1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
104 // Or the more-general IAllocator interface.
105 void* __cdecl operator new(size_t n, IAllocator* alloc);
106 void* __cdecl operator new[](size_t n, IAllocator* alloc);
108 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
109 // caused these to be ambiguous with the global placement new operators.
110 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
111 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
114 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
115 #include "loopcloning.h"
117 /*****************************************************************************/
119 /* This is included here and not earlier as it needs the definition of "CSE"
120 * which is defined in the section above */
122 /*****************************************************************************/
124 unsigned genLog2(unsigned value);
125 unsigned genLog2(unsigned __int64 value);
127 var_types genActualType(var_types type);
128 var_types genUnsignedType(var_types type);
129 var_types genSignedType(var_types type);
131 unsigned ReinterpretHexAsDecimal(unsigned);
133 /*****************************************************************************/
136 #ifdef FEATURE_AVX_SUPPORT
137 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
138 #else // !FEATURE_AVX_SUPPORT
139 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
140 #endif // !FEATURE_AVX_SUPPORT
141 #else // !FEATURE_SIMD
142 const unsigned TEMP_MAX_SIZE = sizeof(double);
143 #endif // !FEATURE_SIMD
144 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
146 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
149 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
152 // The following holds the Local var info (scope information)
153 typedef const char* VarName; // Actual ASCII string
156 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
157 IL_OFFSET vsdLifeEnd; // instr offset of end of life
158 unsigned vsdVarNum; // (remapped) LclVarDsc number
161 VarName vsdName; // name of the var
164 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
165 // Also, it is the index of this entry in the info.compVarScopes array,
166 // which is useful since the array is also accessed via the
167 // compEnterScopeList and compExitScopeList sorted arrays.
170 /*****************************************************************************
172 * The following holds the local variable counts and the descriptor table.
175 // This is the location of a definition.
181 DefLoc() : m_blk(nullptr), m_tree(nullptr)
186 // This class encapsulates all info about a local variable that may vary for different SSA names
191 ValueNumPair m_vnPair;
199 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
204 // The constructor. Most things can just be zero'ed.
205 LclVarDsc(Compiler* comp);
207 // note this only packs because var_types is a typedef of unsigned char
208 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
210 unsigned char lvIsParam : 1; // is this a parameter?
211 unsigned char lvIsRegArg : 1; // is this a register argument?
212 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
214 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
215 // is to help determine whether to use block init in the prolog.
216 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
217 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
218 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
219 // variable is in the same register for the entire function.
220 unsigned char lvTracked : 1; // is this a tracked variable?
221 bool lvTrackedNonStruct()
223 return lvTracked && lvType != TYP_STRUCT;
225 unsigned char lvPinned : 1; // is this a pinned variable?
227 unsigned char lvMustInit : 1; // must be initialized
228 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
229 // global location, etc.
230 // We cannot reason reliably about the value of the variable.
231 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
232 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
236 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
238 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
239 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
240 // read/write its value.
241 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
242 // the variable to be
243 // in the stack (at least at those boundaries.)
244 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
245 // particular byte from an int).
246 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
247 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
249 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
250 unsigned char lvRefAssign : 1; // involved in pointer assignment
251 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
252 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
255 unsigned char lvHasILStoreOp : 1; // there is at least one STLOC or STARG on this local
256 unsigned char lvHasMultipleILStoreOp : 1; // there is more than one STLOC on this local
258 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp (if lvIsParam is false), or implicit byref parameter
259 // (if lvIsParam is true)
261 unsigned char lvIsBoolean : 1; // set if variable is boolean
263 unsigned char lvRngOptDone : 1; // considered for range check opt?
264 unsigned char lvLoopInc : 1; // incremented in the loop?
265 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
266 unsigned char lvArrIndx : 1; // used as an array index?
267 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
268 unsigned char lvArrIndxDom : 1; // index dominates loop exit
270 unsigned char lvSingleDef : 1; // variable has a single def
271 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
272 unsigned char lvVolatileHint : 1; // hint for AssertionProp
275 unsigned char lvSpilled : 1; // enregistered variable was spilled
276 #ifndef _TARGET_64BIT_
277 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
278 #endif // !_TARGET_64BIT_
279 #ifdef _TARGET_64BIT_
280 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
283 unsigned char lvKeepType : 1; // Don't change the type of this variable
284 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
286 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
288 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
289 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
290 // 32-bit target. For implicit byref parameters, this gets hijacked between
291 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to indicate whether
292 // references to the arg are being rewritten as references to a promoted shadow local.
293 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
294 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
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(lvType == TYP_STRUCT);
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 unsigned lvSlotNum; // original slot # (if remapped)
723 typeInfo lvVerTypeInfo; // type info needed for verification
725 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
727 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
729 BYTE* lvGcLayout; // GC layout info for structs
732 BlockSet lvRefBlks; // Set of blocks that contain refs
733 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
734 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
736 var_types TypeGet() const
738 return (var_types)lvType;
740 bool lvStackAligned() const
742 assert(lvIsStructField);
743 return ((lvFldOffset % sizeof(void*)) == 0);
745 bool lvNormalizeOnLoad() const
747 return varTypeIsSmall(TypeGet()) &&
748 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
749 (lvIsParam || lvAddrExposed || lvIsStructField);
752 bool lvNormalizeOnStore()
754 return varTypeIsSmall(TypeGet()) &&
755 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
756 !(lvIsParam || lvAddrExposed || lvIsStructField);
759 void lvaResetSortAgainFlag(Compiler* pComp);
760 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
761 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
762 void setPrefReg(regNumber regNum, Compiler* pComp);
763 void addPrefReg(regMaskTP regMask, Compiler* pComp);
764 bool IsFloatRegType() const
766 return isFloatRegType(lvType) || lvIsHfaRegArg();
768 var_types GetHfaType() const
770 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
772 void SetHfaType(var_types type)
774 assert(varTypeIsFloating(type));
775 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
778 #ifndef LEGACY_BACKEND
779 var_types lvaArgType();
782 PerSsaArray lvPerSsaData;
785 // Keep track of the # of SsaNames, for a bounds check.
786 unsigned lvNumSsaNames;
789 // Returns the address of the per-Ssa data for the given ssaNum (which is required
790 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
791 // not an SSA variable).
792 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
794 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
795 assert(SsaConfig::RESERVED_SSA_NUM == 0);
796 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
797 assert(zeroBased < lvNumSsaNames);
798 return &lvPerSsaData.GetRef(zeroBased);
803 void PrintVarReg() const
805 if (isRegPairType(TypeGet()))
807 printf("%s:%s", getRegName(lvOtherReg), // hi32
808 getRegName(lvRegNum)); // lo32
812 printf("%s", getRegName(lvRegNum));
817 }; // class LclVarDsc
820 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
821 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
825 XX The temporary lclVars allocated by the compiler for code generation XX
827 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
828 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
831 /*****************************************************************************
833 * The following keeps track of temporaries allocated in the stack frame
834 * during code-generation (after register allocation). These spill-temps are
835 * only used if we run out of registers while evaluating a tree.
837 * These are different from the more common temps allocated by lvaGrabTemp().
848 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
856 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
860 0); // temps must have a negative number (so they have a different number from all local variables)
861 tdOffs = BAD_TEMP_OFFSET;
865 IMPL_LIMITATION("too many spill temps");
870 bool tdLegalOffset() const
872 return tdOffs != BAD_TEMP_OFFSET;
876 int tdTempOffs() const
878 assert(tdLegalOffset());
881 void tdSetTempOffs(int offs)
884 assert(tdLegalOffset());
886 void tdAdjustTempOffs(int offs)
889 assert(tdLegalOffset());
892 int tdTempNum() const
897 unsigned tdTempSize() const
901 var_types tdTempType() const
907 // interface to hide linearscan implementation from rest of compiler
908 class LinearScanInterface
911 virtual void doLinearScan() = 0;
912 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
915 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
917 // Information about arrays: their element type and size, and the offset of the first element.
918 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
919 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
920 // for example, in value numbering of array index expressions.
923 var_types m_elemType;
924 CORINFO_CLASS_HANDLE m_elemStructType;
926 unsigned m_elemOffset;
928 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
932 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
933 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
938 // This enumeration names the phases into which we divide compilation. The phases should completely
939 // partition a compilation.
942 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
943 #include "compphases.h"
947 extern const char* PhaseNames[];
948 extern const char* PhaseEnums[];
949 extern const LPCWSTR PhaseShortNames[];
951 // The following enum provides a simple 1:1 mapping to CLR API's
952 enum API_ICorJitInfo_Names
954 #define DEF_CLR_API(name) API_##name,
955 #include "ICorJitInfo_API_names.h"
959 //---------------------------------------------------------------
963 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
964 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
965 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
966 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
967 // by "m_timerFailure" being true.
968 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
971 #ifdef FEATURE_JIT_METHOD_PERF
972 // The string names of the phases.
973 static const char* PhaseNames[];
975 static bool PhaseHasChildren[];
976 static int PhaseParent[];
977 static bool PhaseReportsIRSize[];
979 unsigned m_byteCodeBytes;
980 unsigned __int64 m_totalCycles;
981 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
982 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
983 #if MEASURE_CLRAPI_CALLS
984 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
985 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
988 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
990 // For better documentation, we call EndPhase on
991 // non-leaf phases. We should also call EndPhase on the
992 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
993 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
994 // We add all such "redundant end phase" intervals to this variable below; we print
995 // it out in a report, so we can verify that it is, indeed, very small. If it ever
996 // isn't, this means that we're doing something significant between the end of the last
997 // declared subphase and the end of its parent.
998 unsigned __int64 m_parentPhaseEndSlop;
1001 #if MEASURE_CLRAPI_CALLS
1002 // The following measures the time spent inside each individual CLR API call.
1003 unsigned m_allClrAPIcalls;
1004 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1005 unsigned __int64 m_allClrAPIcycles;
1006 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1007 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1008 #endif // MEASURE_CLRAPI_CALLS
1010 CompTimeInfo(unsigned byteCodeBytes);
1014 #ifdef FEATURE_JIT_METHOD_PERF
1016 #if MEASURE_CLRAPI_CALLS
1017 struct WrapICorJitInfo;
1020 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1021 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1022 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1023 // threads, so it is protected by a lock.
1024 // This class is intended to be used as a singleton type, with only a single instance.
1025 class CompTimeSummaryInfo
1027 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1028 static CritSecObject s_compTimeSummaryLock;
1032 CompTimeInfo m_total;
1033 CompTimeInfo m_maximum;
1035 int m_numFilteredMethods;
1036 CompTimeInfo m_filtered;
1038 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1039 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1040 // If any OS interaction fails, returns 0.0.
1041 double CyclesPerSecond();
1043 // This can use what ever data you want to determine if the value to be added
1044 // belongs in the filtered section (it's always included in the unfiltered section)
1045 bool IncludedInFilteredData(CompTimeInfo& info);
1048 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1049 static CompTimeSummaryInfo s_compTimeSummary;
1051 CompTimeSummaryInfo()
1052 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1056 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1057 // This is thread safe.
1058 void AddInfo(CompTimeInfo& info, bool includePhases);
1060 // Print the summary information to "f".
1061 // This is not thread-safe; assumed to be called by only one thread.
1062 void Print(FILE* f);
1065 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1066 // and when the current phase started. This is intended to be part of a Compilation object. This is
1067 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1071 unsigned __int64 m_start; // Start of the compilation.
1072 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1073 #if MEASURE_CLRAPI_CALLS
1074 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1075 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1076 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1077 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1078 static double s_cyclesPerSec; // Cached for speedier measurements
1081 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1083 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1085 static CritSecObject s_csvLock; // Lock to protect the time log file.
1086 void PrintCsvMethodStats(Compiler* comp);
1089 void* operator new(size_t);
1090 void* operator new[](size_t);
1091 void operator delete(void*);
1092 void operator delete[](void*);
1095 // Initialized the timer instance
1096 JitTimer(unsigned byteCodeSize);
1098 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1100 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1103 static void PrintCsvHeader();
1105 // Ends the current phase (argument is for a redundant check).
1106 void EndPhase(Compiler* compiler, Phases phase);
1108 #if MEASURE_CLRAPI_CALLS
1109 // Start and end a timed CLR API call.
1110 void CLRApiCallEnter(unsigned apix);
1111 void CLRApiCallLeave(unsigned apix);
1112 #endif // MEASURE_CLRAPI_CALLS
1114 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1115 // and adds it to "sum".
1116 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1118 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1119 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1120 // "m_info" to true.
1121 bool GetThreadCycles(unsigned __int64* cycles)
1123 bool res = CycleTimer::GetThreadCyclesS(cycles);
1126 m_info.m_timerFailure = true;
1131 #endif // FEATURE_JIT_METHOD_PERF
1133 //------------------- Function/Funclet info -------------------------------
1134 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1136 FUNC_ROOT, // The main/root function (always id==0)
1137 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1138 FUNC_FILTER, // a funclet associated with an EH filter
1141 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1148 BYTE funFlags; // Currently unused, just here for padding
1149 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1150 // funclet. It is only valid if funKind field indicates this is a
1151 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1153 #if defined(_TARGET_AMD64_)
1155 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1156 emitLocation* startLoc;
1157 emitLocation* endLoc;
1158 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1159 emitLocation* coldEndLoc;
1160 UNWIND_INFO unwindHeader;
1161 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1162 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1163 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1164 unsigned unwindCodeSlot;
1166 #ifdef UNIX_AMD64_ABI
1167 jitstd::vector<CFI_CODE>* cfiCodes;
1168 #endif // UNIX_AMD64_ABI
1170 #elif defined(_TARGET_ARMARCH_)
1172 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1173 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1174 // Note: we only have a pointer here instead of the actual object,
1175 // to save memory in the JIT case (compared to the NGEN case),
1176 // where we don't have any cold section.
1177 // Note 2: we currently don't support hot/cold splitting in functions
1178 // with EH, so uwiCold will be NULL for all funclets.
1180 #endif // _TARGET_ARMARCH_
1182 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1183 // that isn't shared between the main function body and funclets.
1186 struct fgArgTabEntry
1189 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1192 otherRegNum = REG_NA;
1193 isStruct = false; // is this a struct arg
1195 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1197 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1199 // it will point at the actual argument in the gtCallLateArgs list.
1200 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1202 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1204 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1206 unsigned numRegs; // Count of number of registers that this argument uses
1208 // A slot is a pointer sized region in the OutArg area.
1209 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1210 unsigned numSlots; // Count of number of slots that this argument uses
1212 unsigned alignment; // 1 or 2 (slots/registers)
1213 unsigned lateArgInx; // index into gtCallLateArgs list
1214 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1216 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1217 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1218 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1219 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1220 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1221 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1222 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1223 // previous arguments.
1224 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1225 // to be on the stack despite its arg list position.
1227 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1228 bool isStruct : 1; // True if this is a struct arg
1230 regNumber otherRegNum; // The (second) register to use when passing this argument.
1232 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1233 #elif !defined(_TARGET_64BIT_)
1234 __declspec(property(get = getIsStruct)) bool isStruct;
1237 return varTypeIsStruct(node);
1239 #endif // !_TARGET_64BIT_
1242 void SetIsHfaRegArg(bool hfaRegArg)
1244 isHfaRegArg = hfaRegArg;
1247 void SetIsBackFilled(bool backFilled)
1249 isBackFilled = backFilled;
1252 bool IsBackFilled() const
1254 return isBackFilled;
1256 #else // !_TARGET_ARM_
1257 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1259 void SetIsHfaRegArg(bool hfaRegArg)
1263 void SetIsBackFilled(bool backFilled)
1267 bool IsBackFilled() const
1271 #endif // !_TARGET_ARM_
1277 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1279 //-------------------------------------------------------------------------
1281 // The class fgArgInfo is used to handle the arguments
1282 // when morphing a GT_CALL node.
1287 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1288 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1289 unsigned argCount; // Updatable arg count value
1290 unsigned nextSlotNum; // Updatable slot count value
1291 unsigned stkLevel; // Stack depth when we make this call (for x86)
1293 #if defined(UNIX_X86_ABI)
1294 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1295 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1296 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1297 // Computed dynamically during codegen, based on stkSizeBytes and the current
1298 // stack level (genStackLevel) when the first stack adjustment is made for
1302 #if FEATURE_FIXED_OUT_ARGS
1303 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1306 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1307 bool hasRegArgs; // true if we have one or more register arguments
1308 bool hasStackArgs; // true if we have one or more stack arguments
1309 bool argsComplete; // marker for state
1310 bool argsSorted; // marker for state
1311 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1314 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1317 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1318 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1320 fgArgTabEntryPtr AddRegArg(
1321 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1323 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1324 fgArgTabEntryPtr AddRegArg(
1331 const bool isStruct,
1332 const regNumber otherRegNum = REG_NA,
1333 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1334 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1336 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1340 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1342 void RemorphReset();
1343 fgArgTabEntryPtr RemorphRegArg(
1344 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1346 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1348 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1350 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1352 void ArgsComplete();
1356 void EvalArgsToTemps();
1358 void RecordStkLevel(unsigned stkLvl);
1359 unsigned RetrieveStkLevel();
1365 fgArgTabEntryPtr* ArgTable()
1369 unsigned GetNextSlotNum()
1379 return hasStackArgs;
1381 bool AreArgsComplete() const
1383 return argsComplete;
1385 #if FEATURE_FIXED_OUT_ARGS
1386 unsigned GetOutArgSize() const
1390 void SetOutArgSize(unsigned newVal)
1392 outArgSize = newVal;
1394 #endif // FEATURE_FIXED_OUT_ARGS
1396 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1398 #if defined(UNIX_X86_ABI)
1399 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1400 #endif // defined(UNIX_X86_ABI)
1403 void SetStkSizeBytes(unsigned newStkSizeBytes)
1405 #if defined(UNIX_X86_ABI)
1406 stkSizeBytes = newStkSizeBytes;
1407 #endif // defined(UNIX_X86_ABI)
1410 #if defined(UNIX_X86_ABI)
1411 unsigned GetStkAlign()
1415 unsigned GetStkSizeBytes() const
1417 return stkSizeBytes;
1419 bool IsStkAlignmentDone() const
1421 return alignmentDone;
1423 void SetStkAlignmentDone()
1425 alignmentDone = true;
1427 #endif // defined(UNIX_X86_ABI)
1429 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1430 GenTreePtr GetLateArg(unsigned argIndex);
1434 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1435 // We have the ability to mark source expressions with "Test Labels."
1436 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1437 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1439 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1442 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1443 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1444 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1445 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1446 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1449 struct TestLabelAndNum
1454 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1459 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1461 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1464 // This class implements the "IAllocator" interface, so that we can use
1465 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1467 class CompAllocator : public IAllocator
1470 #if MEASURE_MEM_ALLOC
1474 CompAllocator(Compiler* comp, CompMemKind cmk)
1476 #if MEASURE_MEM_ALLOC
1482 inline void* Alloc(size_t sz);
1484 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1486 // For the compiler's no-release allocator, free operations are no-ops.
1493 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1494 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1496 XX The big guy. The sections are currently organized as : XX
1498 XX o GenTree and BasicBlock XX
1510 XX o PrologScopeInfo XX
1511 XX o CodeGenerator XX
1516 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1517 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1522 friend class emitter;
1523 friend class UnwindInfo;
1524 friend class UnwindFragmentInfo;
1525 friend class UnwindEpilogInfo;
1526 friend class JitTimer;
1527 friend class LinearScan;
1528 friend class fgArgInfo;
1529 friend class Rationalizer;
1531 friend class Lowering;
1532 friend class CSE_DataFlow;
1533 friend class CSE_Heuristic;
1534 friend class CodeGenInterface;
1535 friend class CodeGen;
1536 friend class LclVarDsc;
1537 friend class TempDsc;
1539 friend class ObjectAllocator;
1541 #ifndef _TARGET_64BIT_
1542 friend class DecomposeLongs;
1543 #endif // !_TARGET_64BIT_
1546 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1547 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1549 XX Misc structs definitions XX
1551 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1556 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1575 bool dumpIRDataflow;
1576 bool dumpIRBlockHeaders;
1578 LPCWSTR dumpIRPhase;
1579 LPCWSTR dumpIRFormat;
1581 bool shouldUseVerboseTrees();
1582 bool asciiTrees; // If true, dump trees using only ASCII characters
1583 bool shouldDumpASCIITrees();
1584 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1585 bool shouldUseVerboseSsa();
1586 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1587 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1589 const char* VarNameToStr(VarName name)
1594 DWORD expensiveDebugCheckLevel;
1597 #if FEATURE_MULTIREG_RET
1598 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1599 #endif // FEATURE_MULTIREG_RET
1602 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1603 #endif // ARM_SOFTFP
1605 //-------------------------------------------------------------------------
1606 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1607 // HFAs are one to four element structs where each element is the same
1608 // type, either all float or all double. They are treated specially
1609 // in the ARM Procedure Call Standard, specifically, they are passed in
1610 // floating-point registers instead of the general purpose registers.
1613 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1614 bool IsHfa(GenTreePtr tree);
1616 var_types GetHfaType(GenTreePtr tree);
1617 unsigned GetHfaCount(GenTreePtr tree);
1619 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1620 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1622 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1623 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1625 //-------------------------------------------------------------------------
1626 // The following is used for validating format of EH table
1630 typedef struct EHNodeDsc* pEHNodeDsc;
1632 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1633 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1646 EHBlockType ehnBlockType; // kind of EH block
1647 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1648 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1649 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1651 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1652 pEHNodeDsc ehnChild; // leftmost nested block
1654 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1655 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1657 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1658 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1660 inline void ehnSetTryNodeType()
1662 ehnBlockType = TryNode;
1664 inline void ehnSetFilterNodeType()
1666 ehnBlockType = FilterNode;
1668 inline void ehnSetHandlerNodeType()
1670 ehnBlockType = HandlerNode;
1672 inline void ehnSetFinallyNodeType()
1674 ehnBlockType = FinallyNode;
1676 inline void ehnSetFaultNodeType()
1678 ehnBlockType = FaultNode;
1681 inline BOOL ehnIsTryBlock()
1683 return ehnBlockType == TryNode;
1685 inline BOOL ehnIsFilterBlock()
1687 return ehnBlockType == FilterNode;
1689 inline BOOL ehnIsHandlerBlock()
1691 return ehnBlockType == HandlerNode;
1693 inline BOOL ehnIsFinallyBlock()
1695 return ehnBlockType == FinallyNode;
1697 inline BOOL ehnIsFaultBlock()
1699 return ehnBlockType == FaultNode;
1702 // returns true if there is any overlap between the two nodes
1703 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1705 if (node1->ehnStartOffset < node2->ehnStartOffset)
1707 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1711 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1715 // fails with BADCODE if inner is not completely nested inside outer
1716 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1718 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1722 //-------------------------------------------------------------------------
1723 // Exception handling functions
1726 #if !FEATURE_EH_FUNCLETS
1728 bool ehNeedsShadowSPslots()
1730 return (info.compXcptnsCount || opts.compDbgEnC);
1733 // 0 for methods with no EH
1734 // 1 for methods with non-nested EH, or where only the try blocks are nested
1735 // 2 for a method with a catch within a catch
1737 unsigned ehMaxHndNestingCount;
1739 #endif // !FEATURE_EH_FUNCLETS
1741 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1742 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1744 bool bbInCatchHandlerILRange(BasicBlock* blk);
1745 bool bbInFilterILRange(BasicBlock* blk);
1746 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1747 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1748 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1749 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1750 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1752 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1753 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1755 // Returns true if "block" is the start of a try region.
