE
8095 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
8099 // State information - which phases have completed?
8100 // These are kept together for easy discoverability
8102 bool bRangeAllowStress;
8103 bool compCodeGenDone;
8104 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
8105 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
8106 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
8107 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
8110 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
8111 bool fgLocalVarLivenessChanged;
8113 bool compStackProbePrologDone;
8115 #ifndef LEGACY_BACKEND
8117 #endif // !LEGACY_BACKEND
8118 bool compRationalIRForm;
8120 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
8122 bool compGeneratingProlog;
8123 bool compGeneratingEpilog;
8124 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
8125 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
8126 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
8127 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
8128 bool getNeedsGSSecurityCookie() const
8130 return compNeedsGSSecurityCookie;
8132 void setNeedsGSSecurityCookie()
8134 compNeedsGSSecurityCookie = true;
8137 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
8138 // frame layout calculations, this is the level we are currently
8141 //---------------------------- JITing options -----------------------------
8154 JitFlags* jitFlags; // all flags passed from the EE
8155 unsigned compFlags; // method attributes
8157 codeOptimize compCodeOpt; // what type of code optimizations
8161 #ifdef _TARGET_XARCH_
8162 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
8163 #endif // _TARGET_XARCH_
8165 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
8166 uint64_t compSupportsISA;
8167 void setSupportedISA(InstructionSet isa)
8169 compSupportsISA |= 1ULL << isa;
8173 // optimize maximally and/or favor speed over size?
8175 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
8176 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
8177 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
8178 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
8179 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
8181 // Maximun number of locals before turning off the inlining
8182 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
8185 unsigned instrCount;
8186 unsigned lvRefCount;
8187 bool compMinOptsIsSet;
8189 bool compMinOptsIsUsed;
8191 inline bool MinOpts()
8193 assert(compMinOptsIsSet);
8194 compMinOptsIsUsed = true;
8197 inline bool IsMinOptsSet()
8199 return compMinOptsIsSet;
8202 inline bool MinOpts()
8206 inline bool IsMinOptsSet()
8208 return compMinOptsIsSet;
8211 inline void SetMinOpts(bool val)
8213 assert(!compMinOptsIsUsed);
8214 assert(!compMinOptsIsSet || (compMinOpts == val));
8216 compMinOptsIsSet = true;
8219 // true if the CLFLG_* for an optimization is set.
8220 inline bool OptEnabled(unsigned optFlag)
8222 return !!(compFlags & optFlag);
8225 #ifdef FEATURE_READYTORUN_COMPILER
8226 inline bool IsReadyToRun()
8228 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
8231 inline bool IsReadyToRun()
8237 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
8238 // PInvoke transitions inline (e.g. when targeting CoreRT).
8239 inline bool ShouldUsePInvokeHelpers()
8241 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
8244 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
8246 inline bool IsReversePInvoke()
8248 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
8251 // true if we must generate code compatible with JIT32 quirks
8252 inline bool IsJit32Compat()
8254 #if defined(_TARGET_X86_)
8255 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
8261 // true if we must generate code compatible with Jit64 quirks
8262 inline bool IsJit64Compat()
8264 #if defined(_TARGET_AMD64_)
8265 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
8266 #elif !defined(FEATURE_CORECLR)
8273 bool compScopeInfo; // Generate the LocalVar info ?
8274 bool compDbgCode; // Generate debugger-friendly code?
8275 bool compDbgInfo; // Gather debugging info?
8278 #ifdef PROFILING_SUPPORTED
8279 bool compNoPInvokeInlineCB;
8281 static const bool compNoPInvokeInlineCB;
8285 bool compGcChecks; // Check arguments and return values to ensure they are sane
8286 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
8287 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
8291 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
8292 // to be allocated on the stack.
8293 // It will be set to true in the following cases:
8294 // 1. When the method being compiled has a declarative security
8295 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
8296 // This is also the case when we inject a prolog and epilog in the method.
8298 // 2. When the method being compiled has imperative security (i.e. the method
8299 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
8301 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
8303 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
8304 // which gets reported as a GC root to stackwalker.
8305 // (See also ICodeManager::GetAddrOfSecurityObject.)
8307 bool compReloc; // Generate relocs for pointers in code, true for all ngen/prejit codegen
8310 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
8311 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
8315 #ifdef UNIX_AMD64_ABI
8316 // This flag is indicating if there is a need to align the frame.
8317 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8318 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8319 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8320 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8321 // there are calls and making sure the frame alignment logic is executed.
8322 bool compNeedToAlignFrame;
8323 #endif // UNIX_AMD64_ABI
8325 bool compProcedureSplitting; // Separate cold code from hot code
8327 bool genFPorder; // Preserve FP order (operations are non-commutative)
8328 bool genFPopt; // Can we do frame-pointer-omission optimization?
8329 bool altJit; // True if we are an altjit and are compiling this method
8332 bool optRepeat; // Repeat optimizer phases k times
8336 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8337 bool dspCode; // Display native code generated
8338 bool dspEHTable; // Display the EH table reported to the VM
8339 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8340 bool dspEmit; // Display emitter output
8341 bool dspLines; // Display source-code lines intermixed with native code output
8342 bool dmpHex; // Display raw bytes in hex of native code output
8343 bool varNames; // Display variables names in native code output
8344 bool disAsm; // Display native code as it is generated
8345 bool disAsmSpilled; // Display native code when any register spilling occurs
8346 bool disDiffable; // Makes the Disassembly code 'diff-able'
8347 bool disAsm2; // Display native code after it is generated using external disassembler
8348 bool dspOrder; // Display names of each of the methods that we ngen/jit
8349 bool dspUnwind; // Display the unwind info output
8350 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8351 bool compLongAddress; // Force using large pseudo instructions for long address
8352 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8353 bool dspGCtbls; // Display the GC tables
8357 bool doLateDisasm; // Run the late disassembler
8358 #endif // LATE_DISASM
8360 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8361 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8362 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8363 static const bool dspGCtbls = true;
8366 // We need stack probes to guarantee that we won't trigger a stack overflow
8367 // when calling unmanaged code until they get a chance to set up a frame, because
8368 // the EE will have no idea where it is.
8370 // We will only be doing this currently for hosted environments. Unfortunately
8371 // we need to take care of stubs, so potentially, we will have to do the probes
8372 // for any call. We have a plan for not needing for stubs though
8373 bool compNeedStackProbes;
8375 #ifdef PROFILING_SUPPORTED
8376 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8377 // This option helps make the JIT behave as if it is running under a profiler.
8378 bool compJitELTHookEnabled;
8379 #endif // PROFILING_SUPPORTED
8381 #if FEATURE_TAILCALL_OPT
8382 // Whether opportunistic or implicit tail call optimization is enabled.
