1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 #ifndef LEGACY_BACKEND
89 class Lowering; // defined in lower.h
92 // The following are defined in this file, Compiler.h
96 /*****************************************************************************
102 /*****************************************************************************/
105 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
108 // Or the more-general IAllocator interface.
109 void* __cdecl operator new(size_t n, IAllocator* alloc);
110 void* __cdecl operator new[](size_t n, IAllocator* alloc);
112 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
113 // caused these to be ambiguous with the global placement new operators.
114 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
115 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
116 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
118 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
119 #include "loopcloning.h"
121 /*****************************************************************************/
123 /* This is included here and not earlier as it needs the definition of "CSE"
124 * which is defined in the section above */
126 /*****************************************************************************/
128 unsigned genLog2(unsigned value);
129 unsigned genLog2(unsigned __int64 value);
131 var_types genActualType(var_types type);
132 var_types genUnsignedType(var_types type);
133 var_types genSignedType(var_types type);
135 unsigned ReinterpretHexAsDecimal(unsigned);
137 /*****************************************************************************/
140 #ifdef FEATURE_AVX_SUPPORT
141 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
142 #else // !FEATURE_AVX_SUPPORT
143 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
144 #endif // !FEATURE_AVX_SUPPORT
145 #else // !FEATURE_SIMD
146 const unsigned TEMP_MAX_SIZE = sizeof(double);
147 #endif // !FEATURE_SIMD
148 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
150 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
153 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
156 // The following holds the Local var info (scope information)
157 typedef const char* VarName; // Actual ASCII string
160 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
161 IL_OFFSET vsdLifeEnd; // instr offset of end of life
162 unsigned vsdVarNum; // (remapped) LclVarDsc number
165 VarName vsdName; // name of the var
168 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
169 // Also, it is the index of this entry in the info.compVarScopes array,
170 // which is useful since the array is also accessed via the
171 // compEnterScopeList and compExitScopeList sorted arrays.
174 /*****************************************************************************
176 * The following holds the local variable counts and the descriptor table.
179 // This is the location of a definition.
185 DefLoc() : m_blk(nullptr), m_tree(nullptr)
190 // This class encapsulates all info about a local variable that may vary for different SSA names
195 ValueNumPair m_vnPair;
203 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
208 // The constructor. Most things can just be zero'ed.
209 LclVarDsc(Compiler* comp);
211 // note this only packs because var_types is a typedef of unsigned char
212 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
214 unsigned char lvIsParam : 1; // is this a parameter?
215 unsigned char lvIsRegArg : 1; // is this a register argument?
216 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
218 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
219 // is to help determine whether to use block init in the prolog.
220 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
221 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
222 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
223 // variable is in the same register for the entire function.
224 unsigned char lvTracked : 1; // is this a tracked variable?
225 bool lvTrackedNonStruct()
227 return lvTracked && lvType != TYP_STRUCT;
229 unsigned char lvPinned : 1; // is this a pinned variable?
231 unsigned char lvMustInit : 1; // must be initialized
232 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
233 // global location, etc.
234 // We cannot reason reliably about the value of the variable.
235 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
236 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
240 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
242 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
243 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
244 // read/write its value.
245 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
246 // the variable to be
247 // in the stack (at least at those boundaries.)
248 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
249 // particular byte from an int).
250 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
251 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
253 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
254 unsigned char lvRefAssign : 1; // involved in pointer assignment
255 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
256 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
259 unsigned char lvHasILStoreOp : 1; // there is at least one STLOC or STARG on this local
260 unsigned char lvHasMultipleILStoreOp : 1; // there is more than one STLOC on this local
262 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp (if lvIsParam is false), or implicit byref parameter
263 // (if lvIsParam is true)
265 unsigned char lvIsBoolean : 1; // set if variable is boolean
267 unsigned char lvRngOptDone : 1; // considered for range check opt?
268 unsigned char lvLoopInc : 1; // incremented in the loop?
269 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
270 unsigned char lvArrIndx : 1; // used as an array index?
271 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
272 unsigned char lvArrIndxDom : 1; // index dominates loop exit
274 unsigned char lvSingleDef : 1; // variable has a single def
275 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
276 unsigned char lvVolatileHint : 1; // hint for AssertionProp
279 unsigned char lvSpilled : 1; // enregistered variable was spilled
280 #ifndef _TARGET_64BIT_
281 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
282 #endif // !_TARGET_64BIT_
283 #ifdef _TARGET_64BIT_
284 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
287 unsigned char lvKeepType : 1; // Don't change the type of this variable
288 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
290 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
292 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
293 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
294 // 32-bit target. For implicit byref parameters, this gets hijacked between
295 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to indicate whether
296 // references to the arg are being rewritten as references to a promoted shadow local.
297 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
298 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
299 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
300 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
301 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
303 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
304 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
307 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
308 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
309 // with (lvIsRegArg && lvIsHfa())
310 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
311 #endif // FEATURE_HFA
314 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
315 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
316 // Consider cleaning this up so this workaround is not required.
317 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
318 // I.e. there is no longer any reference to the struct directly.
319 // In this case we can simply remove this struct local.
321 #ifndef LEGACY_BACKEND
322 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
323 #endif // !LEGACY_BACKEND
326 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
327 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
328 unsigned char lvSIMDType : 1; // This is a SIMD struct
329 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
330 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
331 #endif // FEATURE_SIMD
332 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
334 unsigned char lvClassIsExact : 1; // lvClassHandle is the exact type
337 unsigned char lvClassInfoUpdated : 1; // true if this var has updated class handle or exactness
341 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
342 // local. For implicit byref parameters, this gets hijacked between
343 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to point to the
344 // struct local created to model the parameter's struct promotion, if any.
345 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
346 // Valid on promoted struct local fields.
349 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
350 unsigned char lvFldOffset;
351 unsigned char lvFldOrdinal;
353 #if FEATURE_MULTIREG_ARGS
354 regNumber lvRegNumForSlot(unsigned slotNum)
360 else if (slotNum == 1)
362 return lvOtherArgReg;
366 assert(false && "Invalid slotNum!");
371 #endif // FEATURE_MULTIREG_ARGS
389 bool lvIsHfaRegArg() const
392 return _lvIsHfaRegArg;
398 void lvSetIsHfaRegArg(bool value = true)
401 _lvIsHfaRegArg = value;
405 bool lvHfaTypeIsFloat() const
408 return _lvHfaTypeIsFloat;
414 void lvSetHfaTypeIsFloat(bool value)
417 _lvHfaTypeIsFloat = value;
421 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
422 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
424 unsigned lvHfaSlots() const
427 assert(lvType == TYP_STRUCT);
429 return lvExactSize / sizeof(float);
430 #else // _TARGET_ARM64_
431 if (lvHfaTypeIsFloat())
433 return lvExactSize / sizeof(float);
437 return lvExactSize / sizeof(double);
439 #endif // _TARGET_ARM64_
442 // lvIsMultiRegArgOrRet()
443 // returns true if this is a multireg LclVar struct used in an argument context
444 // or if this is a multireg LclVar struct assigned from a multireg call
445 bool lvIsMultiRegArgOrRet()
447 return lvIsMultiRegArg || lvIsMultiRegRet;
451 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
452 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
453 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
454 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
455 // to non-zero if the variable gets the same register assignment for its entire
457 #if !defined(_TARGET_64BIT_)
458 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
459 #endif // !defined(_TARGET_64BIT_)
461 regNumberSmall _lvArgReg; // The register in which this argument is passed.
463 #if FEATURE_MULTIREG_ARGS
464 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
465 // Note this is defined but not used by ARM32
466 #endif // FEATURE_MULTIREG_ARGS
468 #ifndef LEGACY_BACKEND
470 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
471 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
473 #endif // !LEGACY_BACKEND
476 // The register number is stored in a small format (8 bits), but the getters return and the setters take
477 // a full-size (unsigned) format, to localize the casts here.
479 /////////////////////
481 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
483 regNumber GetRegNum() const
485 return (regNumber)_lvRegNum;
488 void SetRegNum(regNumber reg)
490 _lvRegNum = (regNumberSmall)reg;
491 assert(_lvRegNum == reg);
494 /////////////////////
496 #if defined(_TARGET_64BIT_)
497 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
499 regNumber GetOtherReg() const
501 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
502 // "unreachable code" warnings
506 void SetOtherReg(regNumber reg)
508 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
509 // "unreachable code" warnings
511 #else // !_TARGET_64BIT_
512 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
514 regNumber GetOtherReg() const
516 return (regNumber)_lvOtherReg;
519 void SetOtherReg(regNumber reg)
521 _lvOtherReg = (regNumberSmall)reg;
522 assert(_lvOtherReg == reg);
524 #endif // !_TARGET_64BIT_
526 /////////////////////
528 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
530 regNumber GetArgReg() const
532 return (regNumber)_lvArgReg;
535 void SetArgReg(regNumber reg)
537 _lvArgReg = (regNumberSmall)reg;
538 assert(_lvArgReg == reg);
541 #if FEATURE_MULTIREG_ARGS
542 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
544 regNumber GetOtherArgReg() const
546 return (regNumber)_lvOtherArgReg;
549 void SetOtherArgReg(regNumber reg)
551 _lvOtherArgReg = (regNumberSmall)reg;
552 assert(_lvOtherArgReg == reg);
554 #endif // FEATURE_MULTIREG_ARGS
557 // Is this is a SIMD struct?
558 bool lvIsSIMDType() const
563 // Is this is a SIMD struct which is used for SIMD intrinsic?