1756 bool bbIsTryBeg(BasicBlock* block);
1758 // Returns true if "block" is the start of a handler or filter region.
1759 bool bbIsHandlerBeg(BasicBlock* block);
1761 // Returns true iff "block" is where control flows if an exception is raised in the
1762 // try region, and sets "*regionIndex" to the index of the try for the handler.
1763 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1764 // block of the filter, but not for the filter's handler.
1765 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1767 bool ehHasCallableHandlers();
1769 // Return the EH descriptor for the given region index.
1770 EHblkDsc* ehGetDsc(unsigned regionIndex);
1772 // Return the EH index given a region descriptor.
1773 unsigned ehGetIndex(EHblkDsc* ehDsc);
1775 // Return the EH descriptor index of the enclosing try, for the given region index.
1776 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1778 // Return the EH descriptor index of the enclosing handler, for the given region index.
1779 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1781 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1782 // block is not in a 'try' region).
1783 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1785 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1786 // if this block is not in a filter or handler region).
1787 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1789 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1790 // nullptr if this block's exceptions propagate to caller).
1791 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1793 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1794 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1795 bool ehIsBlockEHLast(BasicBlock* block);
1797 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1799 // Return the region index of the most nested EH region this block is in.
1800 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1802 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1803 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1805 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1806 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1807 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1808 // (It can never be a filter.)
1809 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1811 // A block has been deleted. Update the EH table appropriately.
1812 void ehUpdateForDeletedBlock(BasicBlock* block);
1814 // Determine whether a block can be deleted while preserving the EH normalization rules.
1815 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1817 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1818 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1820 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1821 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1822 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1823 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1824 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1825 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1826 // lives in a filter.)
1827 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1829 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1830 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1831 // (nullptr if the last block is the last block in the program).
1832 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1833 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1836 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1837 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1838 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1841 #if FEATURE_EH_FUNCLETS
1842 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1843 // if there is a filter that protects a region with a nested EH clause (such as a
1844 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1845 // genFuncletProlog() for more details. However, the VM seems to use it for more
1846 // purposes, maybe including debugging. Until we are sure otherwise, always create
1847 // a PSPSym for functions with any EH.
1848 bool ehNeedsPSPSym() const
1852 #else // _TARGET_X86_
1853 return compHndBBtabCount > 0;
1854 #endif // _TARGET_X86_
1857 bool ehAnyFunclets(); // Are there any funclets in this function?
1858 unsigned ehFuncletCount(); // Return the count of funclets in the function
1860 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1861 #else // !FEATURE_EH_FUNCLETS
1862 bool ehAnyFunclets()
1866 unsigned ehFuncletCount()
1871 unsigned bbThrowIndex(BasicBlock* blk)
1873 return blk->bbTryIndex;
1874 } // Get the index to use as the cache key for sharing throw blocks
1875 #endif // !FEATURE_EH_FUNCLETS
1877 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1878 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1879 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1880 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1881 // convenient to also consider it a predecessor.)
1882 flowList* BlockPredsWithEH(BasicBlock* blk);
1884 // This table is useful for memoization of the method above.
1885 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1887 BlockToFlowListMap* m_blockToEHPreds;
1888 BlockToFlowListMap* GetBlockToEHPreds()
1890 if (m_blockToEHPreds == nullptr)
1892 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1894 return m_blockToEHPreds;
1897 void* ehEmitCookie(BasicBlock* block);
1898 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1900 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1902 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1904 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1906 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1908 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1910 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1912 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1914 void fgAllocEHTable();
1916 void fgRemoveEHTableEntry(unsigned XTnum);
1918 #if FEATURE_EH_FUNCLETS
1920 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1922 #endif // FEATURE_EH_FUNCLETS
1926 #endif // !FEATURE_EH
1928 void fgSortEHTable();
1930 // Causes the EH table to obey some well-formedness conditions, by inserting
1931 // empty BB's when necessary:
1932 // * No block is both the first block of a handler and the first block of a try.
1933 // * No block is the first block of multiple 'try' regions.
1934 // * No block is the last block of multiple EH regions.
1935 void fgNormalizeEH();
1936 bool fgNormalizeEHCase1();
1937 bool fgNormalizeEHCase2();
1938 bool fgNormalizeEHCase3();
1941 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1942 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1943 void fgVerifyHandlerTab();
1944 void fgDispHandlerTab();
1947 bool fgNeedToSortEHTable;
1949 void verInitEHTree(unsigned numEHClauses);
1950 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1951 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1952 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1953 void verCheckNestingLevel(EHNodeDsc* initRoot);
1956 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1957 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1959 XX GenTree and BasicBlock XX
1961 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1963 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1964 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1967 // Functions to create nodes
1968 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1971 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1973 // For binary opers.
1974 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1976 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1978 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1979 var_types type = TYP_I_IMPL,
1980 GenTreePtr op1 = nullptr,
1981 GenTreePtr op2 = nullptr);
1983 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1985 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1987 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1989 GenTreePtr gtNewJmpTableNode();
1990 GenTreePtr gtNewIconHandleNode(
1991 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1993 unsigned gtTokenToIconFlags(unsigned token);
1995 GenTreePtr gtNewIconEmbHndNode(void* value,
1998 unsigned handle1 = 0,
1999 void* handle2 = nullptr,
2000 void* compileTimeHandle = nullptr);
2002 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2003 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2004 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2005 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2007 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2009 GenTreePtr gtNewLconNode(__int64 value);
2011 GenTreePtr gtNewDconNode(double value);
2013 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2015 GenTreePtr gtNewZeroConNode(var_types type);
2017 GenTreePtr gtNewOneConNode(var_types type);
2020 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2021 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2024 GenTreeBlk* gtNewBlkOpNode(
2025 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
2027 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2030 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
2033 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2034 void gtSetObjGcInfo(GenTreeObj* objNode);
2035 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2036 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
2038 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2040 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
2042 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2043 CORINFO_METHOD_HANDLE handle,
2045 GenTreeArgList* args,
2046 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2048 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
2050 GenTreeArgList* args,
2051 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2053 GenTreeCall* gtNewHelperCallNode(unsigned helper,
2056 GenTreeArgList* args = nullptr);
2058 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2061 GenTreeSIMD* gtNewSIMDNode(
2062 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2063 GenTreeSIMD* gtNewSIMDNode(var_types type,
2066 SIMDIntrinsicID simdIntrinsicID,
2069 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2072 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2073 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2074 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2076 GenTreePtr gtNewCodeRef(BasicBlock* block);
2078 GenTreePtr gtNewFieldRef(
2079 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2081 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2083 GenTreeArgList* gtNewArgList(GenTreePtr op);
2084 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2085 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2087 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2088 static fgArgTabEntryPtr gtArgEntryByNode(GenTreeCall* call, GenTreePtr node);
2089 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2090 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2092 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2094 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2096 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2097 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2098 CORINFO_ACCESS_FLAGS access,
2099 CORINFO_FIELD_INFO* pFieldInfo,
2101 CORINFO_CLASS_HANDLE structType,
2104 GenTreePtr gtNewNothingNode();
2106 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2108 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2110 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2112 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2114 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2116 //------------------------------------------------------------------------
2117 // Other GenTree functions
2119 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2121 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2122 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2123 // IntCnses with value `deepVarVal`.
2124 GenTreePtr gtCloneExpr(
2125 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2127 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2128 // `varNum` to int constants with value `varVal`.
2129 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2131 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2134 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2136 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2138 // Returns "true" iff the complexity (not formally defined, but first interpretation
2139 // is #of nodes in subtree) of "tree" is greater than "limit".
2140 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2141 // before they have been set.)
2142 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2144 bool gtCompareTree(GenTree* op1, GenTree* op2);
2146 GenTreePtr gtReverseCond(GenTree* tree);
2148 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2150 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2152 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2154 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2157 unsigned gtHashValue(GenTree* tree);
2159 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2162 void gtPrepareCost(GenTree* tree);
2163 bool gtIsLikelyRegVar(GenTree* tree);
2165 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2167 // Returns true iff the secondNode can be swapped with firstNode.
2168 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2170 unsigned gtSetEvalOrder(GenTree* tree);
2172 #if FEATURE_STACK_FP_X87
2174 void gtComputeFPlvls(GenTreePtr tree);
2175 #endif // FEATURE_STACK_FP_X87
2177 void gtSetStmtInfo(GenTree* stmt);
2179 // Returns "true" iff "node" has any of the side effects in "flags".
2180 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2182 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2183 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2185 // Appends 'expr' in front of 'list'
2186 // 'list' will typically start off as 'nullptr'
2187 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2188 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2190 void gtExtractSideEffList(GenTreePtr expr,
2192 unsigned flags = GTF_SIDE_EFFECT,
2193 bool ignoreRoot = false);
2195 GenTreePtr gtGetThisArg(GenTreeCall* call);
2197 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2198 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2199 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2200 // the given "fldHnd", is such an object pointer.
2201 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2203 // Return true if call is a recursive call; return false otherwise.
2204 // Note when inlining, this looks for calls back to the root method.
2205 bool gtIsRecursiveCall(GenTreeCall* call)
2207 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2210 //-------------------------------------------------------------------------
2212 GenTreePtr gtFoldExpr(GenTreePtr tree);
2215 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2216 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2217 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2218 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2219 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2220 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2221 // optimizations for now.
2222 __attribute__((optnone))
2224 gtFoldExprConst(GenTreePtr tree);
2225 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2226 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2228 //-------------------------------------------------------------------------
2229 // Get the handle, if any.
2230 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2231 // Get the handle, and assert if not found.
2232 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2233 // Get the handle for a ref type.
2234 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2236 //-------------------------------------------------------------------------
2237 // Functions to display the trees
2240 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2242 void gtDispVN(GenTreePtr tree);
2243 void gtDispConst(GenTreePtr tree);
2244 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2245 void gtDispNodeName(GenTreePtr tree);
2246 void gtDispRegVal(GenTreePtr tree);
2258 void gtDispChild(GenTreePtr child,
2259 IndentStack* indentStack,
2261 __in_opt const char* msg = nullptr,
2262 bool topOnly = false);
2263 void gtDispTree(GenTreePtr tree,
2264 IndentStack* indentStack = nullptr,
2265 __in_opt const char* msg = nullptr,
2266 bool topOnly = false,
2267 bool isLIR = false);
2268 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2269 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2270 char* gtGetLclVarName(unsigned lclNum);
2271 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2272 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2273 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2274 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2275 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2276 void gtDispFieldSeq(FieldSeqNode* pfsn);
2278 void gtDispRange(LIR::ReadOnlyRange const& range);
2280 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2282 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2294 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2295 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2298 static fgWalkPreFn gtAssertColonCond;
2300 static fgWalkPreFn gtMarkColonCond;
2301 static fgWalkPreFn gtClearColonCond;
2303 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2304 bool gtHasCatchArg(GenTreePtr tree);
2305 bool gtHasUnmanagedCall(GenTreePtr tree);
2307 typedef ArrayStack<GenTree*> GenTreeStack;
2309 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2310 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2312 //=========================================================================
2313 // BasicBlock functions
2315 // This is a debug flag we will use to assert when creating block during codegen
2316 // as this interferes with procedure splitting. If you know what you're doing, set
2317 // it to true before creating the block. (DEBUG only)
2318 bool fgSafeBasicBlockCreation;
2321 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2324 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2325 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2329 XX The variables to be used by the code generator. XX
2331 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2332 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2336 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2337 // be placed in the stack frame and it's fields must be laid out sequentially.
2339 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2340 // a local variable that can be enregistered or placed in the stack frame.
2341 // The fields do not need to be laid out sequentially
2343 enum lvaPromotionType
2345 PROMOTION_TYPE_NONE, // The struct local is not promoted
2346 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2347 // and its field locals are independent of its parent struct local.
2348 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2349 // but its field locals depend on its parent struct local.
2352 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2353 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2355 /*****************************************************************************/
2357 enum FrameLayoutState
2360 INITIAL_FRAME_LAYOUT,
2361 PRE_REGALLOC_FRAME_LAYOUT,
2362 REGALLOC_FRAME_LAYOUT,
2363 TENTATIVE_FRAME_LAYOUT,
2368 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2369 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2370 bool lvaSortAgain; // true: We need to sort the lvaTable
2371 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2372 unsigned lvaCount; // total number of locals
2374 unsigned lvaRefCount; // total number of references to locals
2375 LclVarDsc* lvaTable; // variable descriptor table
2376 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2378 LclVarDsc** lvaRefSorted; // table sorted by refcount
2380 unsigned short lvaTrackedCount; // actual # of locals being tracked
2381 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2383 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2384 // Only for AMD64 System V cache the first caller stack homed argument.
2385 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2386 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2389 VARSET_TP lvaTrackedVars; // set of tracked variables
2391 #ifndef _TARGET_64BIT_
2392 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2394 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2396 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2397 // It that changes, this changes. VarSets from different epochs
2398 // cannot be meaningfully combined.
2400 unsigned GetCurLVEpoch()
2405 // reverse map of tracked number to var number
2406 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2408 #ifdef LEGACY_BACKEND
2409 // variable interference graph
2410 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2413 // variable preference graph
2414 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2418 // # of procs compiled a with double-aligned stack
2419 static unsigned s_lvaDoubleAlignedProcsCount;
2423 // Getters and setters for address-exposed and do-not-enregister local var properties.
2424 bool lvaVarAddrExposed(unsigned varNum);
2425 void lvaSetVarAddrExposed(unsigned varNum);
2426 bool lvaVarDoNotEnregister(unsigned varNum);
2428 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2429 enum DoNotEnregisterReason
2434 DNER_VMNeedsStackAddr,
2435 DNER_LiveInOutOfHandler,
2436 DNER_LiveAcrossUnmanagedCall,
2437 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2438 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2439 #ifdef JIT32_GCENCODER
2444 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2446 unsigned lvaVarargsHandleArg;
2448 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2450 #endif // _TARGET_X86_
2452 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2453 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2454 #if FEATURE_FIXED_OUT_ARGS
2455 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2457 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2458 // that tracks whether the lock has been taken
2460 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2461 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2462 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2464 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2465 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2467 #if FEATURE_FIXED_OUT_ARGS
2468 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2469 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2470 #endif // FEATURE_FIXED_OUT_ARGS
2473 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2474 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2475 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2476 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2477 // this variable to be this scratch word whenever struct promotion occurs.
2478 unsigned lvaPromotedStructAssemblyScratchVar;
2479 #endif // _TARGET_ARM_
2482 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2483 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2486 unsigned lvaGenericsContextUseCount;
2488 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2489 // CORINFO_GENERICS_CTXT_FROM_THIS?
2490 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2492 //-------------------------------------------------------------------------
2493 // All these frame offsets are inter-related and must be kept in sync
2495 #if !FEATURE_EH_FUNCLETS
2496 // This is used for the callable handlers
2497 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2498 #endif // FEATURE_EH_FUNCLETS
2500 unsigned lvaCachedGenericContextArgOffs;
2501 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2504 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2506 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2508 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2509 // after the reg predict we will use a computed maxTmpSize
2510 // which is based upon the number of spill temps predicted by reg predict
2511 // All this is necessary because if we under-estimate the size of the spill
2512 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2514 // Pre codegen max spill temp size.
2515 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2517 //-------------------------------------------------------------------------
2519 unsigned lvaGetMaxSpillTempSize();
2521 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2522 #endif // _TARGET_ARM_
2523 void lvaAssignFrameOffsets(FrameLayoutState curState);
2524 void lvaFixVirtualFrameOffsets();
2526 #ifndef LEGACY_BACKEND
2527 void lvaUpdateArgsWithInitialReg();
2528 #endif // !LEGACY_BACKEND
2530 void lvaAssignVirtualFrameOffsetsToArgs();
2531 #ifdef UNIX_AMD64_ABI
2532 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2533 #else // !UNIX_AMD64_ABI
2534 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2535 #endif // !UNIX_AMD64_ABI
2536 void lvaAssignVirtualFrameOffsetsToLocals();
2537 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2538 #ifdef _TARGET_AMD64_
2539 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2540 bool lvaIsCalleeSavedIntRegCountEven();
2542 void lvaAlignFrame();
2543 void lvaAssignFrameOffsetsToPromotedStructs();
2544 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2547 void lvaDumpRegLocation(unsigned lclNum);
2548 void lvaDumpFrameLocation(unsigned lclNum);
2549 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2550 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2551 // layout state defined by lvaDoneFrameLayout
2554 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2555 // to avoid bugs from borderline cases.
2556 #define MAX_FrameSize 0x3FFFFFFF
2557 void lvaIncrementFrameSize(unsigned size);
2559 unsigned lvaFrameSize(FrameLayoutState curState);
2561 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2562 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2564 // Returns the caller-SP-relative offset for the local variable "varNum."
2565 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2567 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2568 int lvaGetSPRelativeOffset(unsigned varNum);
2570 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2571 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2573 //------------------------ For splitting types ----------------------------
2575 void lvaInitTypeRef();
2577 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2578 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2579 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2580 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2581 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2582 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2584 void lvaInitVarDsc(LclVarDsc* varDsc,
2586 CorInfoType corInfoType,
2587 CORINFO_CLASS_HANDLE typeHnd,
2588 CORINFO_ARG_LIST_HANDLE varList,
2589 CORINFO_SIG_INFO* varSig);
2591 static unsigned lvaTypeRefMask(var_types type);
2593 var_types lvaGetActualType(unsigned lclNum);
2594 var_types lvaGetRealType(unsigned lclNum);
2596 //-------------------------------------------------------------------------
2600 unsigned lvaLclSize(unsigned varNum);
2601 unsigned lvaLclExactSize(unsigned varNum);
2603 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2605 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2606 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2607 // the return result.
2608 bool lvaLclVarRefsAccum(
2609 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2611 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2612 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2613 // and (destructively) unions "trkedVars" into "*result".
2614 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2616 ALLVARSET_VALARG_TP allVars,
2617 VARSET_VALARG_TP trkdVars);
2619 bool lvaHaveManyLocals() const;
2621 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2622 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2623 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2626 void lvaSortByRefCount();
2627 void lvaDumpRefCounts();
2629 void lvaMarkLocalVars(BasicBlock* block);
2631 void lvaMarkLocalVars(); // Local variable ref-counting
2633 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2635 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2637 static fgWalkPreFn lvaIncRefCntsCB;
2638 void lvaIncRefCnts(GenTreePtr tree);
2640 static fgWalkPreFn lvaDecRefCntsCB;
2641 void lvaDecRefCnts(GenTreePtr tree);
2642 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2643 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2644 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2647 struct lvaStressLclFldArgs
2649 Compiler* m_pCompiler;
2653 static fgWalkPreFn lvaStressLclFldCB;
2654 void lvaStressLclFld();
2656 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2657 void lvaDispVarSet(VARSET_VALARG_TP set);
2662 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2664 int lvaFrameAddress(int varNum, bool* pFPbased);
2667 bool lvaIsParameter(unsigned varNum);
2668 bool lvaIsRegArgument(unsigned varNum);
2669 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2670 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2671 // that writes to arg0
2673 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2674 // (this is an overload of lvIsTemp because there are no temp parameters).
2675 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2676 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2677 bool lvaIsImplicitByRefLocal(unsigned varNum)
2679 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2680 LclVarDsc* varDsc = &(lvaTable[varNum]);
2681 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2683 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2686 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2690 // Returns true if this local var is a multireg struct
2691 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2693 // If the local is a TYP_STRUCT, get/set a class handle describing it
2694 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2695 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2697 // If the local is TYP_REF, set or update the associated class information.
2698 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2699 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2700 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2701 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2703 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2705 // Info about struct fields
2706 struct lvaStructFieldInfo
2708 CORINFO_FIELD_HANDLE fldHnd;
2709 unsigned char fldOffset;
2710 unsigned char fldOrdinal;
2713 CORINFO_CLASS_HANDLE fldTypeHnd;
2716 // Info about struct to be promoted.
2717 struct lvaStructPromotionInfo
2719 CORINFO_CLASS_HANDLE typeHnd;
2721 bool requiresScratchVar;
2724 unsigned char fieldCnt;
2725 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2727 lvaStructPromotionInfo()
2728 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2733 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2734 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2735 lvaStructPromotionInfo* StructPromotionInfo,
2737 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2738 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2739 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2740 #if !defined(_TARGET_64BIT_)
2741 void lvaPromoteLongVars();
2742 #endif // !defined(_TARGET_64BIT_)
2743 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2744 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2745 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2746 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2747 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2748 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2749 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2751 #if defined(FEATURE_SIMD)
2752 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2754 assert(varDsc->lvType == TYP_SIMD12);
2755 assert(varDsc->lvExactSize == 12);
2757 #if defined(_TARGET_64BIT_)
2758 assert(varDsc->lvSize() == 16);
2760 #else // !defined(_TARGET_64BIT_)
2762 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2763 // already does this calculation. However, we also need to prevent mapping types if the var is a
2764 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2765 // However, we don't know this until late, so we may have already pretended the field is bigger
2767 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2776 #endif // !defined(_TARGET_64BIT_)
2778 #endif // defined(FEATURE_SIMD)
2780 BYTE* lvaGetGcLayout(unsigned varNum);
2781 bool lvaTypeIsGC(unsigned varNum);
2782 unsigned lvaGSSecurityCookie; // LclVar number
2783 bool lvaTempsHaveLargerOffsetThanVars();
2785 unsigned lvaSecurityObject; // variable representing the security object on the stack
2786 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2788 #if FEATURE_EH_FUNCLETS
2789 unsigned lvaPSPSym; // variable representing the PSPSym
2792 InlineInfo* impInlineInfo;
2793 InlineStrategy* m_inlineStrategy;
2795 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2796 Compiler* impInlineRoot();
2798 #if defined(DEBUG) || defined(INLINE_DATA)
2799 unsigned __int64 getInlineCycleCount()
2801 return m_compCycles;
2803 #endif // defined(DEBUG) || defined(INLINE_DATA)
2805 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2806 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2808 //=========================================================================
2810 //=========================================================================
2813 //---------------- Local variable ref-counting ----------------------------
2816 BasicBlock* lvaMarkRefsCurBlock;
2817 GenTreePtr lvaMarkRefsCurStmt;
2819 BasicBlock::weight_t lvaMarkRefsWeight;
2821 static fgWalkPreFn lvaMarkLclRefsCallback;
2822 void lvaMarkLclRefs(GenTreePtr tree);
2824 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2825 void SetVolatileHint(LclVarDsc* varDsc);
2827 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2828 PerSsaArray lvMemoryPerSsaData;
2829 unsigned lvMemoryNumSsaNames;
2832 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2833 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2834 // not an SSA variable).