8383 bool compTailCallOpt;
8384 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8385 bool compTailCallLoopOpt;
8389 static const bool compUseSoftFP = true;
8390 #else // !ARM_SOFTFP
8391 static const bool compUseSoftFP = false;
8394 GCPollType compGCPollType;
8398 static bool s_pAltJitExcludeAssembliesListInitialized;
8399 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8404 template <typename T>
8407 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8410 template <typename T>
8413 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8416 static int dspTreeID(GenTree* tree)
8418 return tree->gtTreeID;
8420 static void printTreeID(GenTree* tree)
8422 if (tree == nullptr)
8428 printf("[%06d]", dspTreeID(tree));
8435 #define STRESS_MODES \
8439 /* "Variations" stress areas which we try to mix up with each other. */ \
8440 /* These should not be exhaustively used as they might */ \
8441 /* hide/trivialize other areas */ \
8444 STRESS_MODE(DBL_ALN) \
8445 STRESS_MODE(LCL_FLDS) \
8446 STRESS_MODE(UNROLL_LOOPS) \
8447 STRESS_MODE(MAKE_CSE) \
8448 STRESS_MODE(LEGACY_INLINE) \
8449 STRESS_MODE(CLONE_EXPR) \
8450 STRESS_MODE(USE_FCOMI) \
8451 STRESS_MODE(USE_CMOV) \
8453 STRESS_MODE(BB_PROFILE) \
8454 STRESS_MODE(OPT_BOOLS_GC) \
8455 STRESS_MODE(REMORPH_TREES) \
8456 STRESS_MODE(64RSLT_MUL) \
8457 STRESS_MODE(DO_WHILE_LOOPS) \
8458 STRESS_MODE(MIN_OPTS) \
8459 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8460 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8461 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8462 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8463 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8464 STRESS_MODE(NULL_OBJECT_CHECK) \
8465 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8466 STRESS_MODE(RANDOM_INLINE) \
8467 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8468 STRESS_MODE(GENERIC_VARN) \
8470 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8472 STRESS_MODE(COUNT_VARN) \
8474 /* "Check" stress areas that can be exhaustively used if we */ \
8475 /* dont care about performance at all */ \
8477 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8478 STRESS_MODE(CHK_FLOW_UPDATE) \
8479 STRESS_MODE(EMITTER) \
8480 STRESS_MODE(CHK_REIMPORT) \
8481 STRESS_MODE(FLATFP) \
8482 STRESS_MODE(GENERIC_CHECK) \
8487 #define STRESS_MODE(mode) STRESS_##mode,
8494 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8495 BYTE compActiveStressModes[STRESS_COUNT];
8498 #define MAX_STRESS_WEIGHT 100
8500 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8504 bool compInlineStress()
8506 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8509 bool compRandomInlineStress()
8511 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8516 bool compTailCallStress()
8519 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8525 codeOptimize compCodeOpt()
8528 // Switching between size & speed has measurable throughput impact
8529 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8530 // DEBUG, but should generate identical code between CHK & RET builds,
8531 // so that's not acceptable.
8532 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8533 // Investigate the cause of the throughput regression.
8535 return opts.compCodeOpt;
8537 return BLENDED_CODE;
8541 //--------------------- Info about the procedure --------------------------
8545 COMP_HANDLE compCompHnd;
8546 CORINFO_MODULE_HANDLE compScopeHnd;
8547 CORINFO_CLASS_HANDLE compClassHnd;
8548 CORINFO_METHOD_HANDLE compMethodHnd;
8549 CORINFO_METHOD_INFO* compMethodInfo;
8551 BOOL hasCircularClassConstraints;
8552 BOOL hasCircularMethodConstraints;
8554 #if defined(DEBUG) || defined(LATE_DISASM)
8555 const char* compMethodName;
8556 const char* compClassName;
8557 const char* compFullName;
8558 #endif // defined(DEBUG) || defined(LATE_DISASM)
8560 #if defined(DEBUG) || defined(INLINE_DATA)
8561 // Method hash is logcally const, but computed
8563 mutable unsigned compMethodHashPrivate;
8564 unsigned compMethodHash() const;
8565 #endif // defined(DEBUG) || defined(INLINE_DATA)
8567 #ifdef PSEUDORANDOM_NOP_INSERTION
8568 // things for pseudorandom nop insertion
8569 unsigned compChecksum;
8573 // The following holds the FLG_xxxx flags for the method we're compiling.
8576 // The following holds the class attributes for the method we're compiling.
8577 unsigned compClassAttr;
8579 const BYTE* compCode;
8580 IL_OFFSET compILCodeSize; // The IL code size
8581 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8582 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8583 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8584 // (2) the code is hot/cold split, and we issued less code than we expected
8585 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8587 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8588 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8589 bool compIsContextful : 1; // contextful method
8590 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8591 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8592 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8593 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8594 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8596 var_types compRetType; // Return type of the method as declared in IL
8597 var_types compRetNativeType; // Normalized return type as per target arch ABI
8598 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8599 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8601 #if FEATURE_FASTTAILCALL
8602 size_t compArgStackSize; // Incoming argument stack size in bytes
8603 #endif // FEATURE_FASTTAILCALL
8605 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8606 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8607 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8608 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8609 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8610 unsigned compMaxStack;
8611 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8612 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8614 unsigned compCallUnmanaged; // count of unmanaged calls
8615 unsigned compLvFrameListRoot; // lclNum for the Frame root
8616 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8617 // You should generally use compHndBBtabCount instead: it is the
8618 // current number of EH clauses (after additions like synchronized
8619 // methods and funclets, and removals like unreachable code deletion).
8621 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8622 // and the VM expects that, or the JIT is a "self-host" compiler
8623 // (e.g., x86 hosted targeting x86) and the VM expects that.
8625 /* The following holds IL scope information about local variables.
8628 unsigned compVarScopesCount;
8629 VarScopeDsc* compVarScopes;
8631 /* The following holds information about instr offsets for
8632 * which we need to report IP-mappings
8635 IL_OFFSET* compStmtOffsets; // sorted
8636 unsigned compStmtOffsetsCount;
8637 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8639 #define CPU_X86 0x0100 // The generic X86 CPU
8640 #define CPU_X86_PENTIUM_4 0x0110
8642 #define CPU_X64 0x0200 // The generic x64 CPU
8643 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8644 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8646 #define CPU_ARM 0x0300 // The generic ARM CPU
8648 unsigned genCPU; // What CPU are we running on
8651 // Returns true if the method being compiled returns a non-void and non-struct value.
8652 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8653 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8654 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8655 // Methods returning such structs are considered to return non-struct return value and
8656 // this method returns true in that case.
8657 bool compMethodReturnsNativeScalarType()
8659 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8662 // Returns true if the method being compiled returns RetBuf addr as its return value
8663 bool compMethodReturnsRetBufAddr()
8665 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8666 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8668 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8669 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8670 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8671 // methods with hidden RetBufArg.
8673 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8674 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8675 // returning the address of RetBuf.
8677 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8678 // to be returned in RAX.
8679 CLANG_FORMAT_COMMENT_ANCHOR;
8681 #ifdef _TARGET_AMD64_
8682 return (info.compRetBuffArg != BAD_VAR_NUM);
8683 #else // !_TARGET_AMD64_
8684 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8685 #endif // !_TARGET_AMD64_
8688 // Returns true if the method returns a value in more than one return register
8689 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8690 // TODO-ARM64: Does this apply for ARM64 too?
8691 bool compMethodReturnsMultiRegRetType()
8693 #if FEATURE_MULTIREG_RET
8694 #if defined(_TARGET_X86_)
8695 // On x86 only 64-bit longs are returned in multiple registers
8696 return varTypeIsLong(info.compRetNativeType);
8697 #else // targets: X64-UNIX, ARM64 or ARM32
8698 // On all other targets that support multireg return values:
8699 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8700 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8701 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8702 #endif // TARGET_XXX
8704 #else // not FEATURE_MULTIREG_RET
8706 // For this architecture there are no multireg returns
8709 #endif // FEATURE_MULTIREG_RET
8712 #if FEATURE_MULTIREG_ARGS
8713 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8714 // return the gcPtr layout for the pointers sized fields
8715 void getStructGcPtrsFromOp(GenTree* op, BYTE* gcPtrsOut);
8716 #endif // FEATURE_MULTIREG_ARGS
8718 // Returns true if the method being compiled returns a value
8719 bool compMethodHasRetVal()
8721 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8722 compMethodReturnsMultiRegRetType();
8727 void compDispLocalVars();
8731 //-------------------------- Global Compiler Data ------------------------------------
8734 static unsigned s_compMethodsCount; // to produce unique label names
8735 unsigned compGenTreeID;
8736 unsigned compBasicBlockID;
8739 BasicBlock* compCurBB; // the current basic block in process
8740 GenTree* compCurStmt; // the current statement in process
8742 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8745 // The following is used to create the 'method JIT info' block.