564 bool lvIsUsedInSIMDIntrinsic() const
566 return lvUsedInSIMDIntrinsic;
569 // If feature_simd not enabled, return false
570 bool lvIsSIMDType() const
574 bool lvIsUsedInSIMDIntrinsic() const
580 /////////////////////
582 #ifndef LEGACY_BACKEND
583 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
585 regNumber GetArgInitReg() const
587 return (regNumber)_lvArgInitReg;
590 void SetArgInitReg(regNumber reg)
592 _lvArgInitReg = (regNumberSmall)reg;
593 assert(_lvArgInitReg == reg);
596 /////////////////////
598 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
600 regPairNo GetArgInitRegPair() const
602 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
603 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
607 void SetArgInitRegPair(regPairNo regPair)
609 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
610 _lvArgInitRegPair = (regPairNoSmall)regPair;
611 assert(_lvArgInitRegPair == regPair);
614 /////////////////////
616 bool lvIsRegCandidate() const
618 return lvLRACandidate != 0;
621 bool lvIsInReg() const
623 return lvIsRegCandidate() && (lvRegNum != REG_STK);
626 #else // LEGACY_BACKEND
628 bool lvIsRegCandidate() const
630 return lvTracked != 0;
633 bool lvIsInReg() const
635 return lvRegister != 0;
638 #endif // LEGACY_BACKEND
640 regMaskTP lvRegMask() const
642 regMaskTP regMask = RBM_NONE;
643 if (varTypeIsFloating(TypeGet()))
645 if (lvRegNum != REG_STK)
647 regMask = genRegMaskFloat(lvRegNum, TypeGet());
652 if (lvRegNum != REG_STK)
654 regMask = genRegMask(lvRegNum);
657 // For longs we may have two regs
658 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
660 regMask |= genRegMask(lvOtherReg);
666 regMaskSmall lvPrefReg; // set of regs it prefers to live in
668 unsigned short lvVarIndex; // variable tracking index
669 unsigned short lvRefCnt; // unweighted (real) reference count. For implicit by reference
670 // parameters, this gets hijacked from fgMarkImplicitByRefArgs
671 // through fgMarkDemotedImplicitByRefArgs, to provide a static
672 // appearance count (computed during address-exposed analysis)
673 // that fgMakeOutgoingStructArgCopy consults during global morph
674 // to determine if eliding its copy is legal.
675 unsigned lvRefCntWtd; // weighted reference count
676 int lvStkOffs; // stack offset of home
677 unsigned lvExactSize; // (exact) size of the type in bytes
679 // Is this a promoted struct?
680 // This method returns true only for structs (including SIMD structs), not for
681 // locals that are split on a 32-bit target.
682 // It is only necessary to use this:
683 // 1) if only structs are wanted, and
684 // 2) if Lowering has already been done.
685 // Otherwise lvPromoted is valid.
686 bool lvPromotedStruct()
688 #if !defined(_TARGET_64BIT_)
689 return (lvPromoted && !varTypeIsLong(lvType));
690 #else // defined(_TARGET_64BIT_)
692 #endif // defined(_TARGET_64BIT_)
695 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
697 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
698 // where the struct itself is no longer used because all access is via its member fields.
699 // When that happens, the struct is marked as unused and its type has been changed to
700 // TYP_INT (to keep the GC tracking code from looking at it).
701 // See Compiler::raAssignVars() for details. For example:
702 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
703 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
704 // float V03.f1 (offs=0x00) -> V12 tmp7
705 // f8 (last use) (last use) $345
706 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
707 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
709 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
711 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
712 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
713 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
714 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
715 if ((lvType == TYP_SIMD12) && !lvIsParam)
717 assert(lvExactSize == 12);
720 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
722 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
725 unsigned lvSlotNum; // original slot # (if remapped)
727 typeInfo lvVerTypeInfo; // type info needed for verification
729 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
731 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
733 BYTE* lvGcLayout; // GC layout info for structs
736 BlockSet lvRefBlks; // Set of blocks that contain refs
737 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
738 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
740 var_types TypeGet() const
742 return (var_types)lvType;
744 bool lvStackAligned() const
746 assert(lvIsStructField);
747 return ((lvFldOffset % sizeof(void*)) == 0);
749 bool lvNormalizeOnLoad() const
751 return varTypeIsSmall(TypeGet()) &&
752 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
753 (lvIsParam || lvAddrExposed || lvIsStructField);
756 bool lvNormalizeOnStore()
758 return varTypeIsSmall(TypeGet()) &&
759 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
760 !(lvIsParam || lvAddrExposed || lvIsStructField);
763 void lvaResetSortAgainFlag(Compiler* pComp);
764 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
765 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
766 void setPrefReg(regNumber regNum, Compiler* pComp);
767 void addPrefReg(regMaskTP regMask, Compiler* pComp);
768 bool IsFloatRegType() const
770 return isFloatRegType(lvType) || lvIsHfaRegArg();
772 var_types GetHfaType() const
774 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
776 void SetHfaType(var_types type)
778 assert(varTypeIsFloating(type));
779 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
782 #ifndef LEGACY_BACKEND
783 var_types lvaArgType();
786 PerSsaArray lvPerSsaData;
789 // Keep track of the # of SsaNames, for a bounds check.
790 unsigned lvNumSsaNames;
793 // Returns the address of the per-Ssa data for the given ssaNum (which is required
794 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
795 // not an SSA variable).
796 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
798 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
799 assert(SsaConfig::RESERVED_SSA_NUM == 0);
800 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
801 assert(zeroBased < lvNumSsaNames);
802 return &lvPerSsaData.GetRef(zeroBased);
807 void PrintVarReg() const
809 if (isRegPairType(TypeGet()))
811 printf("%s:%s", getRegName(lvOtherReg), // hi32
812 getRegName(lvRegNum)); // lo32
816 printf("%s", getRegName(lvRegNum));
821 }; // class LclVarDsc
824 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
825 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
829 XX The temporary lclVars allocated by the compiler for code generation XX
831 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
832 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
835 /*****************************************************************************
837 * The following keeps track of temporaries allocated in the stack frame
838 * during code-generation (after register allocation). These spill-temps are
839 * only used if we run out of registers while evaluating a tree.
841 * These are different from the more common temps allocated by lvaGrabTemp().
852 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
860 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
864 0); // temps must have a negative number (so they have a different number from all local variables)
865 tdOffs = BAD_TEMP_OFFSET;
869 IMPL_LIMITATION("too many spill temps");
874 bool tdLegalOffset() const
876 return tdOffs != BAD_TEMP_OFFSET;
880 int tdTempOffs() const
882 assert(tdLegalOffset());
885 void tdSetTempOffs(int offs)
888 assert(tdLegalOffset());
890 void tdAdjustTempOffs(int offs)
893 assert(tdLegalOffset());
896 int tdTempNum() const
901 unsigned tdTempSize() const
905 var_types tdTempType() const
911 // interface to hide linearscan implementation from rest of compiler
912 class LinearScanInterface
915 virtual void doLinearScan() = 0;
916 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
917 virtual bool willEnregisterLocalVars() const = 0;
920 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
922 // Information about arrays: their element type and size, and the offset of the first element.
923 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
924 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
925 // for example, in value numbering of array index expressions.
928 var_types m_elemType;
929 CORINFO_CLASS_HANDLE m_elemStructType;
931 unsigned m_elemOffset;
933 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
937 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
938 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
943 // This enumeration names the phases into which we divide compilation. The phases should completely
944 // partition a compilation.
947 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
948 #include "compphases.h"
952 extern const char* PhaseNames[];
953 extern const char* PhaseEnums[];
954 extern const LPCWSTR PhaseShortNames[];
956 // The following enum provides a simple 1:1 mapping to CLR API's
957 enum API_ICorJitInfo_Names
959 #define DEF_CLR_API(name) API_##name,
960 #include "ICorJitInfo_API_names.h"
964 //---------------------------------------------------------------
968 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
969 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
970 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
971 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
972 // by "m_timerFailure" being true.
973 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
976 #ifdef FEATURE_JIT_METHOD_PERF
977 // The string names of the phases.
978 static const char* PhaseNames[];
980 static bool PhaseHasChildren[];
981 static int PhaseParent[];
982 static bool PhaseReportsIRSize[];
984 unsigned m_byteCodeBytes;
985 unsigned __int64 m_totalCycles;
986 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
987 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
988 #if MEASURE_CLRAPI_CALLS
989 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
990 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
993 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
995 // For better documentation, we call EndPhase on
996 // non-leaf phases. We should also call EndPhase on the
997 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
998 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
999 // We add all such "redundant end phase" intervals to this variable below; we print
1000 // it out in a report, so we can verify that it is, indeed, very small. If it ever
1001 // isn't, this means that we're doing something significant between the end of the last
1002 // declared subphase and the end of its parent.
1003 unsigned __int64 m_parentPhaseEndSlop;
1004 bool m_timerFailure;
1006 #if MEASURE_CLRAPI_CALLS
1007 // The following measures the time spent inside each individual CLR API call.
1008 unsigned m_allClrAPIcalls;
1009 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1010 unsigned __int64 m_allClrAPIcycles;
1011 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1012 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1013 #endif // MEASURE_CLRAPI_CALLS
1015 CompTimeInfo(unsigned byteCodeBytes);
1019 #ifdef FEATURE_JIT_METHOD_PERF
1021 #if MEASURE_CLRAPI_CALLS
1022 struct WrapICorJitInfo;
1025 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1026 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1027 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1028 // threads, so it is protected by a lock.
1029 // This class is intended to be used as a singleton type, with only a single instance.
1030 class CompTimeSummaryInfo
1032 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1033 static CritSecObject s_compTimeSummaryLock;
1037 CompTimeInfo m_total;
1038 CompTimeInfo m_maximum;
1040 int m_numFilteredMethods;
1041 CompTimeInfo m_filtered;
1043 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1044 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1045 // If any OS interaction fails, returns 0.0.