2835 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2837 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2838 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2840 assert(ssaNum < lvMemoryNumSsaNames);
2841 return &lvMemoryPerSsaData.GetRef(ssaNum);
2845 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2846 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2850 XX Imports the given method and converts it to semantic trees XX
2852 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2853 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2859 void impImport(BasicBlock* method);
2861 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2862 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2863 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2864 CORINFO_CLASS_HANDLE impGetStringClass();
2865 CORINFO_CLASS_HANDLE impGetObjectClass();
2867 //=========================================================================
2869 //=========================================================================
2872 //-------------------- Stack manipulation ---------------------------------
2874 unsigned impStkSize; // Size of the full stack
2876 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2878 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2880 struct SavedStack // used to save/restore stack contents.
2882 unsigned ssDepth; // number of values on stack
2883 StackEntry* ssTrees; // saved tree values
2886 bool impIsPrimitive(CorInfoType type);
2887 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2889 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2891 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2892 void impPushNullObjRefOnStack();
2893 StackEntry impPopStack();
2894 StackEntry& impStackTop(unsigned n = 0);
2895 unsigned impStackHeight();
2897 void impSaveStackState(SavedStack* savePtr, bool copy);
2898 void impRestoreStackState(SavedStack* savePtr);
2900 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2901 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2902 CORINFO_CALL_INFO* pCallInfo);
2904 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2906 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2908 bool impCanPInvokeInline();
2909 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2910 void impCheckForPInvokeCall(
2911 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2912 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2913 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2915 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2916 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2917 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2919 var_types impImportCall(OPCODE opcode,
2920 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2921 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2923 GenTreePtr newobjThis,
2925 CORINFO_CALL_INFO* callInfo,
2926 IL_OFFSET rawILOffset);
2928 void impDevirtualizeCall(GenTreeCall* call,
2930 CORINFO_CALL_INFO* callInfo,
2931 CORINFO_CONTEXT_HANDLE* exactContextHnd);
2933 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2935 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2937 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2940 var_types impImportJitTestLabelMark(int numArgs);
2943 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2945 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2947 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2948 CORINFO_ACCESS_FLAGS access,
2949 CORINFO_FIELD_INFO* pFieldInfo,
2952 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2954 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2956 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2958 void impImportLeave(BasicBlock* block);
2959 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2960 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2961 CORINFO_CLASS_HANDLE clsHnd,
2962 CORINFO_METHOD_HANDLE method,
2963 CORINFO_SIG_INFO* sig,
2967 CorInfoIntrinsics* pIntrinsicID);
2968 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2969 CORINFO_SIG_INFO* sig,
2972 CorInfoIntrinsics intrinsicID);
2973 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2975 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2977 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2978 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2979 CORINFO_THIS_TRANSFORM transform);
2981 //----------------- Manipulating the trees and stmts ----------------------
2983 GenTreePtr impTreeList; // Trees for the BB being imported
2984 GenTreePtr impTreeLast; // The last tree for the current BB
2988 CHECK_SPILL_ALL = -1,
2989 CHECK_SPILL_NONE = -2
2993 void impBeginTreeList();
2994 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2995 void impEndTreeList(BasicBlock* block);
2996 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2997 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2998 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2999 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3000 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3001 void impAssignTempGen(unsigned tmp,
3004 GenTreePtr* pAfterStmt = nullptr,
3005 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3006 BasicBlock* block = nullptr);
3007 void impAssignTempGen(unsigned tmpNum,
3009 CORINFO_CLASS_HANDLE structHnd,
3011 GenTreePtr* pAfterStmt = nullptr,
3012 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3013 BasicBlock* block = nullptr);
3014 GenTreePtr impCloneExpr(GenTreePtr tree,
3016 CORINFO_CLASS_HANDLE structHnd,
3018 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3019 GenTreePtr impAssignStruct(GenTreePtr dest,
3021 CORINFO_CLASS_HANDLE structHnd,
3023 GenTreePtr* pAfterStmt = nullptr,
3024 BasicBlock* block = nullptr);
3025 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3027 CORINFO_CLASS_HANDLE structHnd,
3029 GenTreePtr* pAfterStmt = nullptr,
3030 BasicBlock* block = nullptr);
3032 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3033 CORINFO_CLASS_HANDLE structHnd,
3037 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3038 BYTE* gcLayout = nullptr,
3039 unsigned* numGCVars = nullptr,
3040 var_types* simdBaseType = nullptr);
3042 GenTreePtr impNormStructVal(GenTreePtr structVal,
3043 CORINFO_CLASS_HANDLE structHnd,
3045 bool forceNormalization = false);
3047 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3048 BOOL* pRuntimeLookup = nullptr,
3049 BOOL mustRestoreHandle = FALSE,
3050 BOOL importParent = FALSE);
3052 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3053 BOOL* pRuntimeLookup = nullptr,
3054 BOOL mustRestoreHandle = FALSE)
3056 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3059 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3060 CORINFO_LOOKUP* pLookup,
3062 void* compileTimeHandle);
3064 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3066 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3067 CORINFO_LOOKUP* pLookup,
3068 void* compileTimeHandle);
3070 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3072 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3073 CorInfoHelpFunc helper,
3075 GenTreeArgList* arg = nullptr,
3076 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3078 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3080 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3083 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3084 CORINFO_CLASS_HANDLE typeClass,
3088 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3089 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3090 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3091 static bool IsMathIntrinsic(GenTreePtr tree);
3094 //----------------- Importing the method ----------------------------------
3096 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3099 unsigned impCurOpcOffs;
3100 const char* impCurOpcName;
3101 bool impNestedStackSpill;
3103 // For displaying instrs with generated native code (-n:B)
3104 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3105 void impNoteLastILoffs();
3108 /* IL offset of the stmt currently being imported. It gets set to
3109 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3110 updated at IL offsets for which we have to report mapping info.
3111 It also includes flag bits, so use jitGetILoffs()
3112 to get the actual IL offset value.
3115 IL_OFFSETX impCurStmtOffs;
3116 void impCurStmtOffsSet(IL_OFFSET offs);
3118 void impNoteBranchOffs();
3120 unsigned impInitBlockLineInfo();
3122 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3123 bool impIsThis(GenTreePtr obj);
3124 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3125 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3126 bool impIsAnySTLOC(OPCODE opcode)
3128 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3129 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3132 GenTreeArgList* impPopList(unsigned count,
3134 CORINFO_SIG_INFO* sig,
3135 GenTreeArgList* prefixTree = nullptr);
3137 GenTreeArgList* impPopRevList(unsigned count,
3139 CORINFO_SIG_INFO* sig,
3140 unsigned skipReverseCount = 0);
3143 * Get current IL offset with stack-empty info incoporated
3145 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3147 //---------------- Spilling the importer stack ----------------------------
3149 // The maximum number of bytes of IL processed without clean stack state.
3150 // It allows to limit the maximum tree size and depth.
3151 static const unsigned MAX_TREE_SIZE = 200;
3152 bool impCanSpillNow(OPCODE prevOpcode);
3158 SavedStack pdSavedStack;
3159 ThisInitState pdThisPtrInit;
3162 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3163 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3165 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3166 ExpandArray<BYTE> impPendingBlockMembers;
3168 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3169 // Operates on the map in the top-level ancestor.
3170 BYTE impGetPendingBlockMember(BasicBlock* blk)
3172 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3175 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3176 // Operates on the map in the top-level ancestor.
3177 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3179 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3182 bool impCanReimport;
3184 bool impSpillStackEntry(unsigned level,
3188 bool bAssertOnRecursion,
3193 void impSpillStackEnsure(bool spillLeaves = false);
3194 void impEvalSideEffects();
3195 void impSpillSpecialSideEff();
3196 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3197 void impSpillValueClasses();
3198 void impSpillEvalStack();
3199 static fgWalkPreFn impFindValueClasses;
3200 void impSpillLclRefs(ssize_t lclNum);
3202 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3204 void impImportBlockCode(BasicBlock* block);
3206 void impReimportMarkBlock(BasicBlock* block);
3207 void impReimportMarkSuccessors(BasicBlock* block);
3209 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3211 void impImportBlockPending(BasicBlock* block);
3213 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3214 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3215 // for the block, but instead, just re-uses the block's existing EntryState.
3216 void impReimportBlockPending(BasicBlock* block);
3218 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3220 void impImportBlock(BasicBlock* block);
3222 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3223 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3224 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3225 // the variables that will be used -- and for all the predecessors of those successors, and the
3226 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3227 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3228 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3229 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3230 // of local variable numbers, so we represent them with the base local variable number), returns that.
3231 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3232 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3233 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3234 // on which kind of member of the clique the block is).
3235 unsigned impGetSpillTmpBase(BasicBlock* block);
3237 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3238 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3239 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3240 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3241 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3242 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3243 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3244 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3245 // successors receive a native int. Similarly float and double are unified to double.
3246 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3247 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3248 // predecessors, so they insert an upcast if needed).
3249 void impReimportSpillClique(BasicBlock* block);
3251 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3252 // block, and represent the predecessor and successor members of the clique currently being computed.
3253 // *** Access to these will need to be locked in a parallel compiler.
3254 ExpandArray<BYTE> impSpillCliquePredMembers;
3255 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3263 // Abstract class for receiving a callback while walking a spill clique
3264 class SpillCliqueWalker
3267 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3270 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3271 class SetSpillTempsBase : public SpillCliqueWalker
3276 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3279 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3282 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3283 class ReimportSpillClique : public SpillCliqueWalker
3288 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3291 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3294 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3295 // predecessor or successor within the spill clique
3296 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3298 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3299 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3300 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3301 void impRetypeEntryStateTemps(BasicBlock* blk);
3303 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3304 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3306 void impPushVar(GenTree* op, typeInfo tiRetVal);
3307 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3308 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3310 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3312 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3313 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3314 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3317 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3320 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3321 struct BlockListNode
3324 BlockListNode* m_next;
3325 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3328 void* operator new(size_t sz, Compiler* comp);
3330 BlockListNode* impBlockListNodeFreeList;
3332 BlockListNode* AllocBlockListNode();
3333 void FreeBlockListNode(BlockListNode* node);
3335 bool impIsValueType(typeInfo* pTypeInfo);
3336 var_types mangleVarArgsType(var_types type);
3339 regNumber getCallArgIntRegister(regNumber floatReg);
3340 regNumber getCallArgFloatRegister(regNumber intReg);
3341 #endif // FEATURE_VARARG
3344 static unsigned jitTotalMethodCompiled;
3348 static LONG jitNestingLevel;
3351 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3353 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3355 // STATIC inlining decision based on the IL code.
3356 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3357 CORINFO_METHOD_INFO* methInfo,
3359 InlineResult* inlineResult);
3361 void impCheckCanInline(GenTreePtr call,
3362 CORINFO_METHOD_HANDLE fncHandle,
3364 CORINFO_CONTEXT_HANDLE exactContextHnd,
3365 InlineCandidateInfo** ppInlineCandidateInfo,
3366 InlineResult* inlineResult);
3368 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3369 GenTreePtr curArgVal,
3371 InlineResult* inlineResult);
3373 void impInlineInitVars(InlineInfo* pInlineInfo);
3375 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3377 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3379 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3381 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3382 GenTreePtr variableBeingDereferenced,
3383 InlArgInfo* inlArgInfo);
3385 void impMarkInlineCandidate(GenTreePtr call,
3386 CORINFO_CONTEXT_HANDLE exactContextHnd,
3387 bool exactContextNeedsRuntimeLookup,
3388 CORINFO_CALL_INFO* callInfo);
3390 bool impTailCallRetTypeCompatible(var_types callerRetType,
3391 CORINFO_CLASS_HANDLE callerRetTypeClass,
3392 var_types calleeRetType,
3393 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3395 bool impIsTailCallILPattern(bool tailPrefixed,
3397 const BYTE* codeAddrOfNextOpcode,
3398 const BYTE* codeEnd,
3400 bool* IsCallPopRet = nullptr);
3402 bool impIsImplicitTailCallCandidate(
3403 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3405 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3408 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3409 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3413 XX Info about the basic-blocks, their contents and the flow analysis XX
3415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3416 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3420 BasicBlock* fgFirstBB; // Beginning of the basic block list
3421 BasicBlock* fgLastBB; // End of the basic block list
3422 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3423 #if FEATURE_EH_FUNCLETS
3424 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3426 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3428 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3429 unsigned fgEdgeCount; // # of control flow edges between the BBs
3430 unsigned fgBBcount; // # of BBs in the method
3432 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3434 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3435 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3436 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3437 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3439 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3440 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3441 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3442 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3443 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3444 // index). The arrays are of size fgBBNumMax + 1.
3445 unsigned* fgDomTreePreOrder;
3446 unsigned* fgDomTreePostOrder;
3448 bool fgBBVarSetsInited;
3450 // Allocate array like T* a = new T[fgBBNumMax + 1];
3451 // Using helper so we don't keep forgetting +1.
3452 template <typename T>
3453 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3455 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3458 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3459 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3460 // cannot be meaningfully combined. Note that new blocks can be created with higher
3461 // block numbers without changing the basic block epoch. These blocks *cannot*
3462 // participate in a block set until the blocks are all renumbered, causing the epoch
3463 // to change. This is useful if continuing to use previous block sets is valuable.
3464 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3465 unsigned fgCurBBEpoch;
3467 unsigned GetCurBasicBlockEpoch()
3469 return fgCurBBEpoch;
3472 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3473 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3474 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3475 unsigned fgCurBBEpochSize;
3477 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3478 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3479 unsigned fgBBSetCountInSizeTUnits;
3481 void NewBasicBlockEpoch()
3483 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3485 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3487 fgCurBBEpochSize = fgBBNumMax + 1;
3488 fgBBSetCountInSizeTUnits =
3489 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3492 // All BlockSet objects are now invalid!
3493 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3494 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3498 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3499 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3500 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3501 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3503 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3504 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3505 // array of size_t bitsets), then print that out.
3506 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3513 void EnsureBasicBlockEpoch()
3515 if (fgCurBBEpochSize != fgBBNumMax + 1)
3517 NewBasicBlockEpoch();
3521 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3522 void fgEnsureFirstBBisScratch();
3523 bool fgFirstBBisScratch();
3524 bool fgBBisScratch(BasicBlock* block);
3526 void fgExtendEHRegionBefore(BasicBlock* block);
3527 void fgExtendEHRegionAfter(BasicBlock* block);
3529 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3531 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3533 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3536 BasicBlock* nearBlk,
3537 bool putInFilter = false,
3538 bool runRarely = false,
3539 bool insertAtEnd = false);
3541 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3543 bool runRarely = false,
3544 bool insertAtEnd = false);
3546 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3548 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3549 BasicBlock* afterBlk,
3550 unsigned xcptnIndex,
3551 bool putInTryRegion);
3553 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3554 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3555 void fgUnlinkBlock(BasicBlock* block);
3557 unsigned fgMeasureIR();
3559 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3560 bool fgMultipleNots;
3563 bool fgModified; // True if the flow graph has been modified recently
3564 bool fgComputePredsDone; // Have we computed the bbPreds list
3565 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3566 bool fgDomsComputed; // Have we computed the dominator sets?
3567 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3569 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3570 bool fgHasPostfix; // any postfix ++/-- found?
3571 unsigned fgIncrCount; // number of increment nodes found
3573 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3577 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3578 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3581 bool fgRemoveRestOfBlock; // true if we know that we will throw
3582 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3584 // There are two modes for ordering of the trees.
3585 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3586 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3587 // by traversing the tree according to the order of the operands.
3588 // - In FGOrderLinear, the dominant ordering is the linear order.
3595 FlowGraphOrder fgOrder;
3597 // The following are boolean flags that keep track of the state of internal data structures
3599 bool fgStmtListThreaded;
3600 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3601 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3602 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3603 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3604 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3605 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3606 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3607 // This is derived from the profile data
3608 // or is BB_UNITY_WEIGHT when we don't have profile data
3610 #if FEATURE_EH_FUNCLETS
3611 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3612 #endif // FEATURE_EH_FUNCLETS
3614 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3615 // since fgMorphTree can be called from several places
3616 bool fgExpandInline; // indicates that we are creating tree for the inliner
3618 bool impBoxTempInUse; // the temp below is valid and available
3619 unsigned impBoxTemp; // a temporary that is used for boxing
3622 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3623 // and we are trying to compile again in a "safer", minopts mode?
3627 unsigned impInlinedCodeSize;
3630 //-------------------------------------------------------------------------
3636 void fgTransformFatCalli();
3640 void fgRemoveEmptyTry();
3642 void fgRemoveEmptyFinally();
3644 void fgMergeFinallyChains();
3646 void fgCloneFinally();
3648 void fgCleanupContinuation(BasicBlock* continuation);
3650 void fgUpdateFinallyTargetFlags();
3652 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3653 BasicBlock* handler,
3654 BlockToBlockMap& continuationMap);
3656 GenTreePtr fgGetCritSectOfStaticMethod();
3658 #if FEATURE_EH_FUNCLETS
3660 void fgAddSyncMethodEnterExit();
3662 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3664 void fgConvertSyncReturnToLeave(BasicBlock* block);
3666 #endif // FEATURE_EH_FUNCLETS
3668 void fgAddReversePInvokeEnterExit();
3670 bool fgMoreThanOneReturnBlock();
3672 // The number of separate return points in the method.
3673 unsigned fgReturnCount;
3675 void fgAddInternal();
3677 bool fgFoldConditional(BasicBlock* block);
3679 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3680 void fgMorphBlocks();
3682 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3684 void fgCheckArgCnt();
3685 void fgSetOptions();
3688 static fgWalkPreFn fgAssertNoQmark;
3689 void fgPreExpandQmarkChecks(GenTreePtr expr);
3690 void fgPostExpandQmarkChecks();
3691 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3694 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3696 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3697 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3698 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3699 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3700 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3702 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3703 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3704 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3705 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3707 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3708 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3709 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3710 void fgExpandQmarkNodes();
3714 // Do "simple lowering." This functionality is (conceptually) part of "general"
3715 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3716 void fgSimpleLowering();
3718 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3720 GenTreePtr fgInitThisClass();
3722 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3724 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3726 void fgLocalVarLiveness();
3728 void fgLocalVarLivenessInit();
3730 #ifdef LEGACY_BACKEND
3731 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3733 void fgPerNodeLocalVarLiveness(GenTree* node);
3735 void fgPerBlockLocalVarLiveness();
3737 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3739 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3741 // This is used in the liveness computation, as a temporary. When we use the
3742 // arbitrary-length VarSet representation, it is better not to allocate a new one
3744 VARSET_TP fgMarkIntfUnionVS;
3746 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3748 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3750 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3752 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3754 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3756 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3758 void fgComputeLife(VARSET_TP& life,
3759 GenTreePtr startNode,
3761 VARSET_VALARG_TP volatileVars,
3762 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3764 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3766 bool fgRemoveDeadStore(GenTree** pTree,
3768 VARSET_VALARG_TP life,
3770 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3772 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3774 // For updating liveset during traversal AFTER fgComputeLife has completed
3775 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3776 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3778 // Returns the set of live variables after endTree,
3779 // assuming that liveSet is the set of live variables BEFORE tree.
3780 // Requires that fgComputeLife has completed, and that tree is in the same
3781 // statement as endTree, and that it comes before endTree in execution order
3783 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3785 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3786 while (tree != nullptr && tree != endTree->gtNext)
3788 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3789 tree = tree->gtNext;
3791 assert(tree == endTree->gtNext);
3795 void fgInterBlockLocalVarLiveness();
3797 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3798 // "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
3799 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3800 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3801 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3802 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3803 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3805 if (m_opAsgnVarDefSsaNums == nullptr)
3807 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3809 return m_opAsgnVarDefSsaNums;
3812 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3813 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3814 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3816 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3818 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3819 // Except: assumes that lcl is a def, and if it is
3820 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3821 // rather than the "use" SSA number recorded in the tree "lcl".
3822 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3824 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3825 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3826 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3827 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3828 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3830 // (byref addrS1 = &s1,
3831 // *(addrS1 * offsetof(f0)) = s2f0,
3833 // *(addrS1 * offsetof(fn)) = s2fn)
3835 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3836 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3837 // give it SSA names and value numbers?
3839 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3840 // end with an instance of the structure below, whose fields are described in the declaration.
3841 struct IndirectAssignmentAnnotation
3843 unsigned m_lclNum; // The local num that is being indirectly assigned.
3844 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3845 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3846 // be the singleton field sequence "g". The individual assignments would
3847 // further append the fields of "s.g" to that.
3848 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3849 // structure has a single field).
3850 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3851 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3854 IndirectAssignmentAnnotation(unsigned lclNum,
3855 FieldSeqNode* fldSeq,
3857 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3858 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3859 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3863 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3864 NodeToIndirAssignMap;
3865 NodeToIndirAssignMap* m_indirAssignMap;
3866 NodeToIndirAssignMap* GetIndirAssignMap()
3868 if (m_indirAssignMap == nullptr)
3870 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3871 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3872 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3874 return m_indirAssignMap;
3877 // Performs SSA conversion.
3880 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3881 void fgResetForSsa();
3883 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3885 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3886 inline bool fgExcludeFromSsa(unsigned lclNum);
3888 // The value numbers for this compilation.
3889 ValueNumStore* vnStore;
3892 ValueNumStore* GetValueNumStore()
3897 // Do value numbering (assign a value number to each
3899 void fgValueNumber();
3901 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3902 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3903 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3904 // match the element type of the array or fldSeq. When this type doesn't match
3905 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3907 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3910 FieldSeqNode* fldSeq,
3914 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3915 // has been parsed to yield the other input arguments. If evaluation of the address
3916 // can raise exceptions, those should be captured in the exception set "excVN."
3917 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3918 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3919 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3920 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3921 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3923 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3924 CORINFO_CLASS_HANDLE elemTypeEq,
3928 FieldSeqNode* fldSeq);
3930 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3931 // by evaluating the array index expression "tree". Returns the value number resulting from
3932 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3933 // "GT_IND" that does the dereference, and it is given the returned value number.
3934 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3936 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3937 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3939 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3941 // Utility functions for fgValueNumber.
3943 // Perform value-numbering for the trees in "blk".
3944 void fgValueNumberBlock(BasicBlock* blk);
3946 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3947 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3948 // assumed for the memoryKind at the start "entryBlk".
3949 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3951 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3952 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3953 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3955 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3957 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3959 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3960 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3961 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3963 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3964 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3966 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3967 // value in that SSA #.