8746 size_t compInfoBlkSize;
8747 BYTE* compInfoBlkAddr;
8749 EHblkDsc* compHndBBtab; // array of EH data
8750 unsigned compHndBBtabCount; // element count of used elements in EH data array
8751 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8753 #if defined(_TARGET_X86_)
8755 //-------------------------------------------------------------------------
8756 // Tracking of region covered by the monitor in synchronized methods
8757 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8758 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8760 #endif // !_TARGET_X86_
8762 Phases previousCompletedPhase; // the most recently completed phase
8764 //-------------------------------------------------------------------------
8765 // The following keeps track of how many bytes of local frame space we've
8766 // grabbed so far in the current function, and how many argument bytes we
8767 // need to pop when we return.
8770 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8772 // Count of callee-saved regs we pushed in the prolog.
8773 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8774 // In case of Amd64 this doesn't include float regs saved on stack.
8775 unsigned compCalleeRegsPushed;
8777 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8778 // Mask of callee saved float regs on stack.
8779 regMaskTP compCalleeFPRegsSavedMask;
8781 #ifdef _TARGET_AMD64_
8782 // Quirk for VS debug-launch scenario to work:
8783 // Bytes of padding between save-reg area and locals.
8784 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8785 unsigned compVSQuirkStackPaddingNeeded;
8786 bool compQuirkForPPPflag;
8789 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8791 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8792 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8793 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8795 //-------------------------------------------------------------------------
8797 static void compStartup(); // One-time initialization
8798 static void compShutdown(); // One-time finalization
8800 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8803 static void compDisplayStaticSizes(FILE* fout);
8805 //------------ Some utility functions --------------
8807 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8808 void** ppIndirection); /* OUT */
8810 // Several JIT/EE interface functions return a CorInfoType, and also return a
8811 // class handle as an out parameter if the type is a value class. Returns the
8812 // size of the type these describe.
8813 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8816 // Components used by the compiler may write unit test suites, and
8817 // have them run within this method. They will be run only once per process, and only
8818 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8819 // These should fail by asserting.
8820 void compDoComponentUnitTestsOnce();
8823 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8824 CORINFO_MODULE_HANDLE classPtr,
8825 COMP_HANDLE compHnd,
8826 CORINFO_METHOD_INFO* methodInfo,
8827 void** methodCodePtr,
8828 ULONG* methodCodeSize,
8829 JitFlags* compileFlags);
8830 void compCompileFinish();
8831 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8832 COMP_HANDLE compHnd,
8833 CORINFO_METHOD_INFO* methodInfo,
8834 void** methodCodePtr,
8835 ULONG* methodCodeSize,
8836 JitFlags* compileFlags,
8837 CorInfoInstantiationVerification instVerInfo);
8839 ArenaAllocator* compGetAllocator();
8841 #if MEASURE_MEM_ALLOC
8843 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8847 unsigned allocCnt; // # of allocs
8848 UINT64 allocSz; // total size of those alloc.
8849 UINT64 allocSzMax; // Maximum single allocation.
8850 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8851 UINT64 nraTotalSizeAlloc;
8852 UINT64 nraTotalSizeUsed;
8854 static const char* s_CompMemKindNames[]; // Names of the kinds.
8856 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8858 for (int i = 0; i < CMK_Count; i++)
8860 allocSzByKind[i] = 0;
8863 MemStats(const MemStats& ms)
8864 : allocCnt(ms.allocCnt)
8865 , allocSz(ms.allocSz)
8866 , allocSzMax(ms.allocSzMax)
8867 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8868 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8870 for (int i = 0; i < CMK_Count; i++)
8872 allocSzByKind[i] = ms.allocSzByKind[i];
8876 // Until we have ubiquitous constructors.
8879 this->MemStats::MemStats();
8882 void AddAlloc(size_t sz, CompMemKind cmk)
8886 if (sz > allocSzMax)
8890 allocSzByKind[cmk] += sz;
8893 void Print(FILE* f); // Print these stats to f.
8894 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8896 MemStats genMemStats;
8898 struct AggregateMemStats : public MemStats
8902 AggregateMemStats() : MemStats(), nMethods(0)
8906 void Add(const MemStats& ms)
8909 allocCnt += ms.allocCnt;
8910 allocSz += ms.allocSz;
8911 allocSzMax = max(allocSzMax, ms.allocSzMax);
8912 for (int i = 0; i < CMK_Count; i++)
8914 allocSzByKind[i] += ms.allocSzByKind[i];
8916 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8917 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8920 void Print(FILE* f); // Print these stats to jitstdout.
8923 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8924 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8925 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8927 #endif // MEASURE_MEM_ALLOC
8929 #if LOOP_HOIST_STATS
8930 unsigned m_loopsConsidered;
8931 bool m_curLoopHasHoistedExpression;
8932 unsigned m_loopsWithHoistedExpressions;
8933 unsigned m_totalHoistedExpressions;
8935 void AddLoopHoistStats();
8936 void PrintPerMethodLoopHoistStats();
8938 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8939 static unsigned s_loopsConsidered;
8940 static unsigned s_loopsWithHoistedExpressions;
8941 static unsigned s_totalHoistedExpressions;
8943 static void PrintAggregateLoopHoistStats(FILE* f);
8944 #endif // LOOP_HOIST_STATS
8946 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8947 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8948 void compFreeMem(void*);
8950 bool compIsForImportOnly();
8951 bool compIsForInlining();
8952 bool compDonotInline();
8955 unsigned char compGetJitDefaultFill(); // Get the default fill char value
8956 // we randomize this value when JitStress is enabled
8958 const char* compLocalVarName(unsigned varNum, unsigned offs);
8959 VarName compVarName(regNumber reg, bool isFloatReg = false);
8960 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8961 const char* compRegPairName(regPairNo regPair);
8962 const char* compRegNameForSize(regNumber reg, size_t size);
8963 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8964 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8965 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8968 //-------------------------------------------------------------------------
8970 struct VarScopeListNode
8973 VarScopeListNode* next;
8974 static VarScopeListNode* Create(VarScopeDsc* value, CompAllocator* alloc)
8976 VarScopeListNode* node = new (alloc) VarScopeListNode;
8978 node->next = nullptr;
8983 struct VarScopeMapInfo
8985 VarScopeListNode* head;
8986 VarScopeListNode* tail;
8987 static VarScopeMapInfo* Create(VarScopeListNode* node, CompAllocator* alloc)
8989 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8996 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8997 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8999 typedef JitHashTable<unsigned, JitSmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*> VarNumToScopeDscMap;
9001 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
9002 VarNumToScopeDscMap* compVarScopeMap;
9004 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
9006 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
9008 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
9010 void compInitVarScopeMap();
9012 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
9013 // enter scope, sorted by instr offset
9014 unsigned compNextEnterScope;
9016 VarScopeDsc** compExitScopeList; // List has the offsets where variables
9017 // go out of scope, sorted by instr offset
9018 unsigned compNextExitScope;
9020 void compInitScopeLists();
9022 void compResetScopeLists();
9024 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
9026 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
9028 void compProcessScopesUntil(unsigned offset,
9030 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
9031 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
9034 void compDispScopeLists();
9037 bool compIsProfilerHookNeeded();
9039 //-------------------------------------------------------------------------
9040 /* Statistical Data Gathering */
9042 void compJitStats(); // call this function and enable
9043 // various ifdef's below for statistical data
9046 void compCallArgStats();
9047 static void compDispCallArgStats(FILE* fout);