1046 double CyclesPerSecond();
1048 // This can use what ever data you want to determine if the value to be added
1049 // belongs in the filtered section (it's always included in the unfiltered section)
1050 bool IncludedInFilteredData(CompTimeInfo& info);
1053 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1054 static CompTimeSummaryInfo s_compTimeSummary;
1056 CompTimeSummaryInfo()
1057 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1061 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1062 // This is thread safe.
1063 void AddInfo(CompTimeInfo& info, bool includePhases);
1065 // Print the summary information to "f".
1066 // This is not thread-safe; assumed to be called by only one thread.
1067 void Print(FILE* f);
1070 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1071 // and when the current phase started. This is intended to be part of a Compilation object. This is
1072 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1076 unsigned __int64 m_start; // Start of the compilation.
1077 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1078 #if MEASURE_CLRAPI_CALLS
1079 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1080 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1081 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1082 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1083 static double s_cyclesPerSec; // Cached for speedier measurements
1086 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1088 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1090 static CritSecObject s_csvLock; // Lock to protect the time log file.
1091 void PrintCsvMethodStats(Compiler* comp);
1094 void* operator new(size_t);
1095 void* operator new[](size_t);
1096 void operator delete(void*);
1097 void operator delete[](void*);
1100 // Initialized the timer instance
1101 JitTimer(unsigned byteCodeSize);
1103 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1105 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1108 static void PrintCsvHeader();
1110 // Ends the current phase (argument is for a redundant check).
1111 void EndPhase(Compiler* compiler, Phases phase);
1113 #if MEASURE_CLRAPI_CALLS
1114 // Start and end a timed CLR API call.
1115 void CLRApiCallEnter(unsigned apix);
1116 void CLRApiCallLeave(unsigned apix);
1117 #endif // MEASURE_CLRAPI_CALLS
1119 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1120 // and adds it to "sum".
1121 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1123 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1124 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1125 // "m_info" to true.
1126 bool GetThreadCycles(unsigned __int64* cycles)
1128 bool res = CycleTimer::GetThreadCyclesS(cycles);
1131 m_info.m_timerFailure = true;
1136 #endif // FEATURE_JIT_METHOD_PERF
1138 //------------------- Function/Funclet info -------------------------------
1139 enum FuncKind : BYTE
1141 FUNC_ROOT, // The main/root function (always id==0)
1142 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1143 FUNC_FILTER, // a funclet associated with an EH filter
1152 BYTE funFlags; // Currently unused, just here for padding
1153 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1154 // funclet. It is only valid if funKind field indicates this is a
1155 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1157 #if defined(_TARGET_AMD64_)
1159 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1160 emitLocation* startLoc;
1161 emitLocation* endLoc;
1162 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1163 emitLocation* coldEndLoc;
1164 UNWIND_INFO unwindHeader;
1165 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1166 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1167 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1168 unsigned unwindCodeSlot;
1170 #ifdef UNIX_AMD64_ABI
1171 jitstd::vector<CFI_CODE>* cfiCodes;
1172 #endif // UNIX_AMD64_ABI
1174 #elif defined(_TARGET_ARMARCH_)
1176 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1177 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1178 // Note: we only have a pointer here instead of the actual object,
1179 // to save memory in the JIT case (compared to the NGEN case),
1180 // where we don't have any cold section.
1181 // Note 2: we currently don't support hot/cold splitting in functions
1182 // with EH, so uwiCold will be NULL for all funclets.
1184 #endif // _TARGET_ARMARCH_
1186 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1187 // that isn't shared between the main function body and funclets.
1190 struct fgArgTabEntry
1193 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1196 otherRegNum = REG_NA;
1197 isStruct = false; // is this a struct arg
1199 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1201 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1203 // it will point at the actual argument in the gtCallLateArgs list.
1204 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1206 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1208 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1210 unsigned numRegs; // Count of number of registers that this argument uses
1212 // A slot is a pointer sized region in the OutArg area.
1213 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1214 unsigned numSlots; // Count of number of slots that this argument uses
1216 unsigned alignment; // 1 or 2 (slots/registers)
1217 unsigned lateArgInx; // index into gtCallLateArgs list
1218 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1220 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1221 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1222 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1223 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1224 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1225 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1226 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1227 // previous arguments.
1228 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1229 // to be on the stack despite its arg list position.
1231 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1232 bool isStruct : 1; // True if this is a struct arg
1234 regNumber otherRegNum; // The (second) register to use when passing this argument.
1236 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1237 #elif !defined(_TARGET_64BIT_)
1238 __declspec(property(get = getIsStruct)) bool isStruct;
1241 return varTypeIsStruct(node);
1243 #endif // !_TARGET_64BIT_
1246 void SetIsHfaRegArg(bool hfaRegArg)
1248 isHfaRegArg = hfaRegArg;
1251 void SetIsBackFilled(bool backFilled)
1253 isBackFilled = backFilled;
1256 bool IsBackFilled() const
1258 return isBackFilled;
1260 #else // !_TARGET_ARM_
1261 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1263 void SetIsHfaRegArg(bool hfaRegArg)
1267 void SetIsBackFilled(bool backFilled)
1271 bool IsBackFilled() const
1275 #endif // !_TARGET_ARM_
1281 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1283 //-------------------------------------------------------------------------
1285 // The class fgArgInfo is used to handle the arguments
1286 // when morphing a GT_CALL node.
1291 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1292 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1293 unsigned argCount; // Updatable arg count value
1294 unsigned nextSlotNum; // Updatable slot count value
1295 unsigned stkLevel; // Stack depth when we make this call (for x86)
1297 #if defined(UNIX_X86_ABI)
1298 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1299 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1300 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1301 // Computed dynamically during codegen, based on stkSizeBytes and the current
1302 // stack level (genStackLevel) when the first stack adjustment is made for
1306 #if FEATURE_FIXED_OUT_ARGS
1307 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1310 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1311 bool hasRegArgs; // true if we have one or more register arguments
1312 bool hasStackArgs; // true if we have one or more stack arguments
1313 bool argsComplete; // marker for state
1314 bool argsSorted; // marker for state
1315 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1318 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1321 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1322 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1324 fgArgTabEntryPtr AddRegArg(
1325 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1327 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1328 fgArgTabEntryPtr AddRegArg(
1335 const bool isStruct,
1336 const regNumber otherRegNum = REG_NA,
1337 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1338 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1340 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1344 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1346 void RemorphReset();
1347 fgArgTabEntryPtr RemorphRegArg(
1348 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1350 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1352 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1354 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1356 void ArgsComplete();
1360 void EvalArgsToTemps();
1362 void RecordStkLevel(unsigned stkLvl);
1363 unsigned RetrieveStkLevel();
1369 fgArgTabEntryPtr* ArgTable()
1373 unsigned GetNextSlotNum()
1383 return hasStackArgs;
1385 bool AreArgsComplete() const
1387 return argsComplete;
1389 #if FEATURE_FIXED_OUT_ARGS
1390 unsigned GetOutArgSize() const
1394 void SetOutArgSize(unsigned newVal)
1396 outArgSize = newVal;
1398 #endif // FEATURE_FIXED_OUT_ARGS
1400 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1402 #if defined(UNIX_X86_ABI)
1403 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1404 #endif // defined(UNIX_X86_ABI)
1407 void SetStkSizeBytes(unsigned newStkSizeBytes)
1409 #if defined(UNIX_X86_ABI)
1410 stkSizeBytes = newStkSizeBytes;
1411 #endif // defined(UNIX_X86_ABI)
1414 #if defined(UNIX_X86_ABI)
1415 unsigned GetStkAlign()
1419 unsigned GetStkSizeBytes() const
1421 return stkSizeBytes;
1423 bool IsStkAlignmentDone() const
1425 return alignmentDone;
1427 void SetStkAlignmentDone()
1429 alignmentDone = true;
1431 #endif // defined(UNIX_X86_ABI)
1433 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1434 GenTreePtr GetLateArg(unsigned argIndex);
1438 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1439 // We have the ability to mark source expressions with "Test Labels."
1440 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1441 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1443 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1446 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1447 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1448 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1449 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1450 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1453 struct TestLabelAndNum
1458 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1463 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1465 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1468 // This class implements the "IAllocator" interface, so that we can use
1469 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1471 class CompAllocator : public IAllocator
1474 #if MEASURE_MEM_ALLOC
1478 CompAllocator(Compiler* comp, CompMemKind cmk)
1480 #if MEASURE_MEM_ALLOC
1486 inline void* Alloc(size_t sz);
1488 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1490 // For the compiler's no-release allocator, free operations are no-ops.
1497 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1498 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1500 XX The big guy. The sections are currently organized as : XX
1502 XX o GenTree and BasicBlock XX
1514 XX o PrologScopeInfo XX
1515 XX o CodeGenerator XX
1520 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1521 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1526 friend class emitter;
1527 friend class UnwindInfo;
1528 friend class UnwindFragmentInfo;
1529 friend class UnwindEpilogInfo;
1530 friend class JitTimer;
1531 friend class LinearScan;
1532 friend class fgArgInfo;
1533 friend class Rationalizer;
1535 friend class Lowering;
1536 friend class CSE_DataFlow;
1537 friend class CSE_Heuristic;
1538 friend class CodeGenInterface;
1539 friend class CodeGen;
1540 friend class LclVarDsc;
1541 friend class TempDsc;
1543 friend class ObjectAllocator;
1545 #ifndef _TARGET_64BIT_
1546 friend class DecomposeLongs;
1547 #endif // !_TARGET_64BIT_
1550 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1551 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1553 XX Misc structs definitions XX
1555 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1556 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1560 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1579 bool dumpIRDataflow;
1580 bool dumpIRBlockHeaders;
1582 LPCWSTR dumpIRPhase;
1583 LPCWSTR dumpIRFormat;
1585 bool shouldUseVerboseTrees();
1586 bool asciiTrees; // If true, dump trees using only ASCII characters
1587 bool shouldDumpASCIITrees();
1588 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1589 bool shouldUseVerboseSsa();
1590 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1591 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1593 const char* VarNameToStr(VarName name)
1598 DWORD expensiveDebugCheckLevel;
1601 #if FEATURE_MULTIREG_RET
1602 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1603 #endif // FEATURE_MULTIREG_RET
1606 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1607 #endif // ARM_SOFTFP
1609 //-------------------------------------------------------------------------
1610 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1611 // HFAs are one to four element structs where each element is the same
1612 // type, either all float or all double. They are treated specially
1613 // in the ARM Procedure Call Standard, specifically, they are passed in
1614 // floating-point registers instead of the general purpose registers.