3968 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3970 // The input 'tree' is a leaf node that is a constant
3971 // Assign the proper value number to the tree
3972 void fgValueNumberTreeConst(GenTreePtr tree);
3974 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3975 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3977 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3979 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3981 // Does value-numbering for a block assignment.
3982 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3984 // Does value-numbering for a cast tree.
3985 void fgValueNumberCastTree(GenTreePtr tree);
3987 // Does value-numbering for an intrinsic tree.
3988 void fgValueNumberIntrinsic(GenTreePtr tree);
3990 // Does value-numbering for a call. We interpret some helper calls.
3991 void fgValueNumberCall(GenTreeCall* call);
3993 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3994 void fgUpdateArgListVNs(GenTreeArgList* args);
3996 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3997 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3999 // Requires "helpCall" to be a helper call. Assigns it a value number;
4000 // we understand the semantics of some of the calls. Returns "true" if
4001 // the call may modify the heap (we assume arbitrary memory side effects if so).
4002 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4004 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4005 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4007 // These are the current value number for the memory implicit variables while
4008 // doing value numbering. These are the value numbers under the "liberal" interpretation
4009 // of memory values; the "conservative" interpretation needs no VN, since every access of
4010 // memory yields an unknown value.
4011 ValueNum fgCurMemoryVN[MemoryKindCount];
4013 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4014 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4015 // is 1, and the rest is an encoding of "elemTyp".
4016 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4018 if (elemStructType != nullptr)
4020 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4021 varTypeIsIntegral(elemTyp));
4022 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4023 return elemStructType;
4027 elemTyp = varTypeUnsignedToSigned(elemTyp);
4028 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4031 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4032 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4033 // the struct type of the element).
4034 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4036 size_t clsHndVal = size_t(clsHnd);
4037 if (clsHndVal & 0x1)
4039 return var_types(clsHndVal >> 1);
4047 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4048 var_types getJitGCType(BYTE gcType);
4050 enum structPassingKind
4052 SPK_Unknown, // Invalid value, never returned
4053 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4054 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4055 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4056 // parameters registers are used, then the stack will be used)
4057 // for X86 passed on the stack, for ARM32 passed in registers
4058 // or the stack or split between registers and the stack.
4059 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4061 }; // The struct is passed/returned by reference to a copy/buffer.
4063 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4064 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4065 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4066 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4068 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4070 // Get the type that is used to pass values of the given struct type.
4071 // If you have already retrieved the struct size then pass it as the optional third argument
4073 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4074 structPassingKind* wbPassStruct,
4075 unsigned structSize = 0);
4077 // Get the type that is used to return values of the given struct type.
4078 // If you have already retrieved the struct size then pass it as the optional third argument
4080 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4081 structPassingKind* wbPassStruct = nullptr,
4082 unsigned structSize = 0);
4085 // Print a representation of "vnp" or "vn" on standard output.
4086 // If "level" is non-zero, we also print out a partial expansion of the value.
4087 void vnpPrint(ValueNumPair vnp, unsigned level);
4088 void vnPrint(ValueNum vn, unsigned level);
4091 // Dominator computation member functions
4092 // Not exposed outside Compiler
4094 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4096 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4098 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4099 // flow graph. We first assume the fields bbIDom on each
4100 // basic block are invalid. This computation is needed later
4101 // by fgBuildDomTree to build the dominance tree structure.
4102 // Based on: A Simple, Fast Dominance Algorithm
4103 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4105 void fgCompDominatedByExceptionalEntryBlocks();
4107 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4108 // Note: this is relatively slow compared to calling fgDominate(),
4109 // especially if dealing with a single block versus block check.
4111 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4113 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4115 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4117 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4119 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4121 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4122 // processed in topological sort, this function takes care of that.
4124 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4126 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4127 // Returns this as a set.
4129 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4130 // root nodes. Returns this as a set.
4133 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4136 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4137 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4140 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4141 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4142 // && postOrder(A) >= postOrder(B) making the computation O(1).
4143 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4145 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4147 void fgUpdateChangedFlowGraph();
4150 // Compute the predecessors of the blocks in the control flow graph.
4151 void fgComputePreds();
4153 // Remove all predecessor information.
4154 void fgRemovePreds();
4156 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4157 // before the full predecessors lists are computed.
4158 void fgComputeCheapPreds();
4161 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4163 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4173 // Initialize the per-block variable sets (used for liveness analysis).
4174 void fgInitBlockVarSets();
4176 // true if we've gone through and created GC Poll calls.
4177 bool fgGCPollsCreated;
4178 void fgMarkGCPollBlocks();
4179 void fgCreateGCPolls();
4180 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4182 // Requires that "block" is a block that returns from
4183 // a finally. Returns the number of successors (jump targets of
4184 // of blocks in the covered "try" that did a "LEAVE".)
4185 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4187 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4188 // a finally. Returns its "i"th successor (jump targets of
4189 // of blocks in the covered "try" that did a "LEAVE".)
4190 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4191 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4194 // Factor out common portions of the impls of the methods above.
4195 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4198 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4199 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4200 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4201 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4202 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4203 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4204 // we leave the entry associated with the block, but it will no longer be accessed.)
4205 struct SwitchUniqueSuccSet
4207 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4208 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4211 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4212 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4213 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4214 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4217 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4218 BlockToSwitchDescMap;
4221 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4222 // iteration over only the distinct successors.
4223 BlockToSwitchDescMap* m_switchDescMap;
4226 BlockToSwitchDescMap* GetSwitchDescMap()
4228 if (m_switchDescMap == nullptr)
4230 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4232 return m_switchDescMap;
4235 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4236 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4237 // we don't accidentally look up and return the wrong switch data.
4238 void InvalidateUniqueSwitchSuccMap()
4240 m_switchDescMap = nullptr;
4243 // Requires "switchBlock" to be a block that ends in a switch. Returns
4244 // the corresponding SwitchUniqueSuccSet.
4245 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4247 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4248 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4249 // remove it from "this", and ensure that "to" is a member.
4250 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4252 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4253 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4255 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4257 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4259 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4261 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4263 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4265 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4267 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4269 void fgRemoveBlockAsPred(BasicBlock* block);
4271 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4273 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4275 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4277 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4279 flowList* fgAddRefPred(BasicBlock* block,
4280 BasicBlock* blockPred,
4281 flowList* oldEdge = nullptr,
4282 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4285 void fgFindBasicBlocks();
4287 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4289 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4291 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4292 bool putInTryRegion,
4293 BasicBlock* startBlk,
4295 BasicBlock* nearBlk,
4296 BasicBlock* jumpBlk,
4299 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4301 void fgRemoveEmptyBlocks();
4303 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4305 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4307 void fgCreateLoopPreHeader(unsigned lnum);
4309 void fgUnreachableBlock(BasicBlock* block);
4311 void fgRemoveConditionalJump(BasicBlock* block);
4313 BasicBlock* fgLastBBInMainFunction();
4315 BasicBlock* fgEndBBAfterMainFunction();
4317 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4319 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4321 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4323 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4325 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4327 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4329 bool fgRenumberBlocks();
4331 bool fgExpandRarelyRunBlocks();
4333 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4335 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4337 enum FG_RELOCATE_TYPE
4339 FG_RELOCATE_TRY, // relocate the 'try' region
4340 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4342 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4344 #if FEATURE_EH_FUNCLETS
4345 #if defined(_TARGET_ARM_)
4346 void fgClearFinallyTargetBit(BasicBlock* block);
4347 #endif // defined(_TARGET_ARM_)
4348 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4349 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4350 void fgInsertFuncletPrologBlock(BasicBlock* block);
4351 void fgCreateFuncletPrologBlocks();
4352 void fgCreateFunclets();
4353 #else // !FEATURE_EH_FUNCLETS
4354 bool fgRelocateEHRegions();
4355 #endif // !FEATURE_EH_FUNCLETS
4357 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4359 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4361 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4363 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4365 bool fgOptimizeEmptyBlock(BasicBlock* block);
4367 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4369 bool fgOptimizeBranch(BasicBlock* bJump);
4371 bool fgOptimizeSwitchBranches(BasicBlock* block);
4373 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4375 bool fgOptimizeSwitchJumps();
4377 void fgPrintEdgeWeights();
4379 void fgComputeEdgeWeights();
4381 void fgReorderBlocks();
4383 void fgDetermineFirstColdBlock();
4385 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4387 bool fgUpdateFlowGraph(bool doTailDup = false);
4389 void fgFindOperOrder();
4391 // method that returns if you should split here
4392 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4394 void fgSetBlockOrder();
4396 void fgRemoveReturnBlock(BasicBlock* block);
4398 /* Helper code that has been factored out */
4399 inline void fgConvertBBToThrowBB(BasicBlock* block);
4401 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4402 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4403 GenTreePtr fgMakeTmpArgNode(
4404 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4406 // The following check for loops that don't execute calls
4407 bool fgLoopCallMarked;
4409 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4410 void fgLoopCallMark();
4412 void fgMarkLoopHead(BasicBlock* block);
4414 unsigned fgGetCodeEstimate(BasicBlock* block);
4417 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4418 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4419 bool fgDumpFlowGraph(Phases phase);
4421 #endif // DUMP_FLOWGRAPHS
4426 void fgDispBBLiveness(BasicBlock* block);
4427 void fgDispBBLiveness();
4428 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4429 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4430 void fgDispBasicBlocks(bool dumpTrees = false);
4431 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4432 void fgDumpBlock(BasicBlock* block);
4433 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4435 static fgWalkPreFn fgStress64RsltMulCB;
4436 void fgStress64RsltMul();
4437 void fgDebugCheckUpdate();
4438 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4439 void fgDebugCheckBlockLinks();
4440 void fgDebugCheckLinks(bool morphTrees = false);
4441 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4442 void fgDebugCheckFlags(GenTreePtr tree);
4443 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4444 void fgDebugCheckTryFinallyExits();
4447 #ifdef LEGACY_BACKEND
4448 static void fgOrderBlockOps(GenTreePtr tree,
4452 GenTreePtr* opsPtr, // OUT
4453 regMaskTP* regsPtr); // OUT
4454 #endif // LEGACY_BACKEND
4456 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4457 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4459 inline bool fgIsInlining()
4461 return fgExpandInline;
4464 void fgTraverseRPO();
4466 //--------------------- Walking the trees in the IR -----------------------
4471 fgWalkPreFn* wtprVisitorFn;
4472 fgWalkPostFn* wtpoVisitorFn;
4473 void* pCallbackData; // user-provided data
4474 bool wtprLclsOnly; // whether to only visit lclvar nodes
4475 GenTreePtr parent; // parent of current node, provided to callback
4476 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4478 bool printModified; // callback can use this
4482 template <bool computeStack>
4483 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4485 // general purpose tree-walker that is capable of doing pre- and post- order
4486 // callbacks at the same time
4487 template <bool doPreOrder, bool doPostOrder>
4488 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4490 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4491 fgWalkPreFn* visitor,
4492 void* pCallBackData = nullptr,
4493 bool lclVarsOnly = false,
4494 bool computeStack = false);
4496 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4497 fgWalkPreFn* preVisitor,
4498 fgWalkPostFn* postVisitor,
4499 void* pCallBackData = nullptr);
4501 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4505 template <bool computeStack>
4506 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4508 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4509 fgWalkPostFn* visitor,
4510 void* pCallBackData = nullptr,
4511 bool computeStack = false);
4513 // An fgWalkPreFn that looks for expressions that have inline throws in
4514 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4515 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4516 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4517 // properly propagated to parent trees). It returns WALK_CONTINUE
4519 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4520 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4521 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4523 /**************************************************************************
4525 *************************************************************************/
4528 friend class SsaBuilder;
4529 friend struct ValueNumberState;
4531 //--------------------- Detect the basic blocks ---------------------------
4533 BasicBlock** fgBBs; // Table of pointers to the BBs
4535 void fgInitBBLookup();
4536 BasicBlock* fgLookupBB(unsigned addr);
4538 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4540 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4542 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4544 void fgLinkBasicBlocks();
4546 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4548 void fgCheckBasicBlockControlFlow();
4550 void fgControlFlowPermitted(BasicBlock* blkSrc,
4551 BasicBlock* blkDest,
4552 BOOL IsLeave = false /* is the src a leave block */);
4554 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4556 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4558 void fgAdjustForAddressExposedOrWrittenThis();
4560 bool fgProfileData_ILSizeMismatch;
4561 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4562 ULONG fgProfileBufferCount;
4563 ULONG fgNumProfileRuns;
4565 unsigned fgStressBBProf()
4568 unsigned result = JitConfig.JitStressBBProf();
4571 if (compStressCompile(STRESS_BB_PROFILE, 15))
4582 bool fgHaveProfileData();
4583 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4584 void fgInstrumentMethod();
4587 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4588 // or if we have some fake profile data for the stress mode
4589 bool fgIsUsingProfileWeights()
4591 return (fgHaveProfileData() || fgStressBBProf());
4594 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4595 // or BB_UNITY_WEIGHT when we aren't using profile data.
4596 unsigned fgProfileRunsCount()
4598 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4601 //-------- Insert a statement at the start or end of a basic block --------
4605 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4609 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4611 public: // Used by linear scan register allocation
4612 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4615 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4616 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4618 public: // Used by linear scan register allocation
4619 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4622 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4624 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4626 // Create a new temporary variable to hold the result of *ppTree,
4627 // and transform the graph accordingly.
4628 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4629 GenTree* fgMakeMultiUse(GenTree** ppTree);
4632 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4633 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4634 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4636 //-------- Determine the order in which the trees will be evaluated -------
4638 unsigned fgTreeSeqNum;
4639 GenTree* fgTreeSeqLst;
4640 GenTree* fgTreeSeqBeg;
4642 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4643 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4644 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4645 void fgSetStmtSeq(GenTree* tree);
4646 void fgSetBlockOrder(BasicBlock* block);
4648 //------------------------- Morphing --------------------------------------
4650 unsigned fgPtrArgCntCur;
4651 unsigned fgPtrArgCntMax;
4652 hashBv* fgOutgoingArgTemps;
4653 hashBv* fgCurrentlyInUseArgTemps;
4655 bool compCanEncodePtrArgCntMax();
4657 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4660 void fgMoveOpsLeft(GenTreePtr tree);
4663 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4665 bool fgIsThrow(GenTreePtr tree);
4667 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4668 bool fgIsBlockCold(BasicBlock* block);
4670 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4672 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4674 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4676 bool fgMorphRelopToQmark(GenTreePtr tree);
4678 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4679 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4680 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4681 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4682 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4683 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4684 // small; hence the other fields of MorphAddrContext.
4685 enum MorphAddrContextKind
4690 struct MorphAddrContext
4692 MorphAddrContextKind m_kind;
4693 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4694 // top-level indirection and here have been constants.
4695 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4696 // In that case, is the sum of those constant offsets.
4698 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4703 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4704 static MorphAddrContext s_CopyBlockMAC;
4707 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4708 var_types* baseTypeOut,
4710 unsigned* simdSizeOut,
4711 bool ignoreUsedInSIMDIntrinsic = false);
4712 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4713 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4714 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4715 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4717 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4718 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4719 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4721 #endif // FEATURE_SIMD
4722 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4723 GenTreePtr fgMorphCast(GenTreePtr tree);
4724 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4725 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4727 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4730 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4731 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4733 void fgFixupStructReturn(GenTreePtr call);
4734 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4735 bool fgAddrCouldBeNull(GenTreePtr addr);
4736 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4737 bool fgCanFastTailCall(GenTreeCall* call);
4738 void fgMorphTailCall(GenTreeCall* call);
4739 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4740 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4741 fgArgTabEntryPtr argTabEntry,
4743 IL_OFFSETX callILOffset,
4744 GenTreePtr tmpAssignmentInsertionPoint,
4745 GenTreePtr paramAssignmentInsertionPoint);
4746 static int fgEstimateCallStackSize(GenTreeCall* call);
4747 GenTreePtr fgMorphCall(GenTreeCall* call);
4748 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4749 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4751 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4752 static fgWalkPreFn fgFindNonInlineCandidate;
4754 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4755 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4756 CORINFO_RESOLVED_TOKEN* ldftnToken);
4757 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4758 void fgAssignSetVarDef(GenTreePtr tree);
4759 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4760 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4761 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4762 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4763 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4764 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4765 void fgMorphUnsafeBlk(GenTreeObj* obj);
4766 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4767 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4768 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4769 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4770 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4771 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4772 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4774 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4775 GenTreePtr fgMorphConst(GenTreePtr tree);
4778 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4781 #if LOCAL_ASSERTION_PROP
4782 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4783 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4785 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4787 GenTreeStmt* fgMorphStmt;
4789 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4790 // used when morphing big offset.
4792 //----------------------- Liveness analysis -------------------------------
4794 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4795 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4797 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4798 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4799 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4801 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4803 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4805 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4806 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4808 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4809 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4811 void fgExtendDbgScopes();
4812 void fgExtendDbgLifetimes();
4815 void fgDispDebugScopes();
4818 //-------------------------------------------------------------------------
4820 // The following keeps track of any code we've added for things like array
4821 // range checking or explicit calls to enable GC, and so on.
4826 AddCodeDsc* acdNext;
4827 BasicBlock* acdDstBlk; // block to which we jump
4829 SpecialCodeKind acdKind; // what kind of a special block is this?
4830 unsigned short acdStkLvl;
4834 static unsigned acdHelper(SpecialCodeKind codeKind);
4836 AddCodeDsc* fgAddCodeList;
4838 bool fgRngChkThrowAdded;
4839 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4841 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4843 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4846 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4849 bool fgIsCodeAdded();
4851 bool fgIsThrowHlpBlk(BasicBlock* block);
4852 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4854 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4856 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4857 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4858 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4859 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4860 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4862 #if FEATURE_MULTIREG_RET
4863 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4864 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4865 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4866 #endif // FEATURE_MULTIREG_RET
4868 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4871 static fgWalkPreFn fgDebugCheckInlineCandidates;
4873 void CheckNoFatPointerCandidatesLeft();
4874 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4877 void fgPromoteStructs();
4878 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4879 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4881 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4882 void fgMarkImplicitByRefArgs();
4884 // Change implicit byrefs' types from struct to pointer, and for any that were
4885 // promoted, create new promoted struct temps.
4886 void fgRetypeImplicitByRefArgs();
4888 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4889 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4890 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4892 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4893 void fgMarkDemotedImplicitByRefArgs();
4895 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4896 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4897 void fgMarkAddressExposedLocals();
4898 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4900 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4902 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4904 // The given local variable, required to be a struct variable, is being assigned via
4905 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4906 // the variable is not enregistered, and is therefore not promoted independently.
4907 void fgLclFldAssign(unsigned lclNum);
4909 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4910 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4911 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4912 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4915 bool fgPrintInlinedMethods;
4918 bool fgIsBigOffset(size_t offset);
4920 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4921 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4922 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4923 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4924 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4927 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4928 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4932 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4933 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4940 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4943 void optRemoveRangeCheck(
4944 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4945 bool optIsRangeCheckRemovable(GenTreePtr tree);
4948 static fgWalkPreFn optValidRangeCheckIndex;
4949 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4952 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4954 /**************************************************************************
4956 *************************************************************************/
4959 // Do hoisting for all loops.
4960 void optHoistLoopCode();
4962 // To represent sets of VN's that have already been hoisted in outer loops.
4963 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4964 typedef VNToBoolMap VNSet;
4966 struct LoopHoistContext
4969 // The set of variables hoisted in the current loop (or nullptr if there are none).
4970 VNSet* m_pHoistedInCurLoop;
4973 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4974 VNSet m_hoistedInParentLoops;
4975 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4976 // Previous decisions on loop-invariance of value numbers in the current loop.
4977 VNToBoolMap m_curLoopVnInvariantCache;
4979 VNSet* GetHoistedInCurLoop(Compiler* comp)
4981 if (m_pHoistedInCurLoop == nullptr)
4983 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4985 return m_pHoistedInCurLoop;
4988 VNSet* ExtractHoistedInCurLoop()
4990 VNSet* res = m_pHoistedInCurLoop;
4991 m_pHoistedInCurLoop = nullptr;
4995 LoopHoistContext(Compiler* comp)
4996 : m_pHoistedInCurLoop(nullptr)
4997 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4998 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5003 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5004 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5005 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5006 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5008 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5009 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5010 // "m_hoistedInParentLoops".
5012 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5014 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5015 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5016 // expressions to "hoistInLoop".
5017 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5019 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5020 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5022 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5023 // that are invariant in loop "lnum" (an index into the optLoopTable)
5024 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5025 // expressions to "hoistInLoop".
5026 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5027 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5028 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5029 bool optHoistLoopExprsForTree(GenTreePtr tree,
5031 LoopHoistContext* hoistCtxt,
5032 bool* firstBlockAndBeforeSideEffect,
5034 bool* pCctorDependent);
5036 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5037 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5039 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5040 // Constants and init values are always loop invariant.
5041 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5042 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5044 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5045 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5046 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5047 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5048 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5050 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5051 // in the loop table.
5052 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5054 // Records the set of "side effects" of all loops: fields (object instance and static)
5055 // written to, and SZ-array element type equivalence classes updated.
5056 void optComputeLoopSideEffects();
5059 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5060 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5061 // static) written to, and SZ-array element type equivalence classes updated.
5062 void optComputeLoopNestSideEffects(unsigned lnum);
5064 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5065 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5067 // Hoist the expression "expr" out of loop "lnum".
5068 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5071 void optOptimizeBools();
5074 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5076 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5079 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5081 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5082 // the loop into a "do-while" loop
5083 // Also finds all natural loops and records them in the loop table
5085 // Optionally clone loops in the loop table.
5086 void optCloneLoops();
5088 // Clone loop "loopInd" in the loop table.
5089 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5091 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5092 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5093 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5095 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5097 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5100 // This enumeration describes what is killed by a call.
5104 CALLINT_NONE, // no interference (most helpers)
5105 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5106 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5107 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5108 CALLINT_ALL, // kills everything (normal method call)
5112 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5113 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5114 // in bbNext order; we use comparisons on the bbNum to decide order.)
5115 // The blocks that define the body are
5116 // first <= top <= entry <= bottom .
5117 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5118 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5119 // Compiler::optFindNaturalLoops().
5122 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5123 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5124 // loop, but not the outer loop.)
5125 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5127 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5128 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5129 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5131 callInterf lpAsgCall; // "callInterf" for calls in the loop
5132 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5133 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5135 unsigned short lpFlags; // Mask of the LPFLG_* constants
5137 unsigned char lpExitCnt; // number of exits from the loop
5139 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5140 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5141 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5142 // (Actually, an "immediately" nested loop --
5143 // no other child of this loop is a parent of lpChild.)
5144 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5145 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5146 // by following "lpChild" then "lpSibling" links.
5148 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5149 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5151 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5152 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5153 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5155 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5156 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5158 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5159 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5160 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5161 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5163 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5164 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5165 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5167 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5168 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5169 // type are assigned to.
5171 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5172 // memory side effects. If this is set, the fields below
5173 // may not be accurate (since they become irrelevant.)