9050 //-------------------------------------------------------------------------
9057 ArenaAllocator* compAllocator;
9060 CompAllocator* compAllocatorGeneric; // An allocator that uses the CMK_Generic tracker.
9061 #if MEASURE_MEM_ALLOC
9062 CompAllocator* compAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
9063 CompAllocator* compAllocatorGC; // An allocator that uses the CMK_GC tracker.
9064 CompAllocator* compAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
9066 CompAllocator* compAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
9068 #endif // MEASURE_MEM_ALLOC
9070 void compFunctionTraceStart();
9071 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
9074 size_t compMaxUncheckedOffsetForNullObject;
9076 void compInitOptions(JitFlags* compileFlags);
9078 void compSetProcessor();
9079 void compInitDebuggingInfo();
9080 void compSetOptimizationLevel();
9081 #ifdef _TARGET_ARMARCH_
9082 bool compRsvdRegCheck(FrameLayoutState curState);
9084 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
9086 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
9087 void ResetOptAnnotations();
9089 // Regenerate loop descriptors; to be used between iterations when repeating opts.
9090 void RecomputeLoopInfo();
9092 #ifdef PROFILING_SUPPORTED
9093 // Data required for generating profiler Enter/Leave/TailCall hooks
9095 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
9096 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
9097 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
9100 #ifdef _TARGET_AMD64_
9101 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
9104 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
9105 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
9107 CompAllocator* getAllocator()
9109 return compAllocatorGeneric;
9112 #if MEASURE_MEM_ALLOC
9113 CompAllocator* getAllocatorBitset()
9115 return compAllocatorBitset;
9117 CompAllocator* getAllocatorGC()
9119 return compAllocatorGC;
9121 CompAllocator* getAllocatorLoopHoist()
9123 return compAllocatorLoopHoist;
9125 #else // !MEASURE_MEM_ALLOC
9126 CompAllocator* getAllocatorBitset()
9128 return compAllocatorGeneric;
9130 CompAllocator* getAllocatorGC()
9132 return compAllocatorGeneric;
9134 CompAllocator* getAllocatorLoopHoist()
9136 return compAllocatorGeneric;
9138 #endif // !MEASURE_MEM_ALLOC
9141 CompAllocator* getAllocatorDebugOnly()
9143 #if MEASURE_MEM_ALLOC
9144 return compAllocatorDebugOnly;
9145 #else // !MEASURE_MEM_ALLOC
9146 return compAllocatorGeneric;
9147 #endif // !MEASURE_MEM_ALLOC
9152 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9153 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9157 XX Checks for type compatibility and merges types XX
9159 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9160 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9164 // Set to TRUE if verification cannot be skipped for this method
9165 // If we detect unverifiable code, we will lazily check
9166 // canSkipMethodVerification() to see if verification is REALLY needed.
9167 BOOL tiVerificationNeeded;
9169 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
9170 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
9171 BOOL tiIsVerifiableCode;
9173 // Set to TRUE if runtime callout is needed for this method
9174 BOOL tiRuntimeCalloutNeeded;
9176 // Set to TRUE if security prolog/epilog callout is needed for this method
9177 // Note: This flag is different than compNeedSecurityCheck.
9178 // compNeedSecurityCheck means whether or not a security object needs
9179 // to be allocated on the stack, which is currently true for EnC as well.
9180 // tiSecurityCalloutNeeded means whether or not security callouts need
9181 // to be inserted in the jitted code.
9182 BOOL tiSecurityCalloutNeeded;
9184 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
9185 // This support is necessary to suport attributes that are not described in
9186 // for example, signatures. For example, the permanent home byref (byref that
9187 // points to the gc heap), isn't a property of method signatures, therefore,
9188 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
9189 // but when deciding if we need to reimport a block, we need to take these
9191 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
9193 // Returns TRUE if child is equal to or a subtype of parent.
9194 // normalisedForStack indicates that both types are normalised for the stack
9195 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
9197 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
9198 // *pDest is modified to represent the merged type. Sets "*changed" to true
9199 // if this changes "*pDest".
9200 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
9202 // Set pDest from the primitive value type.
9203 // Eg. System.Int32 -> ELEMENT_TYPE_I4
9205 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
9208 // <BUGNUM> VSW 471305
9209 // IJW allows assigning REF to BYREF. The following allows us to temporarily
9210 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
9211 // We use a "short" as we need to push/pop this scope.
9213 short compRegSetCheckLevel;
9217 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9218 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9220 XX IL verification stuff XX
9223 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9224 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9228 // The following is used to track liveness of local variables, initialization
9229 // of valueclass constructors, and type safe use of IL instructions.
9231 // dynamic state info needed for verification
9232 EntryState verCurrentState;
9234 // this ptr of object type .ctors are considered intited only after
9235 // the base class ctor is called, or an alternate ctor is called.
9236 // An uninited this ptr can be used to access fields, but cannot
9237 // be used to call a member function.
9238 BOOL verTrackObjCtorInitState;
9240 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
9242 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
9243 void verSetThisInit(BasicBlock* block, ThisInitState tis);
9244 void verInitCurrentState();
9245 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
9247 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
9248 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
9249 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
9251 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
9252 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
9253 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
9254 bool bashStructToRef = false); // converts from jit type representation to typeInfo
9255 typeInfo verMakeTypeInfo(CorInfoType ciType,
9256 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
9257 BOOL verIsSDArray(typeInfo ti);
9258 typeInfo verGetArrayElemType(typeInfo ti);
9260 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
9261 BOOL verNeedsVerification();
9262 BOOL verIsByRefLike(const typeInfo& ti);
9263 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
9265 // generic type variables range over types that satisfy IsBoxable
9266 BOOL verIsBoxable(const typeInfo& ti);
9268 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
9269 DEBUGARG(unsigned line));
9270 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
9271 DEBUGARG(unsigned line));
9272 bool verCheckTailCallConstraint(OPCODE opcode,
9273 CORINFO_RESOLVED_TOKEN* pResolvedToken,
9274 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
9275 // on a type parameter?
9276 bool speculative // If true, won't throw if verificatoin fails. Instead it will
9277 // return false to the caller.
9278 // If false, it will throw.
9280 bool verIsBoxedValueType(typeInfo ti);
9282 void verVerifyCall(OPCODE opcode,
9283 CORINFO_RESOLVED_TOKEN* pResolvedToken,
9284 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
9286 bool readonlyCall, // is this a "readonly." call?