1617 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1618 bool IsHfa(GenTreePtr tree);
1620 var_types GetHfaType(GenTreePtr tree);
1621 unsigned GetHfaCount(GenTreePtr tree);
1623 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1624 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1626 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1627 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1629 //-------------------------------------------------------------------------
1630 // The following is used for validating format of EH table
1634 typedef struct EHNodeDsc* pEHNodeDsc;
1636 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1637 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1650 EHBlockType ehnBlockType; // kind of EH block
1651 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1652 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1653 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1655 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1656 pEHNodeDsc ehnChild; // leftmost nested block
1658 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1659 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1661 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1662 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1664 inline void ehnSetTryNodeType()
1666 ehnBlockType = TryNode;
1668 inline void ehnSetFilterNodeType()
1670 ehnBlockType = FilterNode;
1672 inline void ehnSetHandlerNodeType()
1674 ehnBlockType = HandlerNode;
1676 inline void ehnSetFinallyNodeType()
1678 ehnBlockType = FinallyNode;
1680 inline void ehnSetFaultNodeType()
1682 ehnBlockType = FaultNode;
1685 inline BOOL ehnIsTryBlock()
1687 return ehnBlockType == TryNode;
1689 inline BOOL ehnIsFilterBlock()
1691 return ehnBlockType == FilterNode;
1693 inline BOOL ehnIsHandlerBlock()
1695 return ehnBlockType == HandlerNode;
1697 inline BOOL ehnIsFinallyBlock()
1699 return ehnBlockType == FinallyNode;
1701 inline BOOL ehnIsFaultBlock()
1703 return ehnBlockType == FaultNode;
1706 // returns true if there is any overlap between the two nodes
1707 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1709 if (node1->ehnStartOffset < node2->ehnStartOffset)
1711 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1715 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1719 // fails with BADCODE if inner is not completely nested inside outer
1720 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1722 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1726 //-------------------------------------------------------------------------
1727 // Exception handling functions
1730 #if !FEATURE_EH_FUNCLETS
1732 bool ehNeedsShadowSPslots()
1734 return (info.compXcptnsCount || opts.compDbgEnC);
1737 // 0 for methods with no EH
1738 // 1 for methods with non-nested EH, or where only the try blocks are nested
1739 // 2 for a method with a catch within a catch
1741 unsigned ehMaxHndNestingCount;
1743 #endif // !FEATURE_EH_FUNCLETS
1745 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1746 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1748 bool bbInCatchHandlerILRange(BasicBlock* blk);
1749 bool bbInFilterILRange(BasicBlock* blk);
1750 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1751 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1752 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1753 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1754 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1756 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1757 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1759 // Returns true if "block" is the start of a try region.
1760 bool bbIsTryBeg(BasicBlock* block);
1762 // Returns true if "block" is the start of a handler or filter region.
1763 bool bbIsHandlerBeg(BasicBlock* block);
1765 // Returns true iff "block" is where control flows if an exception is raised in the
1766 // try region, and sets "*regionIndex" to the index of the try for the handler.
1767 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1768 // block of the filter, but not for the filter's handler.
1769 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1771 bool ehHasCallableHandlers();
1773 // Return the EH descriptor for the given region index.
1774 EHblkDsc* ehGetDsc(unsigned regionIndex);
1776 // Return the EH index given a region descriptor.
1777 unsigned ehGetIndex(EHblkDsc* ehDsc);
1779 // Return the EH descriptor index of the enclosing try, for the given region index.
1780 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1782 // Return the EH descriptor index of the enclosing handler, for the given region index.
1783 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1785 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1786 // block is not in a 'try' region).
1787 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1789 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1790 // if this block is not in a filter or handler region).
1791 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1793 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1794 // nullptr if this block's exceptions propagate to caller).
1795 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1797 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1798 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1799 bool ehIsBlockEHLast(BasicBlock* block);
1801 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1803 // Return the region index of the most nested EH region this block is in.
1804 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1806 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1807 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1809 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1810 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1811 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1812 // (It can never be a filter.)
1813 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1815 // A block has been deleted. Update the EH table appropriately.
1816 void ehUpdateForDeletedBlock(BasicBlock* block);
1818 // Determine whether a block can be deleted while preserving the EH normalization rules.
1819 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1821 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1822 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1824 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1825 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1826 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1827 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1828 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1829 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1830 // lives in a filter.)
1831 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1833 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1834 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1835 // (nullptr if the last block is the last block in the program).
1836 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1837 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1840 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1841 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1842 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1845 #if FEATURE_EH_FUNCLETS
1846 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1847 // if there is a filter that protects a region with a nested EH clause (such as a
1848 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1849 // genFuncletProlog() for more details. However, the VM seems to use it for more
1850 // purposes, maybe including debugging. Until we are sure otherwise, always create
1851 // a PSPSym for functions with any EH.
1852 bool ehNeedsPSPSym() const
1856 #else // _TARGET_X86_
1857 return compHndBBtabCount > 0;
1858 #endif // _TARGET_X86_
1861 bool ehAnyFunclets(); // Are there any funclets in this function?
1862 unsigned ehFuncletCount(); // Return the count of funclets in the function
1864 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1865 #else // !FEATURE_EH_FUNCLETS
1866 bool ehAnyFunclets()
1870 unsigned ehFuncletCount()
1875 unsigned bbThrowIndex(BasicBlock* blk)
1877 return blk->bbTryIndex;
1878 } // Get the index to use as the cache key for sharing throw blocks
1879 #endif // !FEATURE_EH_FUNCLETS
1881 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1882 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1883 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1884 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1885 // convenient to also consider it a predecessor.)
1886 flowList* BlockPredsWithEH(BasicBlock* blk);
1888 // This table is useful for memoization of the method above.
1889 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1891 BlockToFlowListMap* m_blockToEHPreds;
1892 BlockToFlowListMap* GetBlockToEHPreds()
1894 if (m_blockToEHPreds == nullptr)
1896 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1898 return m_blockToEHPreds;
1901 void* ehEmitCookie(BasicBlock* block);
1902 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1904 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1906 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1908 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1910 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1912 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1914 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1916 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1918 void fgAllocEHTable();
1920 void fgRemoveEHTableEntry(unsigned XTnum);
1922 #if FEATURE_EH_FUNCLETS
1924 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1926 #endif // FEATURE_EH_FUNCLETS
1930 #endif // !FEATURE_EH
1932 void fgSortEHTable();
1934 // Causes the EH table to obey some well-formedness conditions, by inserting
1935 // empty BB's when necessary:
1936 // * No block is both the first block of a handler and the first block of a try.
1937 // * No block is the first block of multiple 'try' regions.
1938 // * No block is the last block of multiple EH regions.
1939 void fgNormalizeEH();
1940 bool fgNormalizeEHCase1();
1941 bool fgNormalizeEHCase2();
1942 bool fgNormalizeEHCase3();
1945 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1946 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1947 void fgVerifyHandlerTab();
1948 void fgDispHandlerTab();
1951 bool fgNeedToSortEHTable;
1953 void verInitEHTree(unsigned numEHClauses);
1954 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1955 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1956 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1957 void verCheckNestingLevel(EHNodeDsc* initRoot);
1960 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1961 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1963 XX GenTree and BasicBlock XX
1965 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1967 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1968 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1971 // Functions to create nodes
1972 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1975 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1977 // For binary opers.
1978 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1980 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1982 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1983 var_types type = TYP_I_IMPL,
1984 GenTreePtr op1 = nullptr,
1985 GenTreePtr op2 = nullptr);
1987 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1989 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1991 GenTreePtr gtNewJmpTableNode();
1992 GenTreePtr gtNewIconHandleNode(
1993 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1995 unsigned gtTokenToIconFlags(unsigned token);
1997 GenTreePtr gtNewIconEmbHndNode(void* value,
2000 unsigned handle1 = 0,
2001 void* handle2 = nullptr,
2002 void* compileTimeHandle = nullptr);
2004 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2005 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2006 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2007 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2009 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2011 GenTreePtr gtNewLconNode(__int64 value);
2013 GenTreePtr gtNewDconNode(double value);
2015 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2017 GenTreePtr gtNewZeroConNode(var_types type);
2019 GenTreePtr gtNewOneConNode(var_types type);
2022 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2023 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2026 GenTreeBlk* gtNewBlkOpNode(
2027 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
2029 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2031 GenTree* gtNewPutArgReg(var_types type, GenTreePtr arg);
2034 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
2037 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2038 void gtSetObjGcInfo(GenTreeObj* objNode);
2039 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2040 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
2042 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2044 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
2046 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2047 CORINFO_METHOD_HANDLE handle,
2049 GenTreeArgList* args,
2050 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2052 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
2054 GenTreeArgList* args,
2055 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2057 GenTreeCall* gtNewHelperCallNode(unsigned helper,
2060 GenTreeArgList* args = nullptr);
2062 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2065 GenTreeSIMD* gtNewSIMDNode(
2066 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2067 GenTreeSIMD* gtNewSIMDNode(var_types type,
2070 SIMDIntrinsicID simdIntrinsicID,
2073 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2076 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2077 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2078 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2080 GenTreePtr gtNewCodeRef(BasicBlock* block);
2082 GenTreePtr gtNewFieldRef(
2083 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2085 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2087 GenTreeArgList* gtNewArgList(GenTreePtr op);
2088 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2089 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2091 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2092 static fgArgTabEntryPtr gtArgEntryByNode(GenTreeCall* call, GenTreePtr node);
2093 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2094 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2096 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2098 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2100 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2101 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2102 CORINFO_ACCESS_FLAGS access,
2103 CORINFO_FIELD_INFO* pFieldInfo,
2105 CORINFO_CLASS_HANDLE structType,
2108 GenTreePtr gtNewNothingNode();
2110 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2112 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2114 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2116 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2118 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2120 //------------------------------------------------------------------------
2121 // Other GenTree functions
2123 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2125 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2126 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2127 // IntCnses with value `deepVarVal`.