5174 bool lpContainsCall; // True if executing the loop body *may* execute a call
5176 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5177 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5179 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5181 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5182 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5184 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5186 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5187 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5189 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5190 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5192 JitSimplerHashBehavior>
5194 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5195 // instance fields modified
5198 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5199 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5201 JitSimplerHashBehavior>
5203 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5204 // arrays of that type are modified
5207 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5208 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5210 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5211 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5212 // (shifted left, with a low-order bit set to distinguish.)
5213 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5214 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5216 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5218 GenTreePtr lpIterTree; // The "i <op>= const" tree
5219 unsigned lpIterVar(); // iterator variable #
5220 int lpIterConst(); // the constant with which the iterator is incremented
5221 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5222 void VERIFY_lpIterTree();
5224 var_types lpIterOperType(); // For overflow instructions
5227 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5228 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5232 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5234 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5235 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5236 void VERIFY_lpTestTree();
5238 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5239 GenTreePtr lpIterator(); // the iterator node in the loop test
5240 GenTreePtr lpLimit(); // the limit node in the loop test
5242 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5243 // LPFLG_CONST_LIMIT
5244 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5246 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5247 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5248 // LPFLG_ARRLEN_LIMIT
5250 // Returns "true" iff "*this" contains the blk.
5251 bool lpContains(BasicBlock* blk)
5253 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5255 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5256 // to be equal, but requiring bottoms to be different.)
5257 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5259 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5262 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5263 // bottoms to be different.)
5264 bool lpContains(const LoopDsc& lp2)
5266 return lpContains(lp2.lpFirst, lp2.lpBottom);
5269 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5270 // (allowing firsts to be equal, but requiring bottoms to be different.)
5271 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5273 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5276 // Returns "true" iff "*this" is (properly) contained by "lp2"
5277 // (allowing firsts to be equal, but requiring bottoms to be different.)
5278 bool lpContainedBy(const LoopDsc& lp2)
5280 return lpContains(lp2.lpFirst, lp2.lpBottom);
5283 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5284 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5286 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5288 // Returns "true" iff "*this" is disjoint from "lp2".
5289 bool lpDisjoint(const LoopDsc& lp2)
5291 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5293 // Returns "true" iff the loop is well-formed (see code for defn).
5296 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5297 lpEntry->bbNum <= lpBottom->bbNum &&
5298 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5303 bool fgMightHaveLoop(); // returns true if there are any backedges
5304 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5307 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5308 unsigned char optLoopCount; // number of tracked loops
5311 unsigned optCallCount; // number of calls made in the method
5312 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5313 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5314 unsigned optLoopsCloned; // number of loops cloned in the current method.
5317 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5318 void optPrintLoopInfo(unsigned loopNum,
5320 BasicBlock* lpFirst,
5322 BasicBlock* lpEntry,
5323 BasicBlock* lpBottom,
5324 unsigned char lpExitCnt,
5326 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5327 void optPrintLoopInfo(unsigned lnum);
5328 void optPrintLoopRecording(unsigned lnum);
5330 void optCheckPreds();
5333 void optSetBlockWeights();
5335 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5337 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5339 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5341 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5342 unsigned optIsLoopIncrTree(GenTreePtr incr);
5343 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5344 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5345 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5346 bool optExtractInitTestIncr(BasicBlock* head,
5351 GenTreePtr* ppIncr);
5353 void optRecordLoop(BasicBlock* head,
5359 unsigned char exitCnt);
5361 void optFindNaturalLoops();
5363 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5364 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5365 bool optCanonicalizeLoopNest(unsigned char loopInd);
5367 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5368 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5369 bool optCanonicalizeLoop(unsigned char loopInd);
5371 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5372 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5373 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5374 bool optLoopContains(unsigned l1, unsigned l2);
5376 // Requires "loopInd" to be a valid index into the loop table.
5377 // Updates the loop table by changing loop "loopInd", whose head is required
5378 // to be "from", to be "to". Also performs this transformation for any
5379 // loop nested in "loopInd" that shares the same head as "loopInd".
5380 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5382 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5383 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5384 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5386 // Marks the containsCall information to "lnum" and any parent loops.
5387 void AddContainsCallAllContainingLoops(unsigned lnum);
5388 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5389 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5390 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5391 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5392 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5393 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5395 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5396 // of "from".) Copies the jump destination from "from" to "to".
5397 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5399 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5400 unsigned optLoopDepth(unsigned lnum)
5402 unsigned par = optLoopTable[lnum].lpParent;
5403 if (par == BasicBlock::NOT_IN_LOOP)
5409 return 1 + optLoopDepth(par);
5413 void fgOptWhileLoop(BasicBlock* block);
5415 bool optComputeLoopRep(int constInit,
5418 genTreeOps iterOper,
5420 genTreeOps testOper,
5423 unsigned* iterCount);
5424 #if FEATURE_STACK_FP_X87
5427 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5428 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5429 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5430 #endif // FEATURE_STACK_FP_X87
5433 static fgWalkPreFn optIsVarAssgCB;
5436 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5438 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5440 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5442 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5444 /**************************************************************************
5445 * Optimization conditions
5446 *************************************************************************/
5448 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5449 bool optPentium4(void);
5450 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5451 bool optAvoidIntMult(void);
5456 // The following is the upper limit on how many expressions we'll keep track
5457 // of for the CSE analysis.
5459 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5461 static const int MIN_CSE_COST = 2;
5463 // Keeps tracked cse indices
5464 BitVecTraits* cseTraits;
5467 /* Generic list of nodes - used by the CSE logic */
5475 typedef struct treeLst* treeLstPtr;
5479 treeStmtLst* tslNext;
5480 GenTreePtr tslTree; // tree node
5481 GenTreePtr tslStmt; // statement containing the tree
5482 BasicBlock* tslBlock; // block containing the statement
5485 typedef struct treeStmtLst* treeStmtLstPtr;
5487 // The following logic keeps track of expressions via a simple hash table.
5491 CSEdsc* csdNextInBucket; // used by the hash table
5493 unsigned csdHashValue; // the orginal hashkey
5495 unsigned csdIndex; // 1..optCSECandidateCount
5496 char csdLiveAcrossCall; // 0 or 1
5498 unsigned short csdDefCount; // definition count
5499 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5501 unsigned csdDefWtCnt; // weighted def count
5502 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5504 GenTreePtr csdTree; // treenode containing the 1st occurance
5505 GenTreePtr csdStmt; // stmt containing the 1st occurance
5506 BasicBlock* csdBlock; // block containing the 1st occurance
5508 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5509 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5511 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5512 // number, this will reflect it; otherwise, NoVN.
5515 static const size_t s_optCSEhashSize;
5516 CSEdsc** optCSEhash;
5519 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5521 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5522 // re-numbered with the bound to improve range check elimination
5524 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5525 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5529 CSEdsc* optCSEfindDsc(unsigned index);
5530 void optUnmarkCSE(GenTreePtr tree);
5532 // user defined callback data for the tree walk function optCSE_MaskHelper()
5533 struct optCSE_MaskData
5535 EXPSET_TP CSE_defMask;
5536 EXPSET_TP CSE_useMask;
5539 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5540 static fgWalkPreFn optCSE_MaskHelper;
5542 // This function walks all the node for an given tree
5543 // and return the mask of CSE definitions and uses for the tree
5545 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5547 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5548 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5549 bool optCSE_canSwap(GenTree* tree);
5551 static fgWalkPostFn optPropagateNonCSE;
5552 static fgWalkPreFn optHasNonCSEChild;
5554 static fgWalkPreFn optUnmarkCSEs;
5556 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5557 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5559 void optCleanupCSEs();
5562 void optEnsureClearCSEInfo();
5565 #endif // FEATURE_ANYCSE
5567 #if FEATURE_VALNUM_CSE
5568 /**************************************************************************
5569 * Value Number based CSEs
5570 *************************************************************************/
5573 void optOptimizeValnumCSEs();
5576 void optValnumCSE_Init();
5577 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5578 unsigned optValnumCSE_Locate();
5579 void optValnumCSE_InitDataFlow();
5580 void optValnumCSE_DataFlow();
5581 void optValnumCSE_Availablity();
5582 void optValnumCSE_Heuristic();
5583 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5585 #endif // FEATURE_VALNUM_CSE
5588 bool optDoCSE; // True when we have found a duplicate CSE tree
5589 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5590 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5591 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5592 unsigned optCSEstart; // The first local variable number that is a CSE
5593 unsigned optCSEcount; // The total count of CSE's introduced.
5594 unsigned optCSEweight; // The weight of the current block when we are
5595 // scanning for CSE expressions
5597 bool optIsCSEcandidate(GenTreePtr tree);
5599 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5601 bool lclNumIsTrueCSE(unsigned lclNum) const
5603 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5606 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5608 bool lclNumIsCSE(unsigned lclNum) const
5610 return lvaTable[lclNum].lvIsCSE;
5614 bool optConfigDisableCSE();
5615 bool optConfigDisableCSE2();
5617 void optOptimizeCSEs();
5619 #endif // FEATURE_ANYCSE
5627 unsigned ivaVar; // Variable we are interested in, or -1
5628 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5629 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5630 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5631 callInterf ivaMaskCall; // What kind of calls are there?
5634 static callInterf optCallInterf(GenTreeCall* call);
5637 // VN based copy propagation.
5638 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5639 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5640 LclNumToGenTreePtrStack;
5642 // Kill set to track variables with intervening definitions.
5643 VARSET_TP optCopyPropKillSet;
5645 // Copy propagation functions.
5646 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5647 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5648 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5649 bool optIsSsaLocal(GenTreePtr tree);
5650 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5651 void optVnCopyProp();
5653 /**************************************************************************
5654 * Early value propagation
5655 *************************************************************************/
5661 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5665 static unsigned GetHashCode(SSAName ssaNm)
5667 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5670 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5672 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5676 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5677 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5678 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5679 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5680 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5681 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5683 bool doesMethodHaveFatPointer()
5685 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5688 void setMethodHasFatPointer()
5690 optMethodFlags |= OMF_HAS_FATPOINTER;
5693 void clearMethodHasFatPointer()
5695 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5698 void addFatPointerCandidate(GenTreeCall* call)
5700 setMethodHasFatPointer();
5701 call->SetFatPointerCandidate();
5704 unsigned optMethodFlags;
5706 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5707 // No throughput diff was found with backward walk bound between 3-8.
5708 static const int optEarlyPropRecurBound = 5;
5710 enum class optPropKind
5718 bool gtIsVtableRef(GenTreePtr tree);
5719 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5720 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5721 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5722 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5723 bool optEarlyPropRewriteTree(GenTreePtr tree);
5724 bool optDoEarlyPropForBlock(BasicBlock* block);
5725 bool optDoEarlyPropForFunc();
5726 void optEarlyProp();
5727 void optFoldNullCheck(GenTreePtr tree);
5728 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5731 /**************************************************************************
5732 * Value/Assertion propagation
5733 *************************************************************************/
5735 // Data structures for assertion prop
5736 BitVecTraits* apTraits;
5739 enum optAssertionKind
5756 O1K_CONSTANT_LOOP_BND,
5777 optAssertionKind assertionKind;
5780 unsigned lclNum; // assigned to or property of this local var number
5788 struct AssertionDscOp1
5790 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5797 struct AssertionDscOp2
5799 optOp2Kind kind; // a const or copy assignment
5803 ssize_t iconVal; // integer
5804 unsigned iconFlags; // gtFlags
5806 struct Range // integer subrange
5820 bool IsCheckedBoundArithBound()
5822 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5824 bool IsCheckedBoundBound()
5826 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5828 bool IsConstantBound()
5830 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5831 op1.kind == O1K_CONSTANT_LOOP_BND);
5833 bool IsBoundsCheckNoThrow()
5835 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5838 bool IsCopyAssertion()
5840 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5843 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5845 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5846 a1->op2.kind == a2->op2.kind;
5849 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5851 if (kind == OAK_EQUAL)
5853 return kind2 == OAK_NOT_EQUAL;
5855 else if (kind == OAK_NOT_EQUAL)
5857 return kind2 == OAK_EQUAL;
5862 static ssize_t GetLowerBoundForIntegralType(var_types type)
5882 static ssize_t GetUpperBoundForIntegralType(var_types type)
5906 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5908 if (op1.kind != that->op1.kind)
5912 else if (op1.kind == O1K_ARR_BND)
5915 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5919 return ((vnBased && (op1.vn == that->op1.vn)) ||
5920 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5924 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5926 if (op2.kind != that->op2.kind)
5932 case O2K_IND_CNS_INT:
5934 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5936 case O2K_CONST_LONG:
5937 return (op2.lconVal == that->op2.lconVal);
5939 case O2K_CONST_DOUBLE:
5940 // exact match because of positive and negative zero.
5941 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5943 case O2K_LCLVAR_COPY:
5945 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5946 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5949 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5952 // we will return false
5956 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5962 bool Complementary(AssertionDsc* that, bool vnBased)
5964 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5965 HasSameOp2(that, vnBased);
5968 bool Equals(AssertionDsc* that, bool vnBased)
5970 if (assertionKind != that->assertionKind)
5974 else if (assertionKind == OAK_NO_THROW)
5976 assert(op2.kind == O2K_INVALID);
5977 return HasSameOp1(that, vnBased);
5981 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5987 static fgWalkPreFn optAddCopiesCallback;
5988 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5989 unsigned optAddCopyLclNum;
5990 GenTreePtr optAddCopyAsgnNode;
5992 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5993 bool optAssertionPropagated; // set to true if we modified the trees
5994 bool optAssertionPropagatedCurrentStmt;
5996 GenTreePtr optAssertionPropCurrentTree;
5998 AssertionIndex* optComplementaryAssertionMap;
5999 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6000 // using the value of a local var) for each local var
6001 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6002 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6003 AssertionIndex optMaxAssertionCount;
6006 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6007 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6008 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6009 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6010 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6011 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6013 AssertionIndex GetAssertionCount()
6015 return optAssertionCount;
6017 ASSERT_TP* bbJtrueAssertionOut;
6018 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6019 ValueNumToAssertsMap;
6020 ValueNumToAssertsMap* optValueNumToAsserts;
6022 // Assertion prop helpers.
6023 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6024 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6025 void optAssertionInit(bool isLocalProp);
6026 void optAssertionTraitsInit(AssertionIndex assertionCount);
6027 #if LOCAL_ASSERTION_PROP
6028 void optAssertionReset(AssertionIndex limit);
6029 void optAssertionRemove(AssertionIndex index);
6032 // Assertion prop data flow functions.
6033 void optAssertionPropMain();
6034 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6035 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6036 ASSERT_TP* optInitAssertionDataflowFlags();
6037 ASSERT_TP* optComputeAssertionGen();
6039 // Assertion Gen functions.
6040 void optAssertionGen(GenTreePtr tree);
6041 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6042 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6043 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6044 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6045 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6046 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6048 // Assertion creation functions.
6049 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6050 AssertionIndex optCreateAssertion(GenTreePtr op1,
6052 optAssertionKind assertionKind,
6053 AssertionDsc* assertion);
6054 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6056 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6057 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6058 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6060 void optPrintVnAssertionMapping();
6062 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6064 // Used for respective assertion propagations.
6065 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6066 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6067 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6068 bool optAssertionIsNonNull(GenTreePtr op,
6069 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6071 // Used for Relop propagation.
6072 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6073 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6074 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6076 // Assertion prop for lcl var functions.
6077 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6078 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6080 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6081 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6082 const GenTreePtr tree,
6083 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6084 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6086 // Assertion propagation functions.
6087 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6088 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6089 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6090 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6091 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6092 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6093 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6094 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6095 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6096 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6097 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6098 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6100 // Implied assertion functions.
6101 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6102 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6103 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6104 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6107 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6108 void optDebugCheckAssertion(AssertionDsc* assertion);
6109 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6111 void optAddCopies();
6112 #endif // ASSERTION_PROP
6114 /**************************************************************************
6116 *************************************************************************/
6119 struct LoopCloneVisitorInfo
6121 LoopCloneContext* context;
6124 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6125 : context(context), loopNum(loopNum), stmt(nullptr)
6130 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6131 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6132 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6133 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6134 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6135 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6136 void optObtainLoopCloningOpts(LoopCloneContext* context);
6137 bool optIsLoopClonable(unsigned loopInd);
6139 bool optCanCloneLoops();
6142 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6144 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6145 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6146 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6147 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6151 void optInsertLoopCloningStress(BasicBlock* head);
6153 #if COUNT_RANGECHECKS
6154 static unsigned optRangeChkRmv;
6155 static unsigned optRangeChkAll;
6164 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6169 RngChkDsc* rcdNextInBucket; // used by the hash table
6171 unsigned short rcdHashValue; // to make matching faster
6172 unsigned short rcdIndex; // 0..optRngChkCount-1
6174 GenTreePtr rcdTree; // the array index tree
6177 unsigned optRngChkCount;
6178 static const size_t optRngChkHashSize;
6180 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6181 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6183 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6186 bool optLoopsMarked;
6189 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6190 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6194 XX Does the register allocation and puts the remaining lclVars on the stack XX
6196 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6197 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6201 #ifndef LEGACY_BACKEND
6206 #else // LEGACY_BACKEND
6211 #endif // LEGACY_BACKEND
6213 #ifdef LEGACY_BACKEND
6215 void raAssignVars(); // register allocation
6216 #endif // LEGACY_BACKEND
6218 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6220 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6222 void raMarkStkVars();
6225 // Some things are used by both LSRA and regpredict allocators.
6227 FrameType rpFrameType;
6228 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6230 #ifdef LEGACY_BACKEND
6231 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6233 #endif // LEGACY_BACKEND
6235 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6237 #if FEATURE_FP_REGALLOC
6238 enum enumConfigRegisterFP
6240 CONFIG_REGISTER_FP_NONE = 0x0,
6241 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6242 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6243 CONFIG_REGISTER_FP_FULL = 0x3,
6245 enumConfigRegisterFP raConfigRegisterFP();
6246 #endif // FEATURE_FP_REGALLOC
6249 regMaskTP raConfigRestrictMaskFP();
6252 #ifndef LEGACY_BACKEND
6253 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6254 #else // LEGACY_BACKEND
6255 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6256 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6257 bool raNewBlocks; // True is we added killing blocks for FPU registers
6258 unsigned rpPasses; // Number of passes made by the register predicter
6259 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6260 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6261 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6262 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6263 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6264 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6265 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6266 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6267 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6268 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6269 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6270 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6272 bool rpRegAllocDone; // Set to true after we have completed register allocation
6274 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6276 void raSetupArgMasks(RegState* r);
6278 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6280 void raDumpVarIntf(); // Dump the variable to variable interference graph
6281 void raDumpRegIntf(); // Dump the variable to register interference graph
6283 void raAdjustVarIntf();
6285 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6287 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6289 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6290 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6292 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6294 static fgWalkPreFn rpMarkRegIntf;
6296 regMaskTP rpPredictAddressMode(
6297 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6299 void rpPredictRefAssign(unsigned lclNum);
6301 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6303 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6305 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6307 void rpPredictRegUse(); // Entry point
6309 unsigned raPredictTreeRegUse(GenTreePtr tree);
6310 unsigned raPredictListRegUse(GenTreePtr list);
6312 void raSetRegVarOrder(var_types regType,
6313 regNumber* customVarOrder,
6314 unsigned* customVarOrderSize,
6316 regMaskTP avoidReg);
6318 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6319 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6320 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6321 void raAddToStkPredict(unsigned val)
6323 unsigned newStkPredict = rpStkPredict + val;
6324 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6325 rpStkPredict = UINT_MAX - 1;
6327 rpStkPredict = newStkPredict;
6331 #if !FEATURE_FP_REGALLOC
6332 void raDispFPlifeInfo();
6336 regMaskTP genReturnRegForTree(GenTreePtr tree);
6337 #endif // LEGACY_BACKEND
6339 /* raIsVarargsStackArg is called by raMaskStkVars and by
6340 lvaSortByRefCount. It identifies the special case
6341 where a varargs function has a parameter passed on the
6342 stack, other than the special varargs handle. Such parameters
6343 require special treatment, because they cannot be tracked
6344 by the GC (their offsets in the stack are not known
6348 bool raIsVarargsStackArg(unsigned lclNum)
6352 LclVarDsc* varDsc = &lvaTable[lclNum];
6354 assert(varDsc->lvIsParam);
6356 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6358 #else // _TARGET_X86_
6362 #endif // _TARGET_X86_
6365 #ifdef LEGACY_BACKEND
6366 // Records the current prediction, if it's better than any previous recorded prediction.
6367 void rpRecordPrediction();
6368 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6369 void rpUseRecordedPredictionIfBetter();
6371 // Data members used in the methods above.
6372 unsigned rpBestRecordedStkPredict;
6373 struct VarRegPrediction
6375 bool m_isEnregistered;
6376 regNumberSmall m_regNum;
6377 regNumberSmall m_otherReg;
6379 VarRegPrediction* rpBestRecordedPrediction;
6380 #endif // LEGACY_BACKEND
6383 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6384 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6388 XX Get to the class and method info from the Execution Engine given XX
6389 XX tokens for the class and method XX
6391 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6392 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6396 /* These are the different addressing modes used to access a local var.
6397 * The JIT has to report the location of the locals back to the EE
6398 * for debugging purposes.
6404 VLT_REG_BYREF, // this type is currently only used for value types on X64
6407 VLT_STK_BYREF, // this type is currently only used for value types on X64
6421 siVarLocType vlType;
6424 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6426 // VLT_REG_BYREF -- the specified register contains the address of the variable
6434 // VLT_STK -- Any 32 bit value which is on the stack
6435 // eg. [ESP+0x20], or [EBP-0x28]
6436 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6437 // eg. mov EAX, [ESP+0x20]; [EAX]
6441 regNumber vlsBaseReg;
6442 NATIVE_OFFSET vlsOffset;
6445 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6454 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6455 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6463 regNumber vlrssBaseReg;
6464 NATIVE_OFFSET vlrssOffset;
6468 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6469 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6475 regNumber vlsrsBaseReg;
6476 NATIVE_OFFSET vlsrsOffset;
6482 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6483 // eg 2 DWords at [ESP+0x10]
6487 regNumber vls2BaseReg;
6488 NATIVE_OFFSET vls2Offset;
6491 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6492 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6499 // VLT_FIXED_VA -- fixed argument of a varargs function.
6500 // The argument location depends on the size of the variable
6501 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6502 // location of the first arg. This argument can then be accessed
6503 // relative to the position of the first arg
6507 unsigned vlfvOffset;
6514 void* rpValue; // pointer to the in-process
6515 // location of the value.