9287 const BYTE* delegateCreateStart,
9288 const BYTE* codeAddr,
9289 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
9291 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
9293 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
9294 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
9295 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
9296 const CORINFO_FIELD_INFO& fieldInfo,
9297 const typeInfo* tiThis,
9299 BOOL allowPlainStructAsThis = FALSE);
9300 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
9301 void verVerifyThisPtrInitialised();
9302 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
9304 // Register allocator
9305 void raInitStackFP();
9306 void raEnregisterVarsPrePassStackFP();
9307 void raSetRegLclBirthDeath(GenTree* tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
9308 void raEnregisterVarsPostPassStackFP();
9309 void raGenerateFPRefCounts();
9310 void raEnregisterVarsStackFP();
9311 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
9313 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
9314 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
9316 // returns true if enregistering v1 would save more mem accesses than v2
9317 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
9320 void raDumpHeightsStackFP();
9321 void raDumpVariableRegIntfFloat();
9324 #if FEATURE_STACK_FP_X87
9326 // Currently, we use FP transition blocks in only 2 situations:
9328 // -conditional jump on longs where FP stack differs with target: it's not strictly
9329 // necessary, but its low frequency and the code would get complicated if we try to
9330 // inline the FP stack adjustment, as we have a lot of special casing going on to try
9331 // minimize the way we generate the jump code.
9332 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
9333 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
9335 // However, transition blocks have 2 problems
9337 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9338 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9339 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9340 // in the right place without preordering them), this causes us to have to generate the transition
9341 // blocks in the cold area if we want procedure splitting.
9344 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9345 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9346 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9347 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9348 // a big change in the exception.
9350 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9351 // optimizations. For these 2 cases:
9353 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9354 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9355 // a switch statement.
9357 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9358 // current procedure splitting and exception code have.
9359 bool compMayHaveTransitionBlocks;
9361 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9363 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9365 unsigned raCntStkStackFP;
9366 unsigned raCntWtdStkDblStackFP;
9367 unsigned raCntStkParamDblStackFP;
9369 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9370 // TODO: Do we want to put this in LclVarDsc?
9371 unsigned raPayloadStackFP[lclMAX_TRACKED];
9372 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9374 // Useful for debugging
9375 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9377 #endif // FEATURE_STACK_FP_X87
9380 // One line log function. Default level is 0. Increasing it gives you
9381 // more log information
9383 // levels are currently unused: #define JITDUMP(level,...) ();
9384 void JitLogEE(unsigned level, const char* fmt, ...);
9386 bool compDebugBreak;
9388 bool compJitHaltMethod();
9393 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9394 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9396 XX GS Security checks for unsafe buffers XX
9398 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9399 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9402 struct ShadowParamVarInfo
9404 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9405 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9407 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9409 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9410 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9411 // slots and update all trees to refer to shadow slots is done immediately after
9412 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9413 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9414 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9415 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9416 // creating a shadow slot even though this routine returns true.
9418 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9419 // required. There are two cases under which a reg arg could potentially be used from its
9421 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9422 // b) LSRA spills it
9424 // Possible solution to address case (a)
9425 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9426 // in this routine. Note that live out of exception handler is something we may not be
9427 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9428 // Therefore, for methods with exception handling and need GS cookie check we might have
9429 // to take conservative approach.
9431 // Possible solution to address case (b)
9432 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9433 // create a new spill temp if the method needs GS cookie check.
9434 return varDsc->lvIsParam;
9435 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9436 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9443 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9448 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9449 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9450 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9452 void gsGSChecksInitCookie(); // Grabs cookie variable
9453 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9454 bool gsFindVulnerableParams(); // Shadow param analysis code
9455 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9457 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9458 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9460 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9461 // This can be overwritten by setting complus_JITInlineSize env variable.
9463 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9465 #define DEFAULT_MAX_LOCALLOC_TO_LOCAL_SIZE 32 // fixed locallocs of this size or smaller will convert to local buffers
9468 #ifdef FEATURE_JIT_METHOD_PERF
9469 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9470 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9472 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9473 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9475 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9477 #if MEASURE_CLRAPI_CALLS
9478 // Thin wrappers that call into JitTimer (if present).
9479 inline void CLRApiCallEnter(unsigned apix);
9480 inline void CLRApiCallLeave(unsigned apix);
9483 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9484 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9489 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9490 // These variables are associated with maintaining SQM data about compile time.
9491 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9492 // in the current compilation.
9493 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9494 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9495 // the inlining phase in the current compilation.
9496 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9498 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9499 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9500 // type-loading and class initialization).
9501 void RecordStateAtEndOfInlining();
9502 // Assumes being called at the end of compilation. Update the SQM state.
9503 void RecordStateAtEndOfCompilation();
9505 #ifdef FEATURE_CLRSQM
9506 // Does anything SQM related necessary at process shutdown time.
9507 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9508 #endif // FEATURE_CLRSQM
9511 #if FUNC_INFO_LOGGING
9512 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9513 // filename to write it to.
9514 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9515 #endif // FUNC_INFO_LOGGING
9517 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9519 // Is the compilation in a full trust context?
9520 bool compIsFullTrust();
9523 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9524 #endif // MEASURE_NOWAY
9526 #ifndef FEATURE_TRACELOGGING
9527 // Should we actually fire the noway assert body and the exception handler?
9528 bool compShouldThrowOnNoway();
9529 #else // FEATURE_TRACELOGGING
9530 // Should we actually fire the noway assert body and the exception handler?
9531 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9533 // Telemetry instance to use per method compilation.
9534 JitTelemetry compJitTelemetry;
9536 // Get common parameters that have to be logged with most telemetry data.
9537 void compGetTelemetryDefaults(const char** assemblyName,
9538 const char** scopeName,
9539 const char** methodName,
9540 unsigned* methodHash);
9541 #endif // !FEATURE_TRACELOGGING
9545 NodeToTestDataMap* m_nodeTestData;
9547 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9548 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9549 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9550 // Current kept in this.
9552 NodeToTestDataMap* GetNodeTestData()
9554 Compiler* compRoot = impInlineRoot();
9555 if (compRoot->m_nodeTestData == nullptr)
9557 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9559 return compRoot->m_nodeTestData;
9562 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, int> NodeToIntMap;
9564 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9565 // currently occur in the AST graph.
9566 NodeToIntMap* FindReachableNodesInNodeTestData();
9568 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9569 // test data, associate that data with "to".
9570 void TransferTestDataToNode(GenTree* from, GenTree* to);
9572 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9573 // have annotations, attach similar annotations to the corresponding nodes in "to".
9574 void CopyTestDataToCloneTree(GenTree* from, GenTree* to);
9576 // These are the methods that test that the various conditions implied by the
9577 // test attributes are satisfied.
9578 void JitTestCheckSSA(); // SSA builder tests.
9579 void JitTestCheckVN(); // Value numbering tests.
9582 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9584 FieldSeqStore* m_fieldSeqStore;
9586 FieldSeqStore* GetFieldSeqStore()
9588 Compiler* compRoot = impInlineRoot();
9589 if (compRoot->m_fieldSeqStore == nullptr)
9591 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9592 CompAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9593 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9595 return compRoot->m_fieldSeqStore;
9598 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, FieldSeqNode*> NodeToFieldSeqMap;
9600 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9601 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9602 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9603 // attach the field sequence directly to the address node.
9604 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9606 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9608 // Don't need to worry about inlining here
9609 if (m_zeroOffsetFieldMap == nullptr)
9611 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9613 CompAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9614 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9616 return m_zeroOffsetFieldMap;
9619 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9620 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9621 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9622 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9623 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9624 // record the the field sequence using the ZeroOffsetFieldMap described above.
9626 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9627 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9628 // CoreRT. Such case is handled same as the default case.