2128 GenTreePtr gtCloneExpr(
2129 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2131 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2132 // `varNum` to int constants with value `varVal`.
2133 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2135 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2138 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2140 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2142 // Returns "true" iff the complexity (not formally defined, but first interpretation
2143 // is #of nodes in subtree) of "tree" is greater than "limit".
2144 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2145 // before they have been set.)
2146 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2148 bool gtCompareTree(GenTree* op1, GenTree* op2);
2150 GenTreePtr gtReverseCond(GenTree* tree);
2152 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2154 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2156 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2158 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2161 unsigned gtHashValue(GenTree* tree);
2163 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2166 void gtPrepareCost(GenTree* tree);
2167 bool gtIsLikelyRegVar(GenTree* tree);
2169 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2171 // Returns true iff the secondNode can be swapped with firstNode.
2172 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2174 unsigned gtSetEvalOrder(GenTree* tree);
2176 #if FEATURE_STACK_FP_X87
2178 void gtComputeFPlvls(GenTreePtr tree);
2179 #endif // FEATURE_STACK_FP_X87
2181 void gtSetStmtInfo(GenTree* stmt);
2183 // Returns "true" iff "node" has any of the side effects in "flags".
2184 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2186 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2187 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2189 // Appends 'expr' in front of 'list'
2190 // 'list' will typically start off as 'nullptr'
2191 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2192 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2194 void gtExtractSideEffList(GenTreePtr expr,
2196 unsigned flags = GTF_SIDE_EFFECT,
2197 bool ignoreRoot = false);
2199 GenTreePtr gtGetThisArg(GenTreeCall* call);
2201 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2202 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2203 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2204 // the given "fldHnd", is such an object pointer.
2205 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2207 // Return true if call is a recursive call; return false otherwise.
2208 // Note when inlining, this looks for calls back to the root method.
2209 bool gtIsRecursiveCall(GenTreeCall* call)
2211 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2214 //-------------------------------------------------------------------------
2216 GenTreePtr gtFoldExpr(GenTreePtr tree);
2219 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2220 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2221 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2222 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2223 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2224 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2225 // optimizations for now.
2226 __attribute__((optnone))
2228 gtFoldExprConst(GenTreePtr tree);
2229 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2230 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2231 bool gtTryRemoveBoxUpstreamEffects(GenTreePtr tree);
2233 //-------------------------------------------------------------------------
2234 // Get the handle, if any.
2235 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2236 // Get the handle, and assert if not found.
2237 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2238 // Get the handle for a ref type.
2239 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2241 //-------------------------------------------------------------------------
2242 // Functions to display the trees
2245 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2247 void gtDispVN(GenTreePtr tree);
2248 void gtDispConst(GenTreePtr tree);
2249 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2250 void gtDispNodeName(GenTreePtr tree);
2251 void gtDispRegVal(GenTreePtr tree);
2263 void gtDispChild(GenTreePtr child,
2264 IndentStack* indentStack,
2266 __in_opt const char* msg = nullptr,
2267 bool topOnly = false);
2268 void gtDispTree(GenTreePtr tree,
2269 IndentStack* indentStack = nullptr,
2270 __in_opt const char* msg = nullptr,
2271 bool topOnly = false,
2272 bool isLIR = false);
2273 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2274 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2275 char* gtGetLclVarName(unsigned lclNum);
2276 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2277 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2278 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2279 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2280 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2281 void gtDispFieldSeq(FieldSeqNode* pfsn);
2283 void gtDispRange(LIR::ReadOnlyRange const& range);
2285 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2287 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2299 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2300 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2303 static fgWalkPreFn gtAssertColonCond;
2305 static fgWalkPreFn gtMarkColonCond;
2306 static fgWalkPreFn gtClearColonCond;
2308 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2309 bool gtHasCatchArg(GenTreePtr tree);
2310 bool gtHasUnmanagedCall(GenTreePtr tree);
2312 typedef ArrayStack<GenTree*> GenTreeStack;
2314 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2315 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2317 //=========================================================================
2318 // BasicBlock functions
2320 // This is a debug flag we will use to assert when creating block during codegen
2321 // as this interferes with procedure splitting. If you know what you're doing, set
2322 // it to true before creating the block. (DEBUG only)
2323 bool fgSafeBasicBlockCreation;
2326 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2329 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2330 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2334 XX The variables to be used by the code generator. XX
2336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2337 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2341 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2342 // be placed in the stack frame and it's fields must be laid out sequentially.
2344 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2345 // a local variable that can be enregistered or placed in the stack frame.
2346 // The fields do not need to be laid out sequentially
2348 enum lvaPromotionType
2350 PROMOTION_TYPE_NONE, // The struct local is not promoted
2351 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2352 // and its field locals are independent of its parent struct local.
2353 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2354 // but its field locals depend on its parent struct local.
2357 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2358 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2360 /*****************************************************************************/
2362 enum FrameLayoutState
2365 INITIAL_FRAME_LAYOUT,
2366 PRE_REGALLOC_FRAME_LAYOUT,
2367 REGALLOC_FRAME_LAYOUT,
2368 TENTATIVE_FRAME_LAYOUT,
2373 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2374 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2375 bool lvaSortAgain; // true: We need to sort the lvaTable
2376 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2377 unsigned lvaCount; // total number of locals
2379 unsigned lvaRefCount; // total number of references to locals
2380 LclVarDsc* lvaTable; // variable descriptor table
2381 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2383 LclVarDsc** lvaRefSorted; // table sorted by refcount
2385 unsigned short lvaTrackedCount; // actual # of locals being tracked
2386 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2388 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2389 // Only for AMD64 System V cache the first caller stack homed argument.
2390 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2391 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2394 VARSET_TP lvaTrackedVars; // set of tracked variables
2396 #ifndef _TARGET_64BIT_
2397 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2399 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2401 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2402 // It that changes, this changes. VarSets from different epochs
2403 // cannot be meaningfully combined.
2405 unsigned GetCurLVEpoch()
2410 // reverse map of tracked number to var number
2411 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2413 #ifdef LEGACY_BACKEND
2414 // variable interference graph
2415 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2418 // variable preference graph
2419 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2423 // # of procs compiled a with double-aligned stack
2424 static unsigned s_lvaDoubleAlignedProcsCount;
2428 // Getters and setters for address-exposed and do-not-enregister local var properties.
2429 bool lvaVarAddrExposed(unsigned varNum);
2430 void lvaSetVarAddrExposed(unsigned varNum);
2431 bool lvaVarDoNotEnregister(unsigned varNum);
2433 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2434 enum DoNotEnregisterReason
2439 DNER_VMNeedsStackAddr,
2440 DNER_LiveInOutOfHandler,
2441 DNER_LiveAcrossUnmanagedCall,
2442 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2443 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2444 DNER_DepField, // It is a field of a dependently promoted struct
2445 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2446 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2447 #ifdef JIT32_GCENCODER
2452 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2454 unsigned lvaVarargsHandleArg;
2456 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2458 #endif // _TARGET_X86_
2460 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2461 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2462 #if FEATURE_FIXED_OUT_ARGS
2463 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2465 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2466 // that tracks whether the lock has been taken
2468 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2469 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2470 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2472 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2473 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2475 #if FEATURE_FIXED_OUT_ARGS
2476 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2477 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2478 #endif // FEATURE_FIXED_OUT_ARGS
2481 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2482 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2483 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2484 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2485 // this variable to be this scratch word whenever struct promotion occurs.
2486 unsigned lvaPromotedStructAssemblyScratchVar;
2487 #endif // _TARGET_ARM_
2490 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2491 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2494 unsigned lvaGenericsContextUseCount;
2496 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2497 // CORINFO_GENERICS_CTXT_FROM_THIS?
2498 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2500 //-------------------------------------------------------------------------
2501 // All these frame offsets are inter-related and must be kept in sync
2503 #if !FEATURE_EH_FUNCLETS
2504 // This is used for the callable handlers
2505 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2506 #endif // FEATURE_EH_FUNCLETS
2508 unsigned lvaCachedGenericContextArgOffs;
2509 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2512 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2514 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2516 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2517 // after the reg predict we will use a computed maxTmpSize
2518 // which is based upon the number of spill temps predicted by reg predict
2519 // All this is necessary because if we under-estimate the size of the spill
2520 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2522 // Pre codegen max spill temp size.