6521 bool vlIsInReg(regNumber reg);
6522 bool vlIsOnStk(regNumber reg, signed offset);
6525 /*************************************************************************/
6530 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6531 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6532 CORINFO_CALLINFO_FLAGS flags,
6533 CORINFO_CALL_INFO* pResult);
6534 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6536 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6537 CORINFO_ACCESS_FLAGS flags,
6538 CORINFO_FIELD_INFO* pResult);
6542 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6544 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6546 bool IsSuperPMIException(unsigned code)
6548 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6550 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6551 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6552 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6553 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6554 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6555 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6556 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6557 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6561 case EXCEPTIONCODE_DebugBreakorAV:
6562 case EXCEPTIONCODE_MC:
6563 case EXCEPTIONCODE_LWM:
6564 case EXCEPTIONCODE_SASM:
6565 case EXCEPTIONCODE_SSYM:
6566 case EXCEPTIONCODE_CALLUTILS:
6567 case EXCEPTIONCODE_TYPEUTILS:
6568 case EXCEPTIONCODE_ASSERT:
6575 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6576 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6578 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6579 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6582 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6583 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6584 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6586 // VOM info, method sigs
6588 void eeGetSig(unsigned sigTok,
6589 CORINFO_MODULE_HANDLE scope,
6590 CORINFO_CONTEXT_HANDLE context,
6591 CORINFO_SIG_INFO* retSig);
6593 void eeGetCallSiteSig(unsigned sigTok,
6594 CORINFO_MODULE_HANDLE scope,
6595 CORINFO_CONTEXT_HANDLE context,
6596 CORINFO_SIG_INFO* retSig);
6598 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6600 // Method entry-points, instrs
6602 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6604 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6606 CORINFO_EE_INFO eeInfo;
6607 bool eeInfoInitialized;
6609 CORINFO_EE_INFO* eeGetEEInfo();
6611 // Gets the offset of a SDArray's first element
6612 unsigned eeGetArrayDataOffset(var_types type);
6613 // Gets the offset of a MDArray's first element
6614 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6616 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6618 // Returns the page size for the target machine as reported by the EE.
6619 inline size_t eeGetPageSize()
6621 return eeGetEEInfo()->osPageSize;
6624 // Returns the frame size at which we will generate a loop to probe the stack.
6625 inline size_t getVeryLargeFrameSize()
6628 // The looping probe code is 40 bytes, whereas the straight-line probing for
6629 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6630 // or greater, to generate smaller code.
6631 return 2 * eeGetPageSize();
6633 return 3 * eeGetPageSize();
6637 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6639 return eeGetEEInfo()->targetAbi == abi;
6642 inline bool generateCFIUnwindCodes()
6644 #ifdef UNIX_AMD64_ABI
6645 return IsTargetAbi(CORINFO_CORERT_ABI);
6653 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6655 // Debugging support - Line number info
6657 void eeGetStmtOffsets();
6659 unsigned eeBoundariesCount;
6661 struct boundariesDsc
6663 UNATIVE_OFFSET nativeIP;
6665 unsigned sourceReason;
6666 } * eeBoundaries; // Boundaries to report to EE
6667 void eeSetLIcount(unsigned count);
6668 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6672 static void eeDispILOffs(IL_OFFSET offs);
6673 static void eeDispLineInfo(const boundariesDsc* line);
6674 void eeDispLineInfos();
6677 // Debugging support - Local var info
6681 unsigned eeVarsCount;
6683 struct VarResultInfo
6685 UNATIVE_OFFSET startOffset;
6686 UNATIVE_OFFSET endOffset;
6690 void eeSetLVcount(unsigned count);
6691 void eeSetLVinfo(unsigned which,
6692 UNATIVE_OFFSET startOffs,
6693 UNATIVE_OFFSET length,
6698 const siVarLoc& loc);
6702 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6703 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6706 // ICorJitInfo wrappers
6708 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6710 void eeAllocUnwindInfo(BYTE* pHotCode,
6716 CorJitFuncKind funcKind);
6718 void eeSetEHcount(unsigned cEH);
6720 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6722 WORD eeGetRelocTypeHint(void* target);
6724 // ICorStaticInfo wrapper functions
6726 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6728 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6730 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6733 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6734 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6735 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6736 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6738 template <typename ParamType>
6739 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6741 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6744 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6746 // Utility functions
6748 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6751 const wchar_t* eeGetCPString(size_t stringHandle);
6754 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6756 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6757 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6759 static fgWalkPreFn CountSharedStaticHelper;
6760 static bool IsSharedStaticHelper(GenTreePtr tree);
6761 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6763 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6764 // returns true/false if 'field' is a Jit Data offset
6765 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6766 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6767 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6769 /*****************************************************************************/
6774 enum TEMP_USAGE_TYPE
6780 static var_types tmpNormalizeType(var_types type);
6781 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6782 void tmpRlsTemp(TempDsc* temp);
6783 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6786 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6787 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6791 bool tmpAllFree() const;
6794 #ifndef LEGACY_BACKEND
6795 void tmpPreAllocateTemps(var_types type, unsigned count);
6796 #endif // !LEGACY_BACKEND
6799 #ifdef LEGACY_BACKEND
6800 unsigned tmpIntSpillMax; // number of int-sized spill temps
6801 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6802 #endif // LEGACY_BACKEND
6804 unsigned tmpCount; // Number of temps
6805 unsigned tmpSize; // Size of all the temps
6808 // Used by RegSet::rsSpillChk()
6809 unsigned tmpGetCount; // Temps which haven't been released yet
6812 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6814 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6815 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6818 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6819 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6823 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6824 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6828 CodeGenInterface* codeGen;
6830 // The following holds information about instr offsets in terms of generated code.
6834 IPmappingDsc* ipmdNext; // next line# record
6835 IL_OFFSETX ipmdILoffsx; // the instr offset
6836 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6837 bool ipmdIsLabel; // Can this code be a branch label?
6840 // Record the instr offset mapping to the generated code
6842 IPmappingDsc* genIPmappingList;
6843 IPmappingDsc* genIPmappingLast;
6845 // Managed RetVal - A side hash table meant to record the mapping from a
6846 // GT_CALL node to its IL offset. This info is used to emit sequence points
6847 // that can be used by debugger to determine the native offset at which the
6848 // managed RetVal will be available.
6850 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6851 // favor of a side table for two reasons: 1) We need IL offset for only those
6852 // GT_CALL nodes (created during importation) that correspond to an IL call and
6853 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6854 // structure and IL offset is needed only when generating debuggable code. Therefore
6855 // it is desirable to avoid memory size penalty in retail scenarios.
6856 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6857 CallSiteILOffsetTable;
6858 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6860 unsigned genReturnLocal; // Local number for the return value when applicable.
6861 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6863 // The following properties are part of CodeGenContext. Getters are provided here for
6864 // convenience and backward compatibility, but the properties can only be set by invoking
6865 // the setter on CodeGenContext directly.
6867 __declspec(property(get = getEmitter)) emitter* genEmitter;
6868 emitter* getEmitter()
6870 return codeGen->getEmitter();
6873 const bool isFramePointerUsed()
6875 return codeGen->isFramePointerUsed();
6878 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6879 bool getInterruptible()
6881 return codeGen->genInterruptible;
6883 void setInterruptible(bool value)
6885 codeGen->setInterruptible(value);
6889 const bool genDoubleAlign()
6891 return codeGen->doDoubleAlign();
6893 DWORD getCanDoubleAlign();
6894 bool shouldDoubleAlign(unsigned refCntStk,
6896 unsigned refCntWtdReg,
6897 unsigned refCntStkParam,
6898 unsigned refCntWtdStkDbl);
6899 #endif // DOUBLE_ALIGN
6901 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6902 bool getFullPtrRegMap()
6904 return codeGen->genFullPtrRegMap;
6906 void setFullPtrRegMap(bool value)
6908 codeGen->setFullPtrRegMap(value);
6911 // Things that MAY belong either in CodeGen or CodeGenContext
6913 #if FEATURE_EH_FUNCLETS
6914 FuncInfoDsc* compFuncInfos;
6915 unsigned short compCurrFuncIdx;
6916 unsigned short compFuncInfoCount;
6918 unsigned short compFuncCount()
6920 assert(fgFuncletsCreated);
6921 return compFuncInfoCount;
6924 #else // !FEATURE_EH_FUNCLETS
6926 // This is a no-op when there are no funclets!
6927 void genUpdateCurrentFunclet(BasicBlock* block)
6932 FuncInfoDsc compFuncInfoRoot;
6934 static const unsigned compCurrFuncIdx = 0;
6936 unsigned short compFuncCount()
6941 #endif // !FEATURE_EH_FUNCLETS
6943 FuncInfoDsc* funCurrentFunc();
6944 void funSetCurrentFunc(unsigned funcIdx);
6945 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6946 unsigned int funGetFuncIdx(BasicBlock* block);
6950 VARSET_TP compCurLife; // current live variables
6951 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6953 template <bool ForCodeGen>
6954 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6956 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6958 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6961 template <bool ForCodeGen>
6962 void compUpdateLife(GenTreePtr tree);
6964 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6965 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6966 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6967 template <bool ForCodeGen>
6968 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6970 template <bool ForCodeGen>
6971 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6973 // Gets a register mask that represent the kill set for a helper call since
6974 // not all JIT Helper calls follow the standard ABI on the target architecture.
6975 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6977 // Gets a register mask that represent the kill set for a NoGC helper call.
6978 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6981 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6982 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6983 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6984 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6985 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6986 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6987 #endif // _TARGET_ARM_
6989 // 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
6991 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6993 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6994 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6995 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6996 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6997 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6998 // for the tracked var indices of the field vars, as in a live var set).
6999 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7001 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7003 if (m_promotedStructDeathVars == nullptr)
7005 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7007 return m_promotedStructDeathVars;
7011 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7012 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7016 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7017 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7020 #if !defined(__GNUC__)
7021 #pragma region Unwind information
7026 // Infrastructure functions: start/stop/reserve/emit.
7029 void unwindBegProlog();
7030 void unwindEndProlog();
7031 void unwindBegEpilog();
7032 void unwindEndEpilog();
7033 void unwindReserve();
7034 void unwindEmit(void* pHotCode, void* pColdCode);
7037 // Specific unwind information functions: called by code generation to indicate a particular
7038 // prolog or epilog unwindable instruction has been generated.
7041 void unwindPush(regNumber reg);
7042 void unwindAllocStack(unsigned size);
7043 void unwindSetFrameReg(regNumber reg, unsigned offset);
7044 void unwindSaveReg(regNumber reg, unsigned offset);
7046 #if defined(_TARGET_ARM_)
7047 void unwindPushMaskInt(regMaskTP mask);
7048 void unwindPushMaskFloat(regMaskTP mask);
7049 void unwindPopMaskInt(regMaskTP mask);
7050 void unwindPopMaskFloat(regMaskTP mask);
7051 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7052 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7053 // called via unwindPadding().
7054 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7055 // instruction and the current location.
7056 #endif // _TARGET_ARM_
7058 #if defined(_TARGET_ARM64_)
7060 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7061 // instruction and the current location.
7062 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7063 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7064 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7065 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7066 void unwindSaveNext(); // unwind code: save_next
7067 void unwindReturn(regNumber reg); // ret lr
7068 #endif // defined(_TARGET_ARM64_)
7071 // Private "helper" functions for the unwind implementation.
7075 #if FEATURE_EH_FUNCLETS
7076 void unwindGetFuncLocations(FuncInfoDsc* func,
7077 bool getHotSectionData,
7078 /* OUT */ emitLocation** ppStartLoc,
7079 /* OUT */ emitLocation** ppEndLoc);
7080 #endif // FEATURE_EH_FUNCLETS
7082 void unwindReserveFunc(FuncInfoDsc* func);
7083 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7085 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7087 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7088 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7090 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7092 #if defined(_TARGET_AMD64_)
7094 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7096 void unwindBegPrologWindows();
7097 void unwindPushWindows(regNumber reg);
7098 void unwindAllocStackWindows(unsigned size);
7099 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7100 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7102 #ifdef UNIX_AMD64_ABI
7103 void unwindBegPrologCFI();
7104 void unwindPushCFI(regNumber reg);
7105 void unwindAllocStackCFI(unsigned size);
7106 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7107 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7108 int mapRegNumToDwarfReg(regNumber reg);
7109 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7110 #endif // UNIX_AMD64_ABI
7111 #elif defined(_TARGET_ARM_)
7113 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7114 void unwindPushPopMaskFloat(regMaskTP mask);
7115 void unwindSplit(FuncInfoDsc* func);
7117 #endif // _TARGET_ARM_
7119 #if !defined(__GNUC__)
7120 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7124 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7125 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7129 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7130 XX that contains the distinguished, well-known SIMD type definitions). XX
7132 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7133 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7136 // Get highest available instruction set for floating point codegen
7137 InstructionSet getFloatingPointInstructionSet()
7139 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7142 return InstructionSet_AVX;
7147 return InstructionSet_SSE3_4;
7151 assert(canUseSSE2());
7152 return InstructionSet_SSE2;
7154 assert(!"getFPInstructionSet() is not implemented for target arch");
7156 return InstructionSet_NONE;
7160 // Get highest available instruction set for SIMD codegen
7161 InstructionSet getSIMDInstructionSet()
7163 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7164 return getFloatingPointInstructionSet();
7166 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7168 return InstructionSet_NONE;
7174 // Should we support SIMD intrinsics?
7177 // Have we identified any SIMD types?
7178 // This is currently used by struct promotion to avoid getting type information for a struct
7179 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7181 bool _usesSIMDTypes;
7182 bool usesSIMDTypes()
7184 return _usesSIMDTypes;
7186 void setUsesSIMDTypes(bool value)
7188 _usesSIMDTypes = value;
7191 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7192 // that require indexed access to the individual fields of the vector, which is not well supported
7193 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7194 unsigned lvaSIMDInitTempVarNum;
7197 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7198 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7199 CORINFO_CLASS_HANDLE SIMDIntHandle;
7200 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7201 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7202 CORINFO_CLASS_HANDLE SIMDShortHandle;
7203 CORINFO_CLASS_HANDLE SIMDByteHandle;
7204 CORINFO_CLASS_HANDLE SIMDLongHandle;
7205 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7206 CORINFO_CLASS_HANDLE SIMDULongHandle;
7207 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7208 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7209 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7210 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7212 // Get the handle for a SIMD type.
7213 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7215 if (simdBaseType == TYP_FLOAT)
7220 return SIMDVector2Handle;
7222 return SIMDVector3Handle;
7224 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7226 return SIMDVector4Handle;
7235 assert(simdType == getSIMDVectorType());
7236 switch (simdBaseType)
7239 return SIMDFloatHandle;
7241 return SIMDDoubleHandle;
7243 return SIMDIntHandle;
7245 return SIMDUShortHandle;
7247 return SIMDUShortHandle;
7249 return SIMDUByteHandle;
7251 return SIMDShortHandle;
7253 return SIMDByteHandle;
7255 return SIMDLongHandle;
7257 return SIMDUIntHandle;
7259 return SIMDULongHandle;
7261 assert(!"Didn't find a class handle for simdType");
7263 return NO_CLASS_HANDLE;
7267 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7268 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7269 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7271 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7272 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7273 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7274 bool isSIMDTypeLocal(GenTree* tree)
7276 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7279 // Returns true if the type of the tree is a byref of TYP_SIMD
7280 bool isAddrOfSIMDType(GenTree* tree)
7282 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7284 switch (tree->OperGet())
7287 return varTypeIsSIMD(tree->gtGetOp1());
7289 case GT_LCL_VAR_ADDR:
7290 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7293 return isSIMDTypeLocal(tree);
7300 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7302 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7303 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7304 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7307 // Returns base type of a TYP_SIMD local.
7308 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7309 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7311 if (isSIMDTypeLocal(tree))
7313 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7319 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7321 return info.compCompHnd->isInSIMDModule(clsHnd);
7324 bool isSIMDClass(typeInfo* pTypeInfo)
7326 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7329 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7330 // if it is not a SIMD type or is an unsupported base type.
7331 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7333 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7335 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7338 // Get SIMD Intrinsic info given the method handle.
7339 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7340 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7341 CORINFO_METHOD_HANDLE methodHnd,
7342 CORINFO_SIG_INFO* sig,
7345 var_types* baseType,
7346 unsigned* sizeBytes);
7348 // Pops and returns GenTree node from importers type stack.
7349 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7350 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7352 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7353 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7355 // Creates a GT_SIMD tree for Select operation
7356 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7358 unsigned simdVectorSize,
7363 // Creates a GT_SIMD tree for Min/Max operation
7364 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7365 CORINFO_CLASS_HANDLE typeHnd,
7367 unsigned simdVectorSize,
7371 // Transforms operands and returns the SIMD intrinsic to be applied on
7372 // transformed operands to obtain given relop result.
7373 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7374 CORINFO_CLASS_HANDLE typeHnd,
7375 unsigned simdVectorSize,
7376 var_types* baseType,
7380 // Creates a GT_SIMD tree for Abs intrinsic.
7381 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7383 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7384 // Transforms operands and returns the SIMD intrinsic to be applied on
7385 // transformed operands to obtain == comparison result.
7386 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7387 unsigned simdVectorSize,
7391 // Transforms operands and returns the SIMD intrinsic to be applied on
7392 // transformed operands to obtain > comparison result.
7393 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7394 unsigned simdVectorSize,
7398 // Transforms operands and returns the SIMD intrinsic to be applied on
7399 // transformed operands to obtain >= comparison result.
7400 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7401 unsigned simdVectorSize,
7405 // Transforms operands and returns the SIMD intrinsic to be applied on
7406 // transformed operands to obtain >= comparison result in case of int32
7407 // and small int base type vectors.
7408 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7409 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7410 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7412 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7413 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7414 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7415 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7416 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7418 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7419 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7420 GenTreePtr newobjThis,
7421 CORINFO_CLASS_HANDLE clsHnd,
7422 CORINFO_METHOD_HANDLE method,
7423 CORINFO_SIG_INFO* sig,
7426 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7428 // Whether SIMD vector occupies part of SIMD register.
7429 // SSE2: vector2f/3f are considered sub register SIMD types.
7430 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7431 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7433 unsigned sizeBytes = 0;
7434 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7435 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7438 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7440 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7443 // Get the type for the hardware SIMD vector.
7444 // This is the maximum SIMD type supported for this target.
7445 var_types getSIMDVectorType()
7447 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7454 assert(canUseSSE2());
7458 assert(!"getSIMDVectorType() unimplemented on target arch");
7463 // Get the size of the SIMD type in bytes
7464 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7466 unsigned sizeBytes = 0;
7467 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7471 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7472 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7474 // Get the the number of elements of basetype of SIMD vector given by its type handle
7475 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7477 // Get preferred alignment of SIMD type.
7478 int getSIMDTypeAlignment(var_types simdType);
7480 // Get the number of bytes in a SIMD Vector for the current compilation.
7481 unsigned getSIMDVectorRegisterByteLength()
7483 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7486 return YMM_REGSIZE_BYTES;
7490 assert(canUseSSE2());
7491 return XMM_REGSIZE_BYTES;
7494 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7499 // The minimum and maximum possible number of bytes in a SIMD vector.
7500 unsigned int maxSIMDStructBytes()
7502 return getSIMDVectorRegisterByteLength();
7504 unsigned int minSIMDStructBytes()
7506 return emitTypeSize(TYP_SIMD8);
7509 #ifdef FEATURE_AVX_SUPPORT
7510 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7511 static const unsigned maxPossibleSIMDStructBytes = 32;
7512 #else // !FEATURE_AVX_SUPPORT
7513 static const unsigned maxPossibleSIMDStructBytes = 16;
7514 #endif // !FEATURE_AVX_SUPPORT
7516 // Returns the codegen type for a given SIMD size.
7517 var_types getSIMDTypeForSize(unsigned size)
7519 var_types simdType = TYP_UNDEF;
7522 simdType = TYP_SIMD8;
7524 else if (size == 12)
7526 simdType = TYP_SIMD12;
7528 else if (size == 16)
7530 simdType = TYP_SIMD16;
7532 #ifdef FEATURE_AVX_SUPPORT
7533 else if (size == 32)
7535 simdType = TYP_SIMD32;
7537 #endif // FEATURE_AVX_SUPPORT
7540 noway_assert(!"Unexpected size for SIMD type");
7545 unsigned getSIMDInitTempVarNum()
7547 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7549 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7550 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7552 return lvaSIMDInitTempVarNum;
7555 #endif // FEATURE_SIMD
7558 //------------------------------------------------------------------------
7559 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7561 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7562 // candidate for enregistration.
7564 unsigned largestEnregisterableStructSize()
7567 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7568 if (vectorRegSize > TARGET_POINTER_SIZE)
7570 return vectorRegSize;
7573 #endif // FEATURE_SIMD
7575 return TARGET_POINTER_SIZE;
7580 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7581 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7582 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7584 // Is this var is of type simd struct?
7585 bool lclVarIsSIMDType(unsigned varNum)
7587 LclVarDsc* varDsc = lvaTable + varNum;
7588 return varDsc->lvIsSIMDType();
7591 // Is this Local node a SIMD local?
7592 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7594 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7597 // Returns true if the TYP_SIMD locals on stack are aligned at their
7598 // preferred byte boundary specified by getSIMDTypeAlignment().
7600 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7601 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7602 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7603 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7604 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7605 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7606 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7609 bool isSIMDTypeLocalAligned(unsigned varNum)
7611 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7612 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7615 int off = lvaFrameAddress(varNum, &ebpBased);
7616 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7617 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7618 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7621 #endif // FEATURE_SIMD
7626 // Whether SSE2 is available
7627 bool canUseSSE2() const
7629 #ifdef _TARGET_XARCH_
7630 return opts.compCanUseSSE2;
7636 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7637 bool CanUseSSE3_4() const
7639 #ifdef _TARGET_XARCH_
7640 return opts.compCanUseSSE3_4;
7646 bool canUseAVX() const
7648 #ifdef FEATURE_AVX_SUPPORT
7649 return opts.compCanUseAVX;
7656 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7657 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7661 XX Generic info about the compilation and the method being compiled. XX
7662 XX It is responsible for driving the other phases. XX
7663 XX It is also responsible for all the memory management. XX
7665 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7666 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7670 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7672 InlineResult* compInlineResult; // The result of importing the inlinee method.
7674 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7675 bool compJmpOpUsed; // Does the method do a JMP
7676 bool compLongUsed; // Does the method use TYP_LONG
7677 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7678 bool compTailCallUsed; // Does the method do a tailcall
7679 bool compLocallocUsed; // Does the method use localloc.
7680 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7681 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7682 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7684 // NOTE: These values are only reliable after
7685 // the importing is completely finished.
7687 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7688 // we can iterate over these efficiently.
7690 #if CPU_USES_BLOCK_MOVE
7691 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7695 // State information - which phases have completed?