9629 void fgAddFieldSeqForZeroOffset(GenTree* op1, FieldSeqNode* fieldSeq);
9631 typedef JitHashTable<const GenTree*, JitPtrKeyFuncs<GenTree>, ArrayInfo> NodeToArrayInfoMap;
9632 NodeToArrayInfoMap* m_arrayInfoMap;
9634 NodeToArrayInfoMap* GetArrayInfoMap()
9636 Compiler* compRoot = impInlineRoot();
9637 if (compRoot->m_arrayInfoMap == nullptr)
9639 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9640 CompAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9641 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9643 return compRoot->m_arrayInfoMap;
9646 //-----------------------------------------------------------------------------------------------------------------
9647 // Compiler::TryGetArrayInfo:
9648 // Given an indirection node, checks to see whether or not that indirection represents an array access, and
9649 // if so returns information about the array.
9652 // indir - The `GT_IND` node.
9653 // arrayInfo (out) - Information about the accessed array if this function returns true. Undefined otherwise.
9656 // True if the `GT_IND` node represents an array access; false otherwise.
9657 inline bool TryGetArrayInfo(GenTreeIndir* indir, ArrayInfo* arrayInfo)
9659 if ((indir->gtFlags & GTF_IND_ARR_INDEX) == 0)
9664 if (indir->gtOp1->OperIs(GT_INDEX_ADDR))
9666 GenTreeIndexAddr* const indexAddr = indir->gtOp1->AsIndexAddr();
9667 *arrayInfo = ArrayInfo(indexAddr->gtElemType, indexAddr->gtElemSize, indexAddr->gtElemOffset,
9668 indexAddr->gtStructElemClass);
9672 bool found = GetArrayInfoMap()->Lookup(indir, arrayInfo);
9677 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9679 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9680 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9681 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9682 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9684 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9686 // Use the same map for GCHeap and ByrefExposed when their states match.
9687 memoryKind = ByrefExposed;
9690 assert(memoryKind < MemoryKindCount);
9691 Compiler* compRoot = impInlineRoot();
9692 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9694 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9695 CompAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9696 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9698 return compRoot->m_memorySsaMap[memoryKind];
9701 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9702 CORINFO_CLASS_HANDLE m_refAnyClass;
9703 CORINFO_FIELD_HANDLE GetRefanyDataField()
9705 if (m_refAnyClass == nullptr)
9707 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9709 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9711 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9713 if (m_refAnyClass == nullptr)
9715 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9717 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9721 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9723 #if ALLVARSET_COUNTOPS
9724 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9727 static HelperCallProperties s_helperCallProperties;
9729 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9730 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9731 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9734 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9737 unsigned __int8* offset0,
9738 unsigned __int8* offset1);
9740 void GetStructTypeOffset(CORINFO_CLASS_HANDLE typeHnd,
9743 unsigned __int8* offset0,
9744 unsigned __int8* offset1);
9746 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9747 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9749 void fgMorphMultiregStructArgs(GenTreeCall* call);
9750 GenTree* fgMorphMultiregStructArg(GenTree* arg, fgArgTabEntry* fgEntryPtr);
9752 bool killGCRefs(GenTree* tree);
9754 }; // end of class Compiler
9756 // Inline methods of CompAllocator.
9757 void* CompAllocator::Alloc(size_t sz)
9759 #if MEASURE_MEM_ALLOC
9760 return m_comp->compGetMem(sz, m_cmk);
9762 return m_comp->compGetMem(sz);
9766 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9768 #if MEASURE_MEM_ALLOC
9769 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9771 return m_comp->compGetMemArray(elems, elemSize);
9775 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9776 inline LclVarDsc::LclVarDsc(Compiler* comp)
9777 : // Initialize the ArgRegs to REG_STK.
9778 // The morph will do the right thing to change
9779 // to the right register if passed in register.
9782 #if FEATURE_MULTIREG_ARGS
9783 _lvOtherArgReg(REG_STK)
9785 #endif // FEATURE_MULTIREG_ARGS
9787 lvRefBlks(BlockSetOps::UninitVal())
9789 #endif // ASSERTION_PROP
9790 lvPerSsaData(comp->getAllocator())
9794 //---------------------------------------------------------------------------------------------------------------------
9795 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9797 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9798 // shown in parentheses):
9800 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9801 // of a misnomer, as the first entry will always be the current node.
9803 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9804 // argument before visiting the node's operands.
9806 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9807 // argument after visiting the node's operands.
9809 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9810 // `DoPreOrder` must be true if this option is true.
9812 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9813 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9814 // visited before the first).
9816 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9818 // A simple pre-order visitor might look something like the following:
9820 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9825 // DoPreOrder = true
9828 // unsigned m_count;
9830 // CountingVisitor(Compiler* compiler)
9831 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9835 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9841 // This visitor would then be used like so:
9843 // CountingVisitor countingVisitor(compiler);
9844 // countingVisitor.WalkTree(root);
9846 template <typename TVisitor>
9847 class GenTreeVisitor
9850 typedef Compiler::fgWalkResult fgWalkResult;
9854 ComputeStack = false,
9856 DoPostOrder = false,
9857 DoLclVarsOnly = false,
9858 UseExecutionOrder = false,
9861 Compiler* m_compiler;
9862 ArrayStack<GenTree*> m_ancestors;
9864 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9866 assert(compiler != nullptr);
9868 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9869 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9872 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9874 return fgWalkResult::WALK_CONTINUE;
9877 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9879 return fgWalkResult::WALK_CONTINUE;
9883 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9885 assert(use != nullptr);
9887 GenTree* node = *use;
9889 if (TVisitor::ComputeStack)
9891 m_ancestors.Push(node);
9894 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9895 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9897 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9898 if (result == fgWalkResult::WALK_ABORT)
9904 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9910 switch (node->OperGet())
9915 case GT_LCL_VAR_ADDR:
9916 case GT_LCL_FLD_ADDR:
9917 if (TVisitor::DoLclVarsOnly)
9919 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9920 if (result == fgWalkResult::WALK_ABORT)
9936 case GT_MEMORYBARRIER:
9941 case GT_START_NONGC:
9943 #if !FEATURE_EH_FUNCLETS
9945 #endif // !FEATURE_EH_FUNCLETS
9947 #ifndef LEGACY_BACKEND
9949 #endif // LEGACY_BACKEND
9952 case GT_CLS_VAR_ADDR:
9956 case GT_PINVOKE_PROLOG:
9957 case GT_PINVOKE_EPILOG:
9961 // Lclvar unary operators
9962 case GT_STORE_LCL_VAR:
9963 case GT_STORE_LCL_FLD:
9964 if (TVisitor::DoLclVarsOnly)
9966 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9967 if (result == fgWalkResult::WALK_ABORT)
9974 // Standard unary operators
10001 case GT_RUNTIMELOOKUP:
10003 GenTreeUnOp* const unOp = node->AsUnOp();
10004 if (unOp->gtOp1 != nullptr)
10006 result = WalkTree(&unOp->gtOp1, unOp);
10007 if (result == fgWalkResult::WALK_ABORT)
10018 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
10020 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
10021 if (result == fgWalkResult::WALK_ABORT)
10025 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
10026 if (result == fgWalkResult::WALK_ABORT)
10030 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
10031 if (result == fgWalkResult::WALK_ABORT)
10038 case GT_ARR_BOUNDS_CHECK:
10039 #ifdef FEATURE_SIMD
10041 #endif // FEATURE_SIMD
10042 #ifdef FEATURE_HW_INTRINSICS
10043 case GT_HW_INTRINSIC_CHK:
10044 #endif // FEATURE_HW_INTRINSICS
10046 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