2523 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2525 //-------------------------------------------------------------------------
2527 unsigned lvaGetMaxSpillTempSize();
2529 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2530 #endif // _TARGET_ARM_
2531 void lvaAssignFrameOffsets(FrameLayoutState curState);
2532 void lvaFixVirtualFrameOffsets();
2534 #ifndef LEGACY_BACKEND
2535 void lvaUpdateArgsWithInitialReg();
2536 #endif // !LEGACY_BACKEND
2538 void lvaAssignVirtualFrameOffsetsToArgs();
2539 #ifdef UNIX_AMD64_ABI
2540 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2541 #else // !UNIX_AMD64_ABI
2542 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2543 #endif // !UNIX_AMD64_ABI
2544 void lvaAssignVirtualFrameOffsetsToLocals();
2545 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2546 #ifdef _TARGET_AMD64_
2547 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2548 bool lvaIsCalleeSavedIntRegCountEven();
2550 void lvaAlignFrame();
2551 void lvaAssignFrameOffsetsToPromotedStructs();
2552 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2555 void lvaDumpRegLocation(unsigned lclNum);
2556 void lvaDumpFrameLocation(unsigned lclNum);
2557 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2558 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2559 // layout state defined by lvaDoneFrameLayout
2562 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2563 // to avoid bugs from borderline cases.
2564 #define MAX_FrameSize 0x3FFFFFFF
2565 void lvaIncrementFrameSize(unsigned size);
2567 unsigned lvaFrameSize(FrameLayoutState curState);
2569 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2570 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2572 // Returns the caller-SP-relative offset for the local variable "varNum."
2573 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2575 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2576 int lvaGetSPRelativeOffset(unsigned varNum);
2578 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2579 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2581 //------------------------ For splitting types ----------------------------
2583 void lvaInitTypeRef();
2585 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2586 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2587 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2588 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2589 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2590 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2592 void lvaInitVarDsc(LclVarDsc* varDsc,
2594 CorInfoType corInfoType,
2595 CORINFO_CLASS_HANDLE typeHnd,
2596 CORINFO_ARG_LIST_HANDLE varList,
2597 CORINFO_SIG_INFO* varSig);
2599 static unsigned lvaTypeRefMask(var_types type);
2601 var_types lvaGetActualType(unsigned lclNum);
2602 var_types lvaGetRealType(unsigned lclNum);
2604 //-------------------------------------------------------------------------
2608 unsigned lvaLclSize(unsigned varNum);
2609 unsigned lvaLclExactSize(unsigned varNum);
2611 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2613 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2614 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2615 // the return result.
2616 bool lvaLclVarRefsAccum(
2617 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2619 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2620 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2621 // and (destructively) unions "trkedVars" into "*result".
2622 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2624 ALLVARSET_VALARG_TP allVars,
2625 VARSET_VALARG_TP trkdVars);
2627 bool lvaHaveManyLocals() const;
2629 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2630 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2631 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2634 void lvaSortByRefCount();
2635 void lvaDumpRefCounts();
2637 void lvaMarkLocalVars(BasicBlock* block);
2639 void lvaMarkLocalVars(); // Local variable ref-counting
2641 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2643 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2645 void lvaIncRefCnts(GenTreePtr tree);
2646 void lvaDecRefCnts(GenTreePtr tree);
2648 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2649 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2650 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2653 struct lvaStressLclFldArgs
2655 Compiler* m_pCompiler;
2659 static fgWalkPreFn lvaStressLclFldCB;
2660 void lvaStressLclFld();
2662 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2663 void lvaDispVarSet(VARSET_VALARG_TP set);
2668 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2670 int lvaFrameAddress(int varNum, bool* pFPbased);
2673 bool lvaIsParameter(unsigned varNum);
2674 bool lvaIsRegArgument(unsigned varNum);
2675 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2676 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2677 // that writes to arg0
2679 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2680 // (this is an overload of lvIsTemp because there are no temp parameters).
2681 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2682 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2683 bool lvaIsImplicitByRefLocal(unsigned varNum)
2685 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2686 LclVarDsc* varDsc = &(lvaTable[varNum]);
2687 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2689 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2692 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2696 // Returns true if this local var is a multireg struct
2697 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2699 // If the local is a TYP_STRUCT, get/set a class handle describing it
2700 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2701 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2703 // If the local is TYP_REF, set or update the associated class information.
2704 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2705 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2706 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2707 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2709 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2711 // Info about struct fields
2712 struct lvaStructFieldInfo
2714 CORINFO_FIELD_HANDLE fldHnd;
2715 unsigned char fldOffset;
2716 unsigned char fldOrdinal;
2719 CORINFO_CLASS_HANDLE fldTypeHnd;
2722 // Info about struct to be promoted.
2723 struct lvaStructPromotionInfo
2725 CORINFO_CLASS_HANDLE typeHnd;
2727 bool requiresScratchVar;
2730 unsigned char fieldCnt;
2731 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2733 lvaStructPromotionInfo()
2734 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2739 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2740 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2741 lvaStructPromotionInfo* StructPromotionInfo,
2743 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2744 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2745 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2746 #if !defined(_TARGET_64BIT_)
2747 void lvaPromoteLongVars();
2748 #endif // !defined(_TARGET_64BIT_)
2749 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2750 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2751 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2752 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2753 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2754 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2755 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2757 #if defined(FEATURE_SIMD)
2758 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2760 assert(varDsc->lvType == TYP_SIMD12);
2761 assert(varDsc->lvExactSize == 12);
2763 #if defined(_TARGET_64BIT_)
2764 assert(varDsc->lvSize() == 16);
2765 #endif // defined(_TARGET_64BIT_)
2767 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2768 // already does this calculation. However, we also need to prevent mapping types if the var is a
2769 // dependently promoted struct field, which must remain its exact size within its parent struct.
2770 // However, we don't know this until late, so we may have already pretended the field is bigger
2772 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2781 #endif // defined(FEATURE_SIMD)
2783 BYTE* lvaGetGcLayout(unsigned varNum);
2784 bool lvaTypeIsGC(unsigned varNum);
2785 unsigned lvaGSSecurityCookie; // LclVar number
2786 bool lvaTempsHaveLargerOffsetThanVars();
2788 unsigned lvaSecurityObject; // variable representing the security object on the stack
2789 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2791 #if FEATURE_EH_FUNCLETS
2792 unsigned lvaPSPSym; // variable representing the PSPSym
2795 InlineInfo* impInlineInfo;
2796 InlineStrategy* m_inlineStrategy;
2798 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2799 Compiler* impInlineRoot();
2801 #if defined(DEBUG) || defined(INLINE_DATA)
2802 unsigned __int64 getInlineCycleCount()
2804 return m_compCycles;
2806 #endif // defined(DEBUG) || defined(INLINE_DATA)
2808 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2809 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2811 //=========================================================================
2813 //=========================================================================
2816 //---------------- Local variable ref-counting ----------------------------
2819 BasicBlock* lvaMarkRefsCurBlock;
2820 GenTreePtr lvaMarkRefsCurStmt;
2822 BasicBlock::weight_t lvaMarkRefsWeight;
2824 void lvaMarkLclRefs(GenTreePtr tree);
2826 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2827 void SetVolatileHint(LclVarDsc* varDsc);
2829 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2830 PerSsaArray lvMemoryPerSsaData;
2831 unsigned lvMemoryNumSsaNames;
2834 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2835 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2836 // not an SSA variable).
2837 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2839 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2840 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2842 assert(ssaNum < lvMemoryNumSsaNames);
2843 return &lvMemoryPerSsaData.GetRef(ssaNum);
2847 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2848 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2852 XX Imports the given method and converts it to semantic trees XX
2854 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2855 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2861 void impImport(BasicBlock* method);
2863 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2864 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2865 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2866 CORINFO_CLASS_HANDLE impGetStringClass();
2867 CORINFO_CLASS_HANDLE impGetObjectClass();
2869 //=========================================================================
2871 //=========================================================================
2874 //-------------------- Stack manipulation ---------------------------------
2876 unsigned impStkSize; // Size of the full stack
2878 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2880 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2882 struct SavedStack // used to save/restore stack contents.