7696 // These are kept together for easy discoverability
7698 bool bRangeAllowStress;
7699 bool compCodeGenDone;
7700 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7701 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7702 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7703 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7706 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7707 bool fgLocalVarLivenessChanged;
7709 bool compStackProbePrologDone;
7711 #ifndef LEGACY_BACKEND
7713 #endif // !LEGACY_BACKEND
7714 bool compRationalIRForm;
7716 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7718 bool compGeneratingProlog;
7719 bool compGeneratingEpilog;
7720 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7721 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7722 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7723 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7724 bool getNeedsGSSecurityCookie() const
7726 return compNeedsGSSecurityCookie;
7728 void setNeedsGSSecurityCookie()
7730 compNeedsGSSecurityCookie = true;
7733 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7734 // frame layout calculations, this is the level we are currently
7737 //---------------------------- JITing options -----------------------------
7750 JitFlags* jitFlags; // all flags passed from the EE
7751 unsigned compFlags; // method attributes
7753 codeOptimize compCodeOpt; // what type of code optimizations
7757 #ifdef _TARGET_XARCH_
7758 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7759 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7761 #ifdef FEATURE_AVX_SUPPORT
7762 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7763 #endif // FEATURE_AVX_SUPPORT
7764 #endif // _TARGET_XARCH_
7766 // optimize maximally and/or favor speed over size?
7768 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7769 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7770 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7771 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7772 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7774 // Maximun number of locals before turning off the inlining
7775 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7778 unsigned instrCount;
7779 unsigned lvRefCount;
7780 bool compMinOptsIsSet;
7782 bool compMinOptsIsUsed;
7784 inline bool MinOpts()
7786 assert(compMinOptsIsSet);
7787 compMinOptsIsUsed = true;
7790 inline bool IsMinOptsSet()
7792 return compMinOptsIsSet;
7795 inline bool MinOpts()
7799 inline bool IsMinOptsSet()
7801 return compMinOptsIsSet;
7804 inline void SetMinOpts(bool val)
7806 assert(!compMinOptsIsUsed);
7807 assert(!compMinOptsIsSet || (compMinOpts == val));
7809 compMinOptsIsSet = true;
7812 // true if the CLFLG_* for an optimization is set.
7813 inline bool OptEnabled(unsigned optFlag)
7815 return !!(compFlags & optFlag);
7818 #ifdef FEATURE_READYTORUN_COMPILER
7819 inline bool IsReadyToRun()
7821 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7824 inline bool IsReadyToRun()
7830 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7831 // PInvoke transitions inline (e.g. when targeting CoreRT).
7832 inline bool ShouldUsePInvokeHelpers()
7834 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7837 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7839 inline bool IsReversePInvoke()
7841 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7844 // true if we must generate code compatible with JIT32 quirks
7845 inline bool IsJit32Compat()
7847 #if defined(_TARGET_X86_)
7848 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7854 // true if we must generate code compatible with Jit64 quirks
7855 inline bool IsJit64Compat()
7857 #if defined(_TARGET_AMD64_)
7858 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7859 #elif !defined(FEATURE_CORECLR)
7866 bool compScopeInfo; // Generate the LocalVar info ?
7867 bool compDbgCode; // Generate debugger-friendly code?
7868 bool compDbgInfo; // Gather debugging info?
7871 #ifdef PROFILING_SUPPORTED
7872 bool compNoPInvokeInlineCB;
7874 static const bool compNoPInvokeInlineCB;
7878 bool compGcChecks; // Check arguments and return values to ensure they are sane
7879 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7880 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7884 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7885 // to be allocated on the stack.
7886 // It will be set to true in the following cases:
7887 // 1. When the method being compiled has a declarative security
7888 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7889 // This is also the case when we inject a prolog and epilog in the method.
7891 // 2. When the method being compiled has imperative security (i.e. the method
7892 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7894 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7896 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7897 // which gets reported as a GC root to stackwalker.
7898 // (See also ICodeManager::GetAddrOfSecurityObject.)
7903 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7904 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7908 #ifdef UNIX_AMD64_ABI
7909 // This flag is indicating if there is a need to align the frame.
7910 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7911 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7912 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7913 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7914 // there are calls and making sure the frame alignment logic is executed.
7915 bool compNeedToAlignFrame;
7916 #endif // UNIX_AMD64_ABI
7918 bool compProcedureSplitting; // Separate cold code from hot code
7920 bool genFPorder; // Preserve FP order (operations are non-commutative)
7921 bool genFPopt; // Can we do frame-pointer-omission optimization?
7922 bool altJit; // True if we are an altjit and are compiling this method
7925 bool optRepeat; // Repeat optimizer phases k times
7929 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7930 bool dspCode; // Display native code generated
7931 bool dspEHTable; // Display the EH table reported to the VM
7932 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7933 bool dspEmit; // Display emitter output
7934 bool dspLines; // Display source-code lines intermixed with native code output
7935 bool dmpHex; // Display raw bytes in hex of native code output
7936 bool varNames; // Display variables names in native code output
7937 bool disAsm; // Display native code as it is generated
7938 bool disAsmSpilled; // Display native code when any register spilling occurs
7939 bool disDiffable; // Makes the Disassembly code 'diff-able'
7940 bool disAsm2; // Display native code after it is generated using external disassembler
7941 bool dspOrder; // Display names of each of the methods that we ngen/jit
7942 bool dspUnwind; // Display the unwind info output
7943 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7944 bool compLongAddress; // Force using large pseudo instructions for long address
7945 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7946 bool dspGCtbls; // Display the GC tables
7950 bool doLateDisasm; // Run the late disassembler
7951 #endif // LATE_DISASM
7953 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7954 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7955 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7956 static const bool dspGCtbls = true;
7959 // We need stack probes to guarantee that we won't trigger a stack overflow
7960 // when calling unmanaged code until they get a chance to set up a frame, because
7961 // the EE will have no idea where it is.
7963 // We will only be doing this currently for hosted environments. Unfortunately
7964 // we need to take care of stubs, so potentially, we will have to do the probes
7965 // for any call. We have a plan for not needing for stubs though
7966 bool compNeedStackProbes;
7968 #ifdef PROFILING_SUPPORTED
7969 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7970 // This option helps make the JIT behave as if it is running under a profiler.
7971 bool compJitELTHookEnabled;
7972 #endif // PROFILING_SUPPORTED
7974 #if FEATURE_TAILCALL_OPT
7975 // Whether opportunistic or implicit tail call optimization is enabled.
7976 bool compTailCallOpt;
7977 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7978 bool compTailCallLoopOpt;
7982 static const bool compUseSoftFP = true;
7983 #else // !ARM_SOFTFP
7984 static const bool compUseSoftFP = false;
7987 GCPollType compGCPollType;
7991 static bool s_pAltJitExcludeAssembliesListInitialized;
7992 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7997 template <typename T>
8000 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8003 template <typename T>
8006 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8009 static int dspTreeID(GenTree* tree)
8011 return tree->gtTreeID;
8013 static void printTreeID(GenTree* tree)
8015 if (tree == nullptr)
8021 printf("[%06d]", dspTreeID(tree));
8028 #define STRESS_MODES \
8032 /* "Variations" stress areas which we try to mix up with each other. */ \
8033 /* These should not be exhaustively used as they might */ \
8034 /* hide/trivialize other areas */ \
8037 STRESS_MODE(DBL_ALN) \
8038 STRESS_MODE(LCL_FLDS) \
8039 STRESS_MODE(UNROLL_LOOPS) \
8040 STRESS_MODE(MAKE_CSE) \
8041 STRESS_MODE(LEGACY_INLINE) \
8042 STRESS_MODE(CLONE_EXPR) \
8043 STRESS_MODE(USE_FCOMI) \
8044 STRESS_MODE(USE_CMOV) \
8046 STRESS_MODE(BB_PROFILE) \
8047 STRESS_MODE(OPT_BOOLS_GC) \
8048 STRESS_MODE(REMORPH_TREES) \
8049 STRESS_MODE(64RSLT_MUL) \
8050 STRESS_MODE(DO_WHILE_LOOPS) \
8051 STRESS_MODE(MIN_OPTS) \
8052 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8053 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8054 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8055 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8056 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8057 STRESS_MODE(NULL_OBJECT_CHECK) \
8058 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8059 STRESS_MODE(RANDOM_INLINE) \
8060 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8061 STRESS_MODE(GENERIC_VARN) \
8063 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8065 STRESS_MODE(COUNT_VARN) \
8067 /* "Check" stress areas that can be exhaustively used if we */ \
8068 /* dont care about performance at all */ \
8070 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8071 STRESS_MODE(CHK_FLOW_UPDATE) \
8072 STRESS_MODE(EMITTER) \
8073 STRESS_MODE(CHK_REIMPORT) \
8074 STRESS_MODE(FLATFP) \
8075 STRESS_MODE(GENERIC_CHECK) \
8080 #define STRESS_MODE(mode) STRESS_##mode,
8087 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8088 BYTE compActiveStressModes[STRESS_COUNT];
8091 #define MAX_STRESS_WEIGHT 100
8093 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8097 bool compInlineStress()
8099 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8102 bool compRandomInlineStress()
8104 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8109 bool compTailCallStress()
8112 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8118 codeOptimize compCodeOpt()
8121 // Switching between size & speed has measurable throughput impact
8122 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8123 // DEBUG, but should generate identical code between CHK & RET builds,
8124 // so that's not acceptable.
8125 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8126 // Investigate the cause of the throughput regression.
8128 return opts.compCodeOpt;
8130 return BLENDED_CODE;
8134 //--------------------- Info about the procedure --------------------------
8138 COMP_HANDLE compCompHnd;
8139 CORINFO_MODULE_HANDLE compScopeHnd;
8140 CORINFO_CLASS_HANDLE compClassHnd;
8141 CORINFO_METHOD_HANDLE compMethodHnd;
8142 CORINFO_METHOD_INFO* compMethodInfo;
8144 BOOL hasCircularClassConstraints;
8145 BOOL hasCircularMethodConstraints;
8147 #if defined(DEBUG) || defined(LATE_DISASM)
8148 const char* compMethodName;
8149 const char* compClassName;
8150 const char* compFullName;
8151 #endif // defined(DEBUG) || defined(LATE_DISASM)
8153 #if defined(DEBUG) || defined(INLINE_DATA)
8154 // Method hash is logcally const, but computed
8156 mutable unsigned compMethodHashPrivate;
8157 unsigned compMethodHash() const;
8158 #endif // defined(DEBUG) || defined(INLINE_DATA)
8160 #ifdef PSEUDORANDOM_NOP_INSERTION
8161 // things for pseudorandom nop insertion
8162 unsigned compChecksum;
8166 // The following holds the FLG_xxxx flags for the method we're compiling.
8169 // The following holds the class attributes for the method we're compiling.
8170 unsigned compClassAttr;
8172 const BYTE* compCode;
8173 IL_OFFSET compILCodeSize; // The IL code size
8174 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8175 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8176 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8177 // (2) the code is hot/cold split, and we issued less code than we expected
8178 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8180 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8181 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8182 bool compIsContextful : 1; // contextful method
8183 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8184 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8185 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8186 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8187 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8189 var_types compRetType; // Return type of the method as declared in IL
8190 var_types compRetNativeType; // Normalized return type as per target arch ABI
8191 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8192 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8193 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8194 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8195 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8196 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8197 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8198 unsigned compMaxStack;
8199 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8200 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8202 unsigned compCallUnmanaged; // count of unmanaged calls
8203 unsigned compLvFrameListRoot; // lclNum for the Frame root
8204 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8205 // You should generally use compHndBBtabCount instead: it is the
8206 // current number of EH clauses (after additions like synchronized
8207 // methods and funclets, and removals like unreachable code deletion).
8209 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8210 // and the VM expects that, or the JIT is a "self-host" compiler
8211 // (e.g., x86 hosted targeting x86) and the VM expects that.
8213 /* The following holds IL scope information about local variables.
8216 unsigned compVarScopesCount;
8217 VarScopeDsc* compVarScopes;
8219 /* The following holds information about instr offsets for
8220 * which we need to report IP-mappings
8223 IL_OFFSET* compStmtOffsets; // sorted
8224 unsigned compStmtOffsetsCount;
8225 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8227 #define CPU_X86 0x0100 // The generic X86 CPU
8228 #define CPU_X86_PENTIUM_4 0x0110
8230 #define CPU_X64 0x0200 // The generic x64 CPU
8231 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8232 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8234 #define CPU_ARM 0x0300 // The generic ARM CPU
8236 unsigned genCPU; // What CPU are we running on
8239 // Returns true if the method being compiled returns a non-void and non-struct value.
8240 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8241 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8242 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8243 // Methods returning such structs are considered to return non-struct return value and
8244 // this method returns true in that case.
8245 bool compMethodReturnsNativeScalarType()
8247 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8250 // Returns true if the method being compiled returns RetBuf addr as its return value
8251 bool compMethodReturnsRetBufAddr()
8253 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8254 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8256 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8257 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8258 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8259 // methods with hidden RetBufArg.
8261 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8262 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8263 // returning the address of RetBuf.
8265 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8266 // to be returned in RAX.
8267 CLANG_FORMAT_COMMENT_ANCHOR;
8269 #ifdef _TARGET_AMD64_
8270 return (info.compRetBuffArg != BAD_VAR_NUM);
8271 #else // !_TARGET_AMD64_
8272 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8273 #endif // !_TARGET_AMD64_
8276 // Returns true if the method returns a value in more than one return register
8277 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8278 // TODO-ARM64: Does this apply for ARM64 too?
8279 bool compMethodReturnsMultiRegRetType()
8281 #if FEATURE_MULTIREG_RET
8282 #if defined(_TARGET_X86_)
8283 // On x86 only 64-bit longs are returned in multiple registers
8284 return varTypeIsLong(info.compRetNativeType);
8285 #else // targets: X64-UNIX, ARM64 or ARM32
8286 // On all other targets that support multireg return values:
8287 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8288 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8289 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8290 #endif // TARGET_XXX
8292 #else // not FEATURE_MULTIREG_RET
8294 // For this architecture there are no multireg returns
8297 #endif // FEATURE_MULTIREG_RET
8300 #if FEATURE_MULTIREG_ARGS
8301 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8302 // return the gcPtr layout for the pointers sized fields
8303 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8304 #endif // FEATURE_MULTIREG_ARGS
8306 // Returns true if the method being compiled returns a value
8307 bool compMethodHasRetVal()
8309 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8310 compMethodReturnsMultiRegRetType();
8315 void compDispLocalVars();
8319 //-------------------------- Global Compiler Data ------------------------------------
8322 static unsigned s_compMethodsCount; // to produce unique label names
8323 unsigned compGenTreeID;
8326 BasicBlock* compCurBB; // the current basic block in process
8327 GenTreePtr compCurStmt; // the current statement in process
8329 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8332 // The following is used to create the 'method JIT info' block.
8333 size_t compInfoBlkSize;
8334 BYTE* compInfoBlkAddr;
8336 EHblkDsc* compHndBBtab; // array of EH data
8337 unsigned compHndBBtabCount; // element count of used elements in EH data array
8338 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8340 #if defined(_TARGET_X86_)
8342 //-------------------------------------------------------------------------
8343 // Tracking of region covered by the monitor in synchronized methods
8344 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8345 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8347 #endif // !_TARGET_X86_
8349 Phases previousCompletedPhase; // the most recently completed phase
8351 //-------------------------------------------------------------------------
8352 // The following keeps track of how many bytes of local frame space we've
8353 // grabbed so far in the current function, and how many argument bytes we
8354 // need to pop when we return.
8357 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8359 // Count of callee-saved regs we pushed in the prolog.
8360 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8361 // In case of Amd64 this doesn't include float regs saved on stack.
8362 unsigned compCalleeRegsPushed;
8364 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8365 // Mask of callee saved float regs on stack.
8366 regMaskTP compCalleeFPRegsSavedMask;
8368 #ifdef _TARGET_AMD64_
8369 // Quirk for VS debug-launch scenario to work:
8370 // Bytes of padding between save-reg area and locals.
8371 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8372 unsigned compVSQuirkStackPaddingNeeded;
8373 bool compQuirkForPPPflag;
8376 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8378 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8379 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8380 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8382 //-------------------------------------------------------------------------
8384 static void compStartup(); // One-time initialization
8385 static void compShutdown(); // One-time finalization
8387 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8390 static void compDisplayStaticSizes(FILE* fout);
8392 //------------ Some utility functions --------------
8394 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8395 void** ppIndirection); /* OUT */
8397 // Several JIT/EE interface functions return a CorInfoType, and also return a
8398 // class handle as an out parameter if the type is a value class. Returns the
8399 // size of the type these describe.
8400 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8403 // Components used by the compiler may write unit test suites, and
8404 // have them run within this method. They will be run only once per process, and only
8405 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8406 // These should fail by asserting.
8407 void compDoComponentUnitTestsOnce();
8410 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8411 CORINFO_MODULE_HANDLE classPtr,
8412 COMP_HANDLE compHnd,
8413 CORINFO_METHOD_INFO* methodInfo,
8414 void** methodCodePtr,
8415 ULONG* methodCodeSize,
8416 JitFlags* compileFlags);
8417 void compCompileFinish();
8418 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8419 COMP_HANDLE compHnd,
8420 CORINFO_METHOD_INFO* methodInfo,
8421 void** methodCodePtr,
8422 ULONG* methodCodeSize,
8423 JitFlags* compileFlags,
8424 CorInfoInstantiationVerification instVerInfo);
8426 ArenaAllocator* compGetAllocator();
8428 #if MEASURE_MEM_ALLOC
8430 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8434 unsigned allocCnt; // # of allocs
8435 UINT64 allocSz; // total size of those alloc.
8436 UINT64 allocSzMax; // Maximum single allocation.
8437 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8438 UINT64 nraTotalSizeAlloc;
8439 UINT64 nraTotalSizeUsed;
8441 static const char* s_CompMemKindNames[]; // Names of the kinds.
8443 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8445 for (int i = 0; i < CMK_Count; i++)
8447 allocSzByKind[i] = 0;
8450 MemStats(const MemStats& ms)
8451 : allocCnt(ms.allocCnt)
8452 , allocSz(ms.allocSz)
8453 , allocSzMax(ms.allocSzMax)
8454 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8455 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8457 for (int i = 0; i < CMK_Count; i++)
8459 allocSzByKind[i] = ms.allocSzByKind[i];
8463 // Until we have ubiquitous constructors.
8466 this->MemStats::MemStats();
8469 void AddAlloc(size_t sz, CompMemKind cmk)
8473 if (sz > allocSzMax)
8477 allocSzByKind[cmk] += sz;
8480 void Print(FILE* f); // Print these stats to f.
8481 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8483 MemStats genMemStats;
8485 struct AggregateMemStats : public MemStats
8489 AggregateMemStats() : MemStats(), nMethods(0)
8493 void Add(const MemStats& ms)
8496 allocCnt += ms.allocCnt;
8497 allocSz += ms.allocSz;
8498 allocSzMax = max(allocSzMax, ms.allocSzMax);
8499 for (int i = 0; i < CMK_Count; i++)
8501 allocSzByKind[i] += ms.allocSzByKind[i];
8503 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8504 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8507 void Print(FILE* f); // Print these stats to jitstdout.
8510 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8511 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8512 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8514 #endif // MEASURE_MEM_ALLOC
8516 #if LOOP_HOIST_STATS
8517 unsigned m_loopsConsidered;
8518 bool m_curLoopHasHoistedExpression;
8519 unsigned m_loopsWithHoistedExpressions;
8520 unsigned m_totalHoistedExpressions;
8522 void AddLoopHoistStats();
8523 void PrintPerMethodLoopHoistStats();
8525 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8526 static unsigned s_loopsConsidered;
8527 static unsigned s_loopsWithHoistedExpressions;
8528 static unsigned s_totalHoistedExpressions;
8530 static void PrintAggregateLoopHoistStats(FILE* f);
8531 #endif // LOOP_HOIST_STATS
8533 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8534 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8535 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8536 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8537 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8538 void compFreeMem(void*);
8540 bool compIsForImportOnly();
8541 bool compIsForInlining();
8542 bool compDonotInline();
8545 const char* compLocalVarName(unsigned varNum, unsigned offs);
8546 VarName compVarName(regNumber reg, bool isFloatReg = false);
8547 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8548 const char* compRegPairName(regPairNo regPair);
8549 const char* compRegNameForSize(regNumber reg, size_t size);
8550 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8551 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8552 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8555 //-------------------------------------------------------------------------
8557 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8559 struct VarScopeMapInfo
8561 VarScopeListNode* head;
8562 VarScopeListNode* tail;
8563 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8565 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8572 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8573 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8575 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8576 VarNumToScopeDscMap;
8578 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8579 VarNumToScopeDscMap* compVarScopeMap;
8581 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8583 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8585 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8587 void compInitVarScopeMap();
8589 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8590 // enter scope, sorted by instr offset
8591 unsigned compNextEnterScope;
8593 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8594 // go out of scope, sorted by instr offset
8595 unsigned compNextExitScope;
8597 void compInitScopeLists();
8599 void compResetScopeLists();
8601 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8603 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8605 void compProcessScopesUntil(unsigned offset,
8607 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8608 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8611 void compDispScopeLists();
8614 bool compIsProfilerHookNeeded();
8616 //-------------------------------------------------------------------------
8617 /* Statistical Data Gathering */
8619 void compJitStats(); // call this function and enable
8620 // various ifdef's below for statistical data
8623 void compCallArgStats();
8624 static void compDispCallArgStats(FILE* fout);
8627 //-------------------------------------------------------------------------
8634 ArenaAllocator* compAllocator;
8637 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8638 // suitable for use by utilcode collection types.
8639 IAllocator* compAsIAllocator;
8641 #if MEASURE_MEM_ALLOC
8642 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8643 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8644 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8646 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8648 #endif // MEASURE_MEM_ALLOC
8650 void compFunctionTraceStart();
8651 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8654 size_t compMaxUncheckedOffsetForNullObject;
8656 void compInitOptions(JitFlags* compileFlags);
8658 void compSetProcessor();
8659 void compInitDebuggingInfo();
8660 void compSetOptimizationLevel();
8661 #ifdef _TARGET_ARMARCH_
8662 bool compRsvdRegCheck(FrameLayoutState curState);
8664 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8666 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8667 void ResetOptAnnotations();
8669 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8670 void RecomputeLoopInfo();
8672 #ifdef PROFILING_SUPPORTED
8673 // Data required for generating profiler Enter/Leave/TailCall hooks
8675 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8676 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8677 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8680 #ifdef _TARGET_AMD64_
8681 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8684 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8685 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8687 IAllocator* getAllocator()
8689 return compAsIAllocator;
8692 #if MEASURE_MEM_ALLOC
8693 IAllocator* getAllocatorBitset()
8695 return compAsIAllocatorBitset;
8697 IAllocator* getAllocatorGC()
8699 return compAsIAllocatorGC;
8701 IAllocator* getAllocatorLoopHoist()
8703 return compAsIAllocatorLoopHoist;
8705 #else // !MEASURE_MEM_ALLOC
8706 IAllocator* getAllocatorBitset()
8708 return compAsIAllocator;
8710 IAllocator* getAllocatorGC()
8712 return compAsIAllocator;
8714 IAllocator* getAllocatorLoopHoist()
8716 return compAsIAllocator;
8718 #endif // !MEASURE_MEM_ALLOC
8721 IAllocator* getAllocatorDebugOnly()
8723 #if MEASURE_MEM_ALLOC
8724 return compAsIAllocatorDebugOnly;
8725 #else // !MEASURE_MEM_ALLOC
8726 return compAsIAllocator;
8727 #endif // !MEASURE_MEM_ALLOC
8732 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8737 XX Checks for type compatibility and merges types XX
8739 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8740 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8744 // Set to TRUE if verification cannot be skipped for this method
8745 // If we detect unverifiable code, we will lazily check
8746 // canSkipMethodVerification() to see if verification is REALLY needed.