10048 result = WalkTree(&boundsChk->gtIndex, boundsChk);
10049 if (result == fgWalkResult::WALK_ABORT)
10053 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
10054 if (result == fgWalkResult::WALK_ABORT)
10063 GenTreeField* const field = node->AsField();
10065 if (field->gtFldObj != nullptr)
10067 result = WalkTree(&field->gtFldObj, field);
10068 if (result == fgWalkResult::WALK_ABORT)
10078 GenTreeArrElem* const arrElem = node->AsArrElem();
10080 result = WalkTree(&arrElem->gtArrObj, arrElem);
10081 if (result == fgWalkResult::WALK_ABORT)
10086 const unsigned rank = arrElem->gtArrRank;
10087 for (unsigned dim = 0; dim < rank; dim++)
10089 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
10090 if (result == fgWalkResult::WALK_ABORT)
10098 case GT_ARR_OFFSET:
10100 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
10102 result = WalkTree(&arrOffs->gtOffset, arrOffs);
10103 if (result == fgWalkResult::WALK_ABORT)
10107 result = WalkTree(&arrOffs->gtIndex, arrOffs);
10108 if (result == fgWalkResult::WALK_ABORT)
10112 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
10113 if (result == fgWalkResult::WALK_ABORT)
10122 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
10124 GenTree** op1Use = &dynBlock->gtOp1;
10125 GenTree** op2Use = &dynBlock->gtDynamicSize;
10127 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
10129 std::swap(op1Use, op2Use);
10132 result = WalkTree(op1Use, dynBlock);
10133 if (result == fgWalkResult::WALK_ABORT)
10137 result = WalkTree(op2Use, dynBlock);
10138 if (result == fgWalkResult::WALK_ABORT)
10145 case GT_STORE_DYN_BLK:
10147 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
10149 GenTree** op1Use = &dynBlock->gtOp1;
10150 GenTree** op2Use = &dynBlock->gtOp2;
10151 GenTree** op3Use = &dynBlock->gtDynamicSize;
10153 if (TVisitor::UseExecutionOrder)
10155 if (dynBlock->IsReverseOp())
10157 std::swap(op1Use, op2Use);
10159 if (dynBlock->gtEvalSizeFirst)
10161 std::swap(op3Use, op2Use);
10162 std::swap(op2Use, op1Use);
10166 result = WalkTree(op1Use, dynBlock);
10167 if (result == fgWalkResult::WALK_ABORT)
10171 result = WalkTree(op2Use, dynBlock);
10172 if (result == fgWalkResult::WALK_ABORT)
10176 result = WalkTree(op3Use, dynBlock);
10177 if (result == fgWalkResult::WALK_ABORT)
10186 GenTreeCall* const call = node->AsCall();
10188 if (call->gtCallObjp != nullptr)
10190 result = WalkTree(&call->gtCallObjp, call);
10191 if (result == fgWalkResult::WALK_ABORT)
10197 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
10199 result = WalkTree(args->pCurrent(), call);
10200 if (result == fgWalkResult::WALK_ABORT)
10206 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
10208 result = WalkTree(args->pCurrent(), call);
10209 if (result == fgWalkResult::WALK_ABORT)
10215 if (call->gtCallType == CT_INDIRECT)
10217 if (call->gtCallCookie != nullptr)
10219 result = WalkTree(&call->gtCallCookie, call);
10220 if (result == fgWalkResult::WALK_ABORT)
10226 result = WalkTree(&call->gtCallAddr, call);
10227 if (result == fgWalkResult::WALK_ABORT)
10233 if (call->gtControlExpr != nullptr)
10235 result = WalkTree(&call->gtControlExpr, call);
10236 if (result == fgWalkResult::WALK_ABORT)
10248 assert(node->OperIsBinary());
10250 GenTreeOp* const op = node->AsOp();
10252 GenTree** op1Use = &op->gtOp1;
10253 GenTree** op2Use = &op->gtOp2;
10255 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
10257 std::swap(op1Use, op2Use);
10260 if (*op1Use != nullptr)
10262 result = WalkTree(op1Use, op);
10263 if (result == fgWalkResult::WALK_ABORT)
10269 if (*op2Use != nullptr)
10271 result = WalkTree(op2Use, op);
10272 if (result == fgWalkResult::WALK_ABORT)
10282 // Finally, visit the current node
10283 if (TVisitor::DoPostOrder)
10285 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
10288 if (TVisitor::ComputeStack)
10297 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
10298 class GenericTreeWalker final
10299 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
10304 ComputeStack = computeStack,
10305 DoPreOrder = doPreOrder,
10306 DoPostOrder = doPostOrder,
10307 DoLclVarsOnly = doLclVarsOnly,
10308 UseExecutionOrder = useExecutionOrder,
10312 Compiler::fgWalkData* m_walkData;
10315 GenericTreeWalker(Compiler::fgWalkData* walkData)
10316 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
10317 walkData->compiler)
10318 , m_walkData(walkData)
10320 assert(walkData != nullptr);
10324 walkData->parentStack = &this->m_ancestors;
10328 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
10330 m_walkData->parent = user;
10331 return m_walkData->wtprVisitorFn(use, m_walkData);
10334 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
10336 m_walkData->parent = user;
10337 return m_walkData->wtpoVisitorFn(use, m_walkData);
10341 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
10347 DoLclVarsOnly = true
10350 IncLclVarRefCountsVisitor(Compiler* compiler);
10351 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
10353 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10356 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
10362 DoLclVarsOnly = true
10365 DecLclVarRefCountsVisitor(Compiler* compiler);
10366 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
10368 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10372 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10373 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10375 XX Miscellaneous Compiler stuff XX
10377 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10378 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10381 // Values used to mark the types a stack slot is used for
10383 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10384 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10385 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10386 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10387 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10388 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10389 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10390 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10392 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10394 /*****************************************************************************
10396 * Variables to keep track of total code amounts.
10401 extern size_t grossVMsize;
10402 extern size_t grossNCsize;
10403 extern size_t totalNCsize;
10405 extern unsigned genMethodICnt;
10406 extern unsigned genMethodNCnt;
10407 extern size_t gcHeaderISize;
10408 extern size_t gcPtrMapISize;
10409 extern size_t gcHeaderNSize;
10410 extern size_t gcPtrMapNSize;
10412 #endif // DISPLAY_SIZES
10414 /*****************************************************************************
10416 * Variables to keep track of basic block counts (more data on 1 BB methods)
10419 #if COUNT_BASIC_BLOCKS
10420 extern Histogram bbCntTable;
10421 extern Histogram bbOneBBSizeTable;
10424 /*****************************************************************************
10426 * Used by optFindNaturalLoops to gather statistical information such as
10427 * - total number of natural loops
10428 * - number of loops with 1, 2, ... exit conditions
10429 * - number of loops that have an iterator (for like)
10430 * - number of loops that have a constant iterator
10435 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10436 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10437 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10438 extern unsigned totalLoopCount; // counts the total number of natural loops
10439 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10440 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10441 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10442 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10444 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10445 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10446 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10447 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10448 extern Histogram loopCountTable; // Histogram of loop counts
10449 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10451 #endif // COUNT_LOOPS
10453 /*****************************************************************************
10454 * variables to keep track of how many iterations we go in a dataflow pass
10459 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10460 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10462 #endif // DATAFLOW_ITER
10464 #if MEASURE_BLOCK_SIZE
10465 extern size_t genFlowNodeSize;
10466 extern size_t genFlowNodeCnt;
10467 #endif // MEASURE_BLOCK_SIZE
10469 #if MEASURE_NODE_SIZE
10470 struct NodeSizeStats
10474 genTreeNodeCnt = 0;
10475 genTreeNodeSize = 0;
10476 genTreeNodeActualSize = 0;
10479 // Count of tree nodes allocated.