2884 unsigned ssDepth; // number of values on stack
2885 StackEntry* ssTrees; // saved tree values
2888 bool impIsPrimitive(CorInfoType type);
2889 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2891 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2893 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2894 void impPushNullObjRefOnStack();
2895 StackEntry impPopStack();
2896 StackEntry& impStackTop(unsigned n = 0);
2897 unsigned impStackHeight();
2899 void impSaveStackState(SavedStack* savePtr, bool copy);
2900 void impRestoreStackState(SavedStack* savePtr);
2902 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2903 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2904 CORINFO_CALL_INFO* pCallInfo);
2906 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2908 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2910 bool impCanPInvokeInline();
2911 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2912 void impCheckForPInvokeCall(
2913 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2914 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2915 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2917 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2918 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2919 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2921 var_types impImportCall(OPCODE opcode,
2922 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2923 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2925 GenTreePtr newobjThis,
2927 CORINFO_CALL_INFO* callInfo,
2928 IL_OFFSET rawILOffset);
2930 void impDevirtualizeCall(GenTreeCall* call,
2932 CORINFO_METHOD_HANDLE* method,
2933 unsigned* methodFlags,
2934 CORINFO_CONTEXT_HANDLE* contextHandle,
2935 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2937 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2939 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2941 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2944 var_types impImportJitTestLabelMark(int numArgs);
2947 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2949 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2951 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2952 CORINFO_ACCESS_FLAGS access,
2953 CORINFO_FIELD_INFO* pFieldInfo,
2956 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2958 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2960 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2962 void impImportLeave(BasicBlock* block);
2963 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2964 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2965 CORINFO_CLASS_HANDLE clsHnd,
2966 CORINFO_METHOD_HANDLE method,
2967 CORINFO_SIG_INFO* sig,
2971 CorInfoIntrinsics* pIntrinsicID);
2972 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2973 CORINFO_SIG_INFO* sig,
2976 CorInfoIntrinsics intrinsicID);
2977 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2979 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2981 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2982 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2983 CORINFO_THIS_TRANSFORM transform);
2985 //----------------- Manipulating the trees and stmts ----------------------
2987 GenTreePtr impTreeList; // Trees for the BB being imported
2988 GenTreePtr impTreeLast; // The last tree for the current BB
2992 CHECK_SPILL_ALL = -1,
2993 CHECK_SPILL_NONE = -2
2997 void impBeginTreeList();
2998 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2999 void impEndTreeList(BasicBlock* block);
3000 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3001 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3002 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3003 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3004 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3005 void impAssignTempGen(unsigned tmp,
3008 GenTreePtr* pAfterStmt = nullptr,
3009 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3010 BasicBlock* block = nullptr);
3011 void impAssignTempGen(unsigned tmpNum,
3013 CORINFO_CLASS_HANDLE structHnd,
3015 GenTreePtr* pAfterStmt = nullptr,
3016 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3017 BasicBlock* block = nullptr);
3018 GenTreePtr impCloneExpr(GenTreePtr tree,
3020 CORINFO_CLASS_HANDLE structHnd,
3022 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3023 GenTreePtr impAssignStruct(GenTreePtr dest,
3025 CORINFO_CLASS_HANDLE structHnd,
3027 GenTreePtr* pAfterStmt = nullptr,
3028 BasicBlock* block = nullptr);
3029 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3031 CORINFO_CLASS_HANDLE structHnd,
3033 GenTreePtr* pAfterStmt = nullptr,
3034 BasicBlock* block = nullptr);
3036 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3037 CORINFO_CLASS_HANDLE structHnd,
3041 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3042 BYTE* gcLayout = nullptr,
3043 unsigned* numGCVars = nullptr,
3044 var_types* simdBaseType = nullptr);
3046 GenTreePtr impNormStructVal(GenTreePtr structVal,
3047 CORINFO_CLASS_HANDLE structHnd,
3049 bool forceNormalization = false);
3051 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3052 BOOL* pRuntimeLookup = nullptr,
3053 BOOL mustRestoreHandle = FALSE,
3054 BOOL importParent = FALSE);
3056 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3057 BOOL* pRuntimeLookup = nullptr,
3058 BOOL mustRestoreHandle = FALSE)
3060 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3063 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3064 CORINFO_LOOKUP* pLookup,
3066 void* compileTimeHandle);
3068 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3070 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3071 CORINFO_LOOKUP* pLookup,
3072 void* compileTimeHandle);
3074 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3076 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3077 CorInfoHelpFunc helper,
3079 GenTreeArgList* arg = nullptr,
3080 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3082 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3084 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3087 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3088 CORINFO_CLASS_HANDLE typeClass,
3092 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3093 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3094 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3095 static bool IsMathIntrinsic(GenTreePtr tree);
3098 //----------------- Importing the method ----------------------------------
3100 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3103 unsigned impCurOpcOffs;
3104 const char* impCurOpcName;
3105 bool impNestedStackSpill;
3107 // For displaying instrs with generated native code (-n:B)
3108 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3109 void impNoteLastILoffs();
3112 /* IL offset of the stmt currently being imported. It gets set to
3113 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3114 updated at IL offsets for which we have to report mapping info.
3115 It also includes flag bits, so use jitGetILoffs()
3116 to get the actual IL offset value.
3119 IL_OFFSETX impCurStmtOffs;
3120 void impCurStmtOffsSet(IL_OFFSET offs);
3122 void impNoteBranchOffs();
3124 unsigned impInitBlockLineInfo();
3126 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3127 bool impIsThis(GenTreePtr obj);
3128 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3129 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3130 bool impIsAnySTLOC(OPCODE opcode)
3132 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3133 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3136 GenTreeArgList* impPopList(unsigned count,
3138 CORINFO_SIG_INFO* sig,
3139 GenTreeArgList* prefixTree = nullptr);
3141 GenTreeArgList* impPopRevList(unsigned count,
3143 CORINFO_SIG_INFO* sig,
3144 unsigned skipReverseCount = 0);
3147 * Get current IL offset with stack-empty info incoporated
3149 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3151 //---------------- Spilling the importer stack ----------------------------
3153 // The maximum number of bytes of IL processed without clean stack state.
3154 // It allows to limit the maximum tree size and depth.
3155 static const unsigned MAX_TREE_SIZE = 200;
3156 bool impCanSpillNow(OPCODE prevOpcode);
3162 SavedStack pdSavedStack;
3163 ThisInitState pdThisPtrInit;
3166 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3167 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3169 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3170 ExpandArray<BYTE> impPendingBlockMembers;
3172 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3173 // Operates on the map in the top-level ancestor.
3174 BYTE impGetPendingBlockMember(BasicBlock* blk)
3176 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3179 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3180 // Operates on the map in the top-level ancestor.
3181 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3183 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3186 bool impCanReimport;
3188 bool impSpillStackEntry(unsigned level,
3192 bool bAssertOnRecursion,
3197 void impSpillStackEnsure(bool spillLeaves = false);
3198 void impEvalSideEffects();
3199 void impSpillSpecialSideEff();
3200 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3201 void impSpillValueClasses();
3202 void impSpillEvalStack();
3203 static fgWalkPreFn impFindValueClasses;
3204 void impSpillLclRefs(ssize_t lclNum);
3206 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3208 void impImportBlockCode(BasicBlock* block);
3210 void impReimportMarkBlock(BasicBlock* block);
3211 void impReimportMarkSuccessors(BasicBlock* block);
3213 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3215 void impImportBlockPending(BasicBlock* block);
3217 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3218 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3219 // for the block, but instead, just re-uses the block's existing EntryState.
3220 void impReimportBlockPending(BasicBlock* block);
3222 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3224 void impImportBlock(BasicBlock* block);
3226 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3227 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3228 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3229 // the variables that will be used -- and for all the predecessors of those successors, and the
3230 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3231 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3232 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3233 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3234 // of local variable numbers, so we represent them with the base local variable number), returns that.
3235 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3236 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3237 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3238 // on which kind of member of the clique the block is).
3239 unsigned impGetSpillTmpBase(BasicBlock* block);
3241 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3242 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3243 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3244 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3245 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3246 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3247 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3248 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3249 // successors receive a native int. Similarly float and double are unified to double.
3250 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3251 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3252 // predecessors, so they insert an upcast if needed).
3253 void impReimportSpillClique(BasicBlock* block);
3255 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3256 // block, and represent the predecessor and successor members of the clique currently being computed.
3257 // *** Access to these will need to be locked in a parallel compiler.
3258 ExpandArray<BYTE> impSpillCliquePredMembers;
3259 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3267 // Abstract class for receiving a callback while walking a spill clique
3268 class SpillCliqueWalker
3271 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3274 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3275 class SetSpillTempsBase : public SpillCliqueWalker
3280 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3283 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3286 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3287 class ReimportSpillClique : public SpillCliqueWalker
3292 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3295 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3298 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3299 // predecessor or successor within the spill clique
3300 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3302 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3303 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3304 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3305 void impRetypeEntryStateTemps(BasicBlock* blk);
3307 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3308 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3310 void impPushVar(GenTree* op, typeInfo tiRetVal);
3311 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3312 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3314 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3316 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3317 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3318 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3321 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3324 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3325 struct BlockListNode
3328 BlockListNode* m_next;
3329 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3332 void* operator new(size_t sz, Compiler* comp);
3334 BlockListNode* impBlockListNodeFreeList;
3336 BlockListNode* AllocBlockListNode();
3337 void FreeBlockListNode(BlockListNode* node);
3339 bool impIsValueType(typeInfo* pTypeInfo);
3340 var_types mangleVarArgsType(var_types type);
3343 regNumber getCallArgIntRegister(regNumber floatReg);
3344 regNumber getCallArgFloatRegister(regNumber intReg);
3345 #endif // FEATURE_VARARG
3348 static unsigned jitTotalMethodCompiled;
3352 static LONG jitNestingLevel;
3355 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3357 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3359 // STATIC inlining decision based on the IL code.
3360 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3361 CORINFO_METHOD_INFO* methInfo,
3363 InlineResult* inlineResult);
3365 void impCheckCanInline(GenTreePtr call,
3366 CORINFO_METHOD_HANDLE fncHandle,
3368 CORINFO_CONTEXT_HANDLE exactContextHnd,
3369 InlineCandidateInfo** ppInlineCandidateInfo,
3370 InlineResult* inlineResult);
3372 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3373 GenTreePtr curArgVal,
3375 InlineResult* inlineResult);
3377 void impInlineInitVars(InlineInfo* pInlineInfo);
3379 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3381 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3383 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3385 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3386 GenTreePtr variableBeingDereferenced,
3387 InlArgInfo* inlArgInfo);
3389 void impMarkInlineCandidate(GenTreePtr call,
3390 CORINFO_CONTEXT_HANDLE exactContextHnd,
3391 bool exactContextNeedsRuntimeLookup,
3392 CORINFO_CALL_INFO* callInfo);
3394 bool impTailCallRetTypeCompatible(var_types callerRetType,
3395 CORINFO_CLASS_HANDLE callerRetTypeClass,
3396 var_types calleeRetType,
3397 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3399 bool impIsTailCallILPattern(bool tailPrefixed,
3401 const BYTE* codeAddrOfNextOpcode,
3402 const BYTE* codeEnd,
3404 bool* IsCallPopRet = nullptr);
3406 bool impIsImplicitTailCallCandidate(
3407 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3409 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3412 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3413 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3417 XX Info about the basic-blocks, their contents and the flow analysis XX
3419 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3420 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3424 BasicBlock* fgFirstBB; // Beginning of the basic block list
3425 BasicBlock* fgLastBB; // End of the basic block list
3426 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3427 #if FEATURE_EH_FUNCLETS
3428 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3430 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3432 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3433 unsigned fgEdgeCount; // # of control flow edges between the BBs
3434 unsigned fgBBcount; // # of BBs in the method
3436 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3438 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3439 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3440 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3441 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3443 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3444 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3445 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3446 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3447 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3448 // index). The arrays are of size fgBBNumMax + 1.