8747 BOOL tiVerificationNeeded;
8749 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8750 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8751 BOOL tiIsVerifiableCode;
8753 // Set to TRUE if runtime callout is needed for this method
8754 BOOL tiRuntimeCalloutNeeded;
8756 // Set to TRUE if security prolog/epilog callout is needed for this method
8757 // Note: This flag is different than compNeedSecurityCheck.
8758 // compNeedSecurityCheck means whether or not a security object needs
8759 // to be allocated on the stack, which is currently true for EnC as well.
8760 // tiSecurityCalloutNeeded means whether or not security callouts need
8761 // to be inserted in the jitted code.
8762 BOOL tiSecurityCalloutNeeded;
8764 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8765 // This support is necessary to suport attributes that are not described in
8766 // for example, signatures. For example, the permanent home byref (byref that
8767 // points to the gc heap), isn't a property of method signatures, therefore,
8768 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8769 // but when deciding if we need to reimport a block, we need to take these
8771 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8773 // Returns TRUE if child is equal to or a subtype of parent.
8774 // normalisedForStack indicates that both types are normalised for the stack
8775 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8777 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8778 // *pDest is modified to represent the merged type. Sets "*changed" to true
8779 // if this changes "*pDest".
8780 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8782 // Set pDest from the primitive value type.
8783 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8785 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8788 // <BUGNUM> VSW 471305
8789 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8790 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8791 // We use a "short" as we need to push/pop this scope.
8793 short compRegSetCheckLevel;
8797 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8800 XX IL verification stuff XX
8803 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8804 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8808 // The following is used to track liveness of local variables, initialization
8809 // of valueclass constructors, and type safe use of IL instructions.
8811 // dynamic state info needed for verification
8812 EntryState verCurrentState;
8814 // this ptr of object type .ctors are considered intited only after
8815 // the base class ctor is called, or an alternate ctor is called.
8816 // An uninited this ptr can be used to access fields, but cannot
8817 // be used to call a member function.
8818 BOOL verTrackObjCtorInitState;
8820 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8822 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8823 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8824 void verInitCurrentState();
8825 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8827 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8828 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8829 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8831 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8832 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8833 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8834 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8835 typeInfo verMakeTypeInfo(CorInfoType ciType,
8836 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8837 BOOL verIsSDArray(typeInfo ti);
8838 typeInfo verGetArrayElemType(typeInfo ti);
8840 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8841 BOOL verNeedsVerification();
8842 BOOL verIsByRefLike(const typeInfo& ti);
8843 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8845 // generic type variables range over types that satisfy IsBoxable
8846 BOOL verIsBoxable(const typeInfo& ti);
8848 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8849 DEBUGARG(unsigned line));
8850 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8851 DEBUGARG(unsigned line));
8852 bool verCheckTailCallConstraint(OPCODE opcode,
8853 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8854 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8855 // on a type parameter?
8856 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8857 // return false to the caller.
8858 // If false, it will throw.
8860 bool verIsBoxedValueType(typeInfo ti);
8862 void verVerifyCall(OPCODE opcode,
8863 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8864 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8866 bool readonlyCall, // is this a "readonly." call?
8867 const BYTE* delegateCreateStart,
8868 const BYTE* codeAddr,
8869 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8871 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8873 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8874 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8875 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8876 const CORINFO_FIELD_INFO& fieldInfo,
8877 const typeInfo* tiThis,
8879 BOOL allowPlainStructAsThis = FALSE);
8880 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8881 void verVerifyThisPtrInitialised();
8882 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8884 // Register allocator
8885 void raInitStackFP();
8886 void raEnregisterVarsPrePassStackFP();
8887 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8888 void raEnregisterVarsPostPassStackFP();
8889 void raGenerateFPRefCounts();
8890 void raEnregisterVarsStackFP();
8891 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8893 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8894 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8896 // returns true if enregistering v1 would save more mem accesses than v2
8897 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8900 void raDumpHeightsStackFP();
8901 void raDumpVariableRegIntfFloat();
8904 #if FEATURE_STACK_FP_X87
8906 // Currently, we use FP transition blocks in only 2 situations:
8908 // -conditional jump on longs where FP stack differs with target: it's not strictly
8909 // necessary, but its low frequency and the code would get complicated if we try to
8910 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8911 // minimize the way we generate the jump code.
8912 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8913 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8915 // However, transition blocks have 2 problems
8917 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8918 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8919 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8920 // in the right place without preordering them), this causes us to have to generate the transition
8921 // blocks in the cold area if we want procedure splitting.
8924 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8925 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8926 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8927 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8928 // a big change in the exception.
8930 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8931 // optimizations. For these 2 cases:
8933 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8934 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8935 // a switch statement.
8937 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8938 // current procedure splitting and exception code have.
8939 bool compMayHaveTransitionBlocks;
8941 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8943 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8945 unsigned raCntStkStackFP;
8946 unsigned raCntWtdStkDblStackFP;
8947 unsigned raCntStkParamDblStackFP;
8949 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8950 // TODO: Do we want to put this in LclVarDsc?
8951 unsigned raPayloadStackFP[lclMAX_TRACKED];
8952 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8954 // Useful for debugging
8955 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8957 #endif // FEATURE_STACK_FP_X87
8960 // One line log function. Default level is 0. Increasing it gives you
8961 // more log information
8963 // levels are currently unused: #define JITDUMP(level,...) ();
8964 void JitLogEE(unsigned level, const char* fmt, ...);
8966 bool compDebugBreak;
8968 bool compJitHaltMethod();
8973 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8974 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8976 XX GS Security checks for unsafe buffers XX
8978 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8979 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8982 struct ShadowParamVarInfo
8984 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8985 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8987 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8989 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8990 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8991 // slots and update all trees to refer to shadow slots is done immediately after
8992 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8993 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8994 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8995 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8996 // creating a shadow slot even though this routine returns true.
8998 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8999 // required. There are two cases under which a reg arg could potentially be used from its
9001 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9002 // b) LSRA spills it
9004 // Possible solution to address case (a)
9005 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9006 // in this routine. Note that live out of exception handler is something we may not be
9007 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9008 // Therefore, for methods with exception handling and need GS cookie check we might have
9009 // to take conservative approach.
9011 // Possible solution to address case (b)
9012 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9013 // create a new spill temp if the method needs GS cookie check.
9014 return varDsc->lvIsParam;
9015 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9016 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9023 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9028 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9029 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9030 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9032 void gsGSChecksInitCookie(); // Grabs cookie variable
9033 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9034 bool gsFindVulnerableParams(); // Shadow param analysis code
9035 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9037 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9038 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9040 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9041 // This can be overwritten by setting complus_JITInlineSize env variable.
9043 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9046 #ifdef FEATURE_JIT_METHOD_PERF
9047 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9048 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9050 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9051 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9053 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9055 #if MEASURE_CLRAPI_CALLS
9056 // Thin wrappers that call into JitTimer (if present).
9057 inline void CLRApiCallEnter(unsigned apix);
9058 inline void CLRApiCallLeave(unsigned apix);
9061 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9062 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9067 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9068 // These variables are associated with maintaining SQM data about compile time.
9069 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9070 // in the current compilation.
9071 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9072 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9073 // the inlining phase in the current compilation.
9074 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9076 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9077 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9078 // type-loading and class initialization).
9079 void RecordStateAtEndOfInlining();
9080 // Assumes being called at the end of compilation. Update the SQM state.
9081 void RecordStateAtEndOfCompilation();
9083 #ifdef FEATURE_CLRSQM
9084 // Does anything SQM related necessary at process shutdown time.
9085 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9086 #endif // FEATURE_CLRSQM
9089 #if FUNC_INFO_LOGGING
9090 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9091 // filename to write it to.
9092 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9093 #endif // FUNC_INFO_LOGGING
9095 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9097 // Is the compilation in a full trust context?
9098 bool compIsFullTrust();
9101 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9102 #endif // MEASURE_NOWAY
9104 #ifndef FEATURE_TRACELOGGING
9105 // Should we actually fire the noway assert body and the exception handler?
9106 bool compShouldThrowOnNoway();
9107 #else // FEATURE_TRACELOGGING
9108 // Should we actually fire the noway assert body and the exception handler?
9109 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9111 // Telemetry instance to use per method compilation.
9112 JitTelemetry compJitTelemetry;
9114 // Get common parameters that have to be logged with most telemetry data.
9115 void compGetTelemetryDefaults(const char** assemblyName,
9116 const char** scopeName,
9117 const char** methodName,
9118 unsigned* methodHash);
9119 #endif // !FEATURE_TRACELOGGING
9123 NodeToTestDataMap* m_nodeTestData;
9125 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9126 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9127 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9128 // Current kept in this.
9130 NodeToTestDataMap* GetNodeTestData()
9132 Compiler* compRoot = impInlineRoot();
9133 if (compRoot->m_nodeTestData == nullptr)
9135 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9137 return compRoot->m_nodeTestData;
9140 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9142 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9143 // currently occur in the AST graph.
9144 NodeToIntMap* FindReachableNodesInNodeTestData();
9146 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9147 // test data, associate that data with "to".
9148 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9150 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9151 // have annotations, attach similar annotations to the corresponding nodes in "to".
9152 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9154 // These are the methods that test that the various conditions implied by the
9155 // test attributes are satisfied.
9156 void JitTestCheckSSA(); // SSA builder tests.
9157 void JitTestCheckVN(); // Value numbering tests.
9160 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9162 FieldSeqStore* m_fieldSeqStore;
9164 FieldSeqStore* GetFieldSeqStore()
9166 Compiler* compRoot = impInlineRoot();
9167 if (compRoot->m_fieldSeqStore == nullptr)
9169 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9170 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9171 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9173 return compRoot->m_fieldSeqStore;
9176 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9178 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9179 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9180 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9181 // attach the field sequence directly to the address node.
9182 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9184 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9186 // Don't need to worry about inlining here
9187 if (m_zeroOffsetFieldMap == nullptr)
9189 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9191 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9192 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9194 return m_zeroOffsetFieldMap;
9197 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9198 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9199 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9200 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9201 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9202 // record the the field sequence using the ZeroOffsetFieldMap described above.
9204 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9205 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9206 // CoreRT. Such case is handled same as the default case.
9207 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9209 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9211 NodeToArrayInfoMap* m_arrayInfoMap;
9213 NodeToArrayInfoMap* GetArrayInfoMap()
9215 Compiler* compRoot = impInlineRoot();
9216 if (compRoot->m_arrayInfoMap == nullptr)
9218 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9219 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9220 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9222 return compRoot->m_arrayInfoMap;
9225 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9227 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9228 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9229 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9230 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9232 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9234 // Use the same map for GCHeap and ByrefExposed when their states match.
9235 memoryKind = ByrefExposed;
9238 assert(memoryKind < MemoryKindCount);
9239 Compiler* compRoot = impInlineRoot();
9240 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9242 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9243 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9244 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9246 return compRoot->m_memorySsaMap[memoryKind];
9249 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9250 CORINFO_CLASS_HANDLE m_refAnyClass;
9251 CORINFO_FIELD_HANDLE GetRefanyDataField()
9253 if (m_refAnyClass == nullptr)
9255 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9257 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9259 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9261 if (m_refAnyClass == nullptr)
9263 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9265 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9269 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9271 #if ALLVARSET_COUNTOPS
9272 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9275 static HelperCallProperties s_helperCallProperties;
9277 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9278 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9279 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9281 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9284 unsigned __int8* offset0,
9285 unsigned __int8* offset1);
9286 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9287 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9289 void fgMorphMultiregStructArgs(GenTreeCall* call);
9290 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9292 }; // end of class Compiler
9294 // Inline methods of CompAllocator.
9295 void* CompAllocator::Alloc(size_t sz)
9297 #if MEASURE_MEM_ALLOC
9298 return m_comp->compGetMem(sz, m_cmk);
9300 return m_comp->compGetMem(sz);
9304 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9306 #if MEASURE_MEM_ALLOC
9307 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9309 return m_comp->compGetMemArray(elems, elemSize);
9313 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9314 inline LclVarDsc::LclVarDsc(Compiler* comp)
9315 : // Initialize the ArgRegs to REG_STK.
9316 // The morph will do the right thing to change
9317 // to the right register if passed in register.
9320 #if FEATURE_MULTIREG_ARGS
9321 _lvOtherArgReg(REG_STK)
9323 #endif // FEATURE_MULTIREG_ARGS
9325 lvRefBlks(BlockSetOps::UninitVal())
9327 #endif // ASSERTION_PROP
9328 lvPerSsaData(comp->getAllocator())
9333 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9334 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9336 XX Miscellaneous Compiler stuff XX
9338 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9339 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9342 // Values used to mark the types a stack slot is used for
9344 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9345 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9346 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9347 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9348 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9349 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9350 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9351 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9353 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9355 /*****************************************************************************
9357 * Variables to keep track of total code amounts.
9362 extern size_t grossVMsize;
9363 extern size_t grossNCsize;
9364 extern size_t totalNCsize;
9366 extern unsigned genMethodICnt;
9367 extern unsigned genMethodNCnt;
9368 extern size_t gcHeaderISize;
9369 extern size_t gcPtrMapISize;
9370 extern size_t gcHeaderNSize;
9371 extern size_t gcPtrMapNSize;
9373 #endif // DISPLAY_SIZES
9375 /*****************************************************************************
9377 * Variables to keep track of basic block counts (more data on 1 BB methods)
9380 #if COUNT_BASIC_BLOCKS
9381 extern Histogram bbCntTable;
9382 extern Histogram bbOneBBSizeTable;
9385 /*****************************************************************************
9387 * Used by optFindNaturalLoops to gather statistical information such as
9388 * - total number of natural loops
9389 * - number of loops with 1, 2, ... exit conditions
9390 * - number of loops that have an iterator (for like)
9391 * - number of loops that have a constant iterator
9396 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9397 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9398 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9399 extern unsigned totalLoopCount; // counts the total number of natural loops
9400 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9401 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9402 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9403 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9405 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9406 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9407 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9408 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9409 extern Histogram loopCountTable; // Histogram of loop counts
9410 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9412 #endif // COUNT_LOOPS
9414 /*****************************************************************************
9415 * variables to keep track of how many iterations we go in a dataflow pass
9420 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9421 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9423 #endif // DATAFLOW_ITER
9425 #if MEASURE_BLOCK_SIZE
9426 extern size_t genFlowNodeSize;
9427 extern size_t genFlowNodeCnt;
9428 #endif // MEASURE_BLOCK_SIZE
9430 #if MEASURE_NODE_SIZE
9431 struct NodeSizeStats
9436 genTreeNodeSize = 0;
9437 genTreeNodeActualSize = 0;
9440 size_t genTreeNodeCnt;
9441 size_t genTreeNodeSize; // The size we allocate
9442 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9443 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9444 // a smaller node to a larger one. TODO-Cleanup: add stats on
9445 // SetOper()/ChangeOper() usage to quanitfy this.
9447 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9448 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9449 extern Histogram genTreeNcntHist;
9450 extern Histogram genTreeNsizHist;
9451 #endif // MEASURE_NODE_SIZE
9453 /*****************************************************************************
9454 * Count fatal errors (including noway_asserts).
9458 extern unsigned fatal_badCode;
9459 extern unsigned fatal_noWay;
9460 extern unsigned fatal_NOMEM;
9461 extern unsigned fatal_noWayAssertBody;
9463 extern unsigned fatal_noWayAssertBodyArgs;
9465 extern unsigned fatal_NYI;
9466 #endif // MEASURE_FATAL
9468 /*****************************************************************************
9472 #ifdef _TARGET_XARCH_
9474 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9475 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9476 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9478 const instruction INS_AND = INS_and;
9479 const instruction INS_OR = INS_or;
9480 const instruction INS_XOR = INS_xor;
9481 const instruction INS_NEG = INS_neg;
9482 const instruction INS_TEST = INS_test;
9483 const instruction INS_MUL = INS_imul;
9484 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9485 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9486 const instruction INS_BREAKPOINT = INS_int3;
9487 const instruction INS_ADDC = INS_adc;
9488 const instruction INS_SUBC = INS_sbb;
9489 const instruction INS_NOT = INS_not;
9495 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9496 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9497 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9499 const instruction INS_AND = INS_and;
9500 const instruction INS_OR = INS_orr;
9501 const instruction INS_XOR = INS_eor;
9502 const instruction INS_NEG = INS_rsb;
9503 const instruction INS_TEST = INS_tst;
9504 const instruction INS_MUL = INS_mul;
9505 const instruction INS_MULADD = INS_mla;
9506 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9507 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9508 const instruction INS_BREAKPOINT = INS_bkpt;
9509 const instruction INS_ADDC = INS_adc;
9510 const instruction INS_SUBC = INS_sbc;
9511 const instruction INS_NOT = INS_mvn;
9513 const instruction INS_ABS = INS_vabs;
9514 const instruction INS_ROUND = INS_invalid;
9515 const instruction INS_SQRT = INS_vsqrt;
9519 #ifdef _TARGET_ARM64_
9521 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9522 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9523 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9525 const instruction INS_AND = INS_and;
9526 const instruction INS_OR = INS_orr;
9527 const instruction INS_XOR = INS_eor;
9528 const instruction INS_NEG = INS_neg;
9529 const instruction INS_TEST = INS_tst;
9530 const instruction INS_MUL = INS_mul;
9531 const instruction INS_MULADD = INS_madd;
9532 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9533 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9534 const instruction INS_BREAKPOINT = INS_bkpt;
9535 const instruction INS_ADDC = INS_adc;
9536 const instruction INS_SUBC = INS_sbc;
9537 const instruction INS_NOT = INS_mvn;
9539 const instruction INS_ABS = INS_fabs;
9540 const instruction INS_ROUND = INS_frintn;
9541 const instruction INS_SQRT = INS_fsqrt;
9545 /*****************************************************************************/
9547 extern const BYTE genTypeSizes[];
9548 extern const BYTE genTypeAlignments[];
9549 extern const BYTE genTypeStSzs[];
9550 extern const BYTE genActualTypes[];
9552 /*****************************************************************************/
9554 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9555 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9558 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9559 #elif defined(_TARGET_ARM64_)
9560 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9563 /*****************************************************************************/
9565 #define REG_CORRUPT regNumber(REG_NA + 1)
9566 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9567 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9569 /*****************************************************************************/
9571 extern BasicBlock dummyBB;
9573 /*****************************************************************************/
9574 /*****************************************************************************/
9576 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9577 // nodes of a statement in execution order.
9578 // __stmt: a GT_STMT type GenTree*
9579 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9581 #define foreach_treenode_execution_order(__node, __stmt) \
9582 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9584 // foreach_block: An iterator over all blocks in the function.
9585 // __compiler: the Compiler* object
9586 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9588 #define foreach_block(__compiler, __block) \
9589 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9591 /*****************************************************************************/
9592 /*****************************************************************************/
9596 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9598 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9599 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9601 XX Debugging helpers XX
9603 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9604 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9607 /*****************************************************************************/
9608 /* The following functions are intended to be called from the debugger, to dump
9609 * various data structures. The can be used in the debugger Watch or Quick Watch
9610 * windows. They are designed to be short to type and take as few arguments as
9611 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9612 * See the function definition comment for more details.
9615 void cBlock(Compiler* comp, BasicBlock* block);
9616 void cBlocks(Compiler* comp);
9617 void cBlocksV(Compiler* comp);
9618 void cTree(Compiler* comp, GenTree* tree);
9619 void cTrees(Compiler* comp);
9620 void cEH(Compiler* comp);
9621 void cVar(Compiler* comp, unsigned lclNum);
9622 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9623 void cVars(Compiler* comp);
9624 void cVarsFinal(Compiler* comp);
9625 void cBlockPreds(Compiler* comp, BasicBlock* block);
9626 void cReach(Compiler* comp);
9627 void cDoms(Compiler* comp);
9628 void cLiveness(Compiler* comp);
9629 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9631 void cFuncIR(Compiler* comp);
9632 void cBlockIR(Compiler* comp, BasicBlock* block);
9633 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9634 void cTreeIR(Compiler* comp, GenTree* tree);
9635 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9636 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9637 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9638 int cOperandIR(Compiler* comp, GenTree* operand);
9639 int cLeafIR(Compiler* comp, GenTree* tree);
9640 int cIndirIR(Compiler* comp, GenTree* tree);
9641 int cListIR(Compiler* comp, GenTree* list);
9642 int cSsaNumIR(Compiler* comp, GenTree* tree);
9643 int cValNumIR(Compiler* comp, GenTree* tree);
9644 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9646 void dBlock(BasicBlock* block);
9649 void dTree(GenTree* tree);
9652 void dVar(unsigned lclNum);
9653 void dVarDsc(LclVarDsc* varDsc);
9656 void dBlockPreds(BasicBlock* block);
9660 void dCVarSet(VARSET_VALARG_TP vars);
9662 void dVarSet(VARSET_VALARG_TP vars);
9663 void dRegMask(regMaskTP mask);
9666 void dBlockIR(BasicBlock* block);
9667 void dTreeIR(GenTree* tree);
9668 void dLoopIR(Compiler::LoopDsc* loop);
9669 void dLoopNumIR(unsigned loopNum);
9670 int dTabStopIR(int curr, int tabstop);
9671 int dTreeTypeIR(GenTree* tree);
9672 int dTreeKindsIR(GenTree* tree);
9673 int dTreeFlagsIR(GenTree* tree);
9674 int dOperandIR(GenTree* operand);
9675 int dLeafIR(GenTree* tree);
9676 int dIndirIR(GenTree* tree);
9677 int dListIR(GenTree* list);
9678 int dSsaNumIR(GenTree* tree);
9679 int dValNumIR(GenTree* tree);
9680 int dDependsIR(GenTree* comma);
9683 GenTree* dFindTree(GenTree* tree, unsigned id);
9684 GenTree* dFindTree(unsigned id);
9685 GenTreeStmt* dFindStmt(unsigned id);
9686 BasicBlock* dFindBlock(unsigned bbNum);
9690 #include "compiler.hpp" // All the shared inline functions
9692 /*****************************************************************************/
9693 #endif //_COMPILER_H_
9694 /*****************************************************************************/