10480 unsigned __int64 genTreeNodeCnt;
10482 // The size we allocate.
10483 unsigned __int64 genTreeNodeSize;
10485 // The actual size of the node. Note that the actual size will likely be smaller
10486 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10487 // a smaller node to a larger one. TODO-Cleanup: add stats on
10488 // SetOper()/ChangeOper() usage to quantify this.
10489 unsigned __int64 genTreeNodeActualSize;
10491 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10492 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10493 extern Histogram genTreeNcntHist;
10494 extern Histogram genTreeNsizHist;
10495 #endif // MEASURE_NODE_SIZE
10497 /*****************************************************************************
10498 * Count fatal errors (including noway_asserts).
10502 extern unsigned fatal_badCode;
10503 extern unsigned fatal_noWay;
10504 extern unsigned fatal_NOMEM;
10505 extern unsigned fatal_noWayAssertBody;
10507 extern unsigned fatal_noWayAssertBodyArgs;
10509 extern unsigned fatal_NYI;
10510 #endif // MEASURE_FATAL
10512 /*****************************************************************************
10516 #ifdef _TARGET_XARCH_
10518 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10519 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10520 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10522 const instruction INS_AND = INS_and;
10523 const instruction INS_OR = INS_or;
10524 const instruction INS_XOR = INS_xor;
10525 const instruction INS_NEG = INS_neg;
10526 const instruction INS_TEST = INS_test;
10527 const instruction INS_MUL = INS_imul;
10528 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10529 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10530 const instruction INS_BREAKPOINT = INS_int3;
10531 const instruction INS_ADDC = INS_adc;
10532 const instruction INS_SUBC = INS_sbb;
10533 const instruction INS_NOT = INS_not;
10535 #endif // _TARGET_XARCH_
10537 #ifdef _TARGET_ARM_
10539 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10540 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10541 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10543 const instruction INS_AND = INS_and;
10544 const instruction INS_OR = INS_orr;
10545 const instruction INS_XOR = INS_eor;
10546 const instruction INS_NEG = INS_rsb;
10547 const instruction INS_TEST = INS_tst;
10548 const instruction INS_MUL = INS_mul;
10549 const instruction INS_MULADD = INS_mla;
10550 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10551 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10552 const instruction INS_BREAKPOINT = INS_bkpt;
10553 const instruction INS_ADDC = INS_adc;
10554 const instruction INS_SUBC = INS_sbc;
10555 const instruction INS_NOT = INS_mvn;
10557 const instruction INS_ABS = INS_vabs;
10558 const instruction INS_SQRT = INS_vsqrt;
10560 #endif // _TARGET_ARM_
10562 #ifdef _TARGET_ARM64_
10564 const instruction INS_MULADD = INS_madd;
10565 const instruction INS_BREAKPOINT = INS_bkpt;
10567 const instruction INS_ABS = INS_fabs;
10568 const instruction INS_SQRT = INS_fsqrt;
10570 #endif // _TARGET_ARM64_
10572 /*****************************************************************************/
10574 extern const BYTE genTypeSizes[];
10575 extern const BYTE genTypeAlignments[];
10576 extern const BYTE genTypeStSzs[];
10577 extern const BYTE genActualTypes[];
10579 /*****************************************************************************/
10581 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10582 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10584 #ifdef _TARGET_ARM_
10585 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10586 #elif defined(_TARGET_ARM64_)
10587 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10590 /*****************************************************************************/
10592 #define REG_CORRUPT regNumber(REG_NA + 1)
10593 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10594 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10596 /*****************************************************************************/
10598 extern BasicBlock dummyBB;
10600 /*****************************************************************************/
10601 /*****************************************************************************/
10603 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10604 // nodes of a statement in execution order.
10605 // __stmt: a GT_STMT type GenTree*
10606 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10608 #define foreach_treenode_execution_order(__node, __stmt) \
10609 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10611 // foreach_block: An iterator over all blocks in the function.
10612 // __compiler: the Compiler* object
10613 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10615 #define foreach_block(__compiler, __block) \
10616 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10618 /*****************************************************************************/
10619 /*****************************************************************************/
10623 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10625 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10626 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10628 XX Debugging helpers XX
10630 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10631 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10634 /*****************************************************************************/
10635 /* The following functions are intended to be called from the debugger, to dump
10636 * various data structures. The can be used in the debugger Watch or Quick Watch
10637 * windows. They are designed to be short to type and take as few arguments as
10638 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10639 * See the function definition comment for more details.
10642 void cBlock(Compiler* comp, BasicBlock* block);
10643 void cBlocks(Compiler* comp);
10644 void cBlocksV(Compiler* comp);
10645 void cTree(Compiler* comp, GenTree* tree);
10646 void cTrees(Compiler* comp);
10647 void cEH(Compiler* comp);
10648 void cVar(Compiler* comp, unsigned lclNum);
10649 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10650 void cVars(Compiler* comp);
10651 void cVarsFinal(Compiler* comp);
10652 void cBlockPreds(Compiler* comp, BasicBlock* block);
10653 void cReach(Compiler* comp);
10654 void cDoms(Compiler* comp);
10655 void cLiveness(Compiler* comp);
10656 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10658 void cFuncIR(Compiler* comp);
10659 void cBlockIR(Compiler* comp, BasicBlock* block);
10660 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10661 void cTreeIR(Compiler* comp, GenTree* tree);
10662 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10663 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10664 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10665 int cOperandIR(Compiler* comp, GenTree* operand);
10666 int cLeafIR(Compiler* comp, GenTree* tree);
10667 int cIndirIR(Compiler* comp, GenTree* tree);
10668 int cListIR(Compiler* comp, GenTree* list);
10669 int cSsaNumIR(Compiler* comp, GenTree* tree);
10670 int cValNumIR(Compiler* comp, GenTree* tree);
10671 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10673 void dBlock(BasicBlock* block);
10676 void dTree(GenTree* tree);
10679 void dVar(unsigned lclNum);
10680 void dVarDsc(LclVarDsc* varDsc);
10683 void dBlockPreds(BasicBlock* block);
10687 void dCVarSet(VARSET_VALARG_TP vars);
10689 void dVarSet(VARSET_VALARG_TP vars);
10690 void dRegMask(regMaskTP mask);
10693 void dBlockIR(BasicBlock* block);
10694 void dTreeIR(GenTree* tree);
10695 void dLoopIR(Compiler::LoopDsc* loop);
10696 void dLoopNumIR(unsigned loopNum);
10697 int dTabStopIR(int curr, int tabstop);
10698 int dTreeTypeIR(GenTree* tree);
10699 int dTreeKindsIR(GenTree* tree);
10700 int dTreeFlagsIR(GenTree* tree);
10701 int dOperandIR(GenTree* operand);
10702 int dLeafIR(GenTree* tree);
10703 int dIndirIR(GenTree* tree);
10704 int dListIR(GenTree* list);
10705 int dSsaNumIR(GenTree* tree);
10706 int dValNumIR(GenTree* tree);
10707 int dDependsIR(GenTree* comma);
10710 GenTree* dFindTree(GenTree* tree, unsigned id);
10711 GenTree* dFindTree(unsigned id);
10712 GenTreeStmt* dFindStmt(unsigned id);
10713 BasicBlock* dFindBlock(unsigned bbNum);
10717 #include "compiler.hpp" // All the shared inline functions
10719 /*****************************************************************************/
10720 #endif //_COMPILER_H_
10721 /*****************************************************************************/