3449 unsigned* fgDomTreePreOrder;
3450 unsigned* fgDomTreePostOrder;
3452 bool fgBBVarSetsInited;
3454 // Allocate array like T* a = new T[fgBBNumMax + 1];
3455 // Using helper so we don't keep forgetting +1.
3456 template <typename T>
3457 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3459 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3462 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3463 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3464 // cannot be meaningfully combined. Note that new blocks can be created with higher
3465 // block numbers without changing the basic block epoch. These blocks *cannot*
3466 // participate in a block set until the blocks are all renumbered, causing the epoch
3467 // to change. This is useful if continuing to use previous block sets is valuable.
3468 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3469 unsigned fgCurBBEpoch;
3471 unsigned GetCurBasicBlockEpoch()
3473 return fgCurBBEpoch;
3476 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3477 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3478 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3479 unsigned fgCurBBEpochSize;
3481 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3482 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3483 unsigned fgBBSetCountInSizeTUnits;
3485 void NewBasicBlockEpoch()
3487 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3489 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3491 fgCurBBEpochSize = fgBBNumMax + 1;
3492 fgBBSetCountInSizeTUnits =
3493 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3496 // All BlockSet objects are now invalid!
3497 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3498 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3502 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3503 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3504 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3505 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3507 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3508 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3509 // array of size_t bitsets), then print that out.
3510 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3517 void EnsureBasicBlockEpoch()
3519 if (fgCurBBEpochSize != fgBBNumMax + 1)
3521 NewBasicBlockEpoch();
3525 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3526 void fgEnsureFirstBBisScratch();
3527 bool fgFirstBBisScratch();
3528 bool fgBBisScratch(BasicBlock* block);
3530 void fgExtendEHRegionBefore(BasicBlock* block);
3531 void fgExtendEHRegionAfter(BasicBlock* block);
3533 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3535 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3537 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3540 BasicBlock* nearBlk,
3541 bool putInFilter = false,
3542 bool runRarely = false,
3543 bool insertAtEnd = false);
3545 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3547 bool runRarely = false,
3548 bool insertAtEnd = false);
3550 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3552 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3553 BasicBlock* afterBlk,
3554 unsigned xcptnIndex,
3555 bool putInTryRegion);
3557 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3558 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3559 void fgUnlinkBlock(BasicBlock* block);
3561 unsigned fgMeasureIR();
3563 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3564 bool fgMultipleNots;
3567 bool fgModified; // True if the flow graph has been modified recently
3568 bool fgComputePredsDone; // Have we computed the bbPreds list
3569 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3570 bool fgDomsComputed; // Have we computed the dominator sets?
3571 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3573 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3574 bool fgHasPostfix; // any postfix ++/-- found?
3575 unsigned fgIncrCount; // number of increment nodes found
3577 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3581 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3582 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3585 bool fgRemoveRestOfBlock; // true if we know that we will throw
3586 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3588 // There are two modes for ordering of the trees.
3589 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3590 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3591 // by traversing the tree according to the order of the operands.
3592 // - In FGOrderLinear, the dominant ordering is the linear order.
3599 FlowGraphOrder fgOrder;
3601 // The following are boolean flags that keep track of the state of internal data structures
3603 bool fgStmtListThreaded;
3604 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3605 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3606 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3607 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3608 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3609 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3610 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3611 // This is derived from the profile data
3612 // or is BB_UNITY_WEIGHT when we don't have profile data
3614 #if FEATURE_EH_FUNCLETS
3615 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3616 #endif // FEATURE_EH_FUNCLETS
3618 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3619 // since fgMorphTree can be called from several places
3621 bool impBoxTempInUse; // the temp below is valid and available
3622 unsigned impBoxTemp; // a temporary that is used for boxing
3625 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3626 // and we are trying to compile again in a "safer", minopts mode?
3630 unsigned impInlinedCodeSize;
3633 //-------------------------------------------------------------------------
3639 void fgTransformFatCalli();
3643 void fgRemoveEmptyTry();
3645 void fgRemoveEmptyFinally();
3647 void fgMergeFinallyChains();
3649 void fgCloneFinally();
3651 void fgCleanupContinuation(BasicBlock* continuation);
3653 void fgUpdateFinallyTargetFlags();
3655 void fgClearAllFinallyTargetBits();
3657 void fgAddFinallyTargetFlags();
3659 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3660 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3661 // when this is necessary.
3662 bool fgNeedToAddFinallyTargetBits;
3663 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3665 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3666 BasicBlock* handler,
3667 BlockToBlockMap& continuationMap);
3669 GenTreePtr fgGetCritSectOfStaticMethod();
3671 #if FEATURE_EH_FUNCLETS
3673 void fgAddSyncMethodEnterExit();
3675 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3677 void fgConvertSyncReturnToLeave(BasicBlock* block);
3679 #endif // FEATURE_EH_FUNCLETS
3681 void fgAddReversePInvokeEnterExit();
3683 bool fgMoreThanOneReturnBlock();
3685 // The number of separate return points in the method.
3686 unsigned fgReturnCount;
3688 void fgAddInternal();
3690 bool fgFoldConditional(BasicBlock* block);
3692 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3693 void fgMorphBlocks();
3695 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3697 void fgCheckArgCnt();
3698 void fgSetOptions();
3701 static fgWalkPreFn fgAssertNoQmark;
3702 void fgPreExpandQmarkChecks(GenTreePtr expr);
3703 void fgPostExpandQmarkChecks();
3704 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3707 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3709 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3710 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3711 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3712 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3713 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3715 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3716 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3717 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3718 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3720 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3721 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3722 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3723 void fgExpandQmarkNodes();
3727 // Do "simple lowering." This functionality is (conceptually) part of "general"
3728 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3729 void fgSimpleLowering();
3731 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3733 GenTreePtr fgInitThisClass();
3735 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3737 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3739 inline bool backendRequiresLocalVarLifetimes()
3741 #if defined(LEGACY_BACKEND)
3744 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3748 void fgLocalVarLiveness();
3750 void fgLocalVarLivenessInit();
3752 #ifdef LEGACY_BACKEND
3753 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3755 void fgPerNodeLocalVarLiveness(GenTree* node);
3757 void fgPerBlockLocalVarLiveness();
3759 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3761 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3763 // This is used in the liveness computation, as a temporary. When we use the
3764 // arbitrary-length VarSet representation, it is better not to allocate a new one
3766 VARSET_TP fgMarkIntfUnionVS;
3768 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3770 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3772 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3774 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3776 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3778 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3780 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3782 void fgComputeLife(VARSET_TP& life,
3783 GenTreePtr startNode,
3785 VARSET_VALARG_TP volatileVars,
3786 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3788 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3790 bool fgRemoveDeadStore(GenTree** pTree,
3792 VARSET_VALARG_TP life,
3794 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3796 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3798 // For updating liveset during traversal AFTER fgComputeLife has completed
3799 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3800 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3802 // Returns the set of live variables after endTree,
3803 // assuming that liveSet is the set of live variables BEFORE tree.
3804 // Requires that fgComputeLife has completed, and that tree is in the same
3805 // statement as endTree, and that it comes before endTree in execution order
3807 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3809 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3810 while (tree != nullptr && tree != endTree->gtNext)
3812 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3813 tree = tree->gtNext;
3815 assert(tree == endTree->gtNext);
3819 void fgInterBlockLocalVarLiveness();
3821 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3822 // "x", and a def of a new SSA name for "x". The tree only has one local variable for "x", so it has to choose
3823 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3824 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3825 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3826 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3827 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3829 if (m_opAsgnVarDefSsaNums == nullptr)
3831 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3833 return m_opAsgnVarDefSsaNums;
3836 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3837 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3838 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3840 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3842 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3843 // Except: assumes that lcl is a def, and if it is
3844 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3845 // rather than the "use" SSA number recorded in the tree "lcl".
3846 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3848 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3849 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3850 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3851 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3852 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3854 // (byref addrS1 = &s1,
3855 // *(addrS1 * offsetof(f0)) = s2f0,
3857 // *(addrS1 * offsetof(fn)) = s2fn)
3859 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3860 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3861 // give it SSA names and value numbers?
3863 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3864 // end with an instance of the structure below, whose fields are described in the declaration.
3865 struct IndirectAssignmentAnnotation
3867 unsigned m_lclNum; // The local num that is being indirectly assigned.
3868 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3869 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3870 // be the singleton field sequence "g". The individual assignments would
3871 // further append the fields of "s.g" to that.
3872 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3873 // structure has a single field).
3874 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3875 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3878 IndirectAssignmentAnnotation(unsigned lclNum,
3879 FieldSeqNode* fldSeq,
3881 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3882 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3883 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3887 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3888 NodeToIndirAssignMap;
3889 NodeToIndirAssignMap* m_indirAssignMap;
3890 NodeToIndirAssignMap* GetIndirAssignMap()
3892 if (m_indirAssignMap == nullptr)
3894 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3895 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3896 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3898 return m_indirAssignMap;
3901 // Performs SSA conversion.
3904 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3905 void fgResetForSsa();
3907 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3909 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3910 inline bool fgExcludeFromSsa(unsigned lclNum);
3912 // The value numbers for this compilation.
3913 ValueNumStore* vnStore;
3916 ValueNumStore* GetValueNumStore()
3921 // Do value numbering (assign a value number to each
3923 void fgValueNumber();
3925 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3926 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3927 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3928 // match the element type of the array or fldSeq. When this type doesn't match
3929 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3931 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3934 FieldSeqNode* fldSeq,
3938 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3939 // has been parsed to yield the other input arguments. If evaluation of the address
3940 // can raise exceptions, those should be captured in the exception set "excVN."
3941 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3942 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3943 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3944 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3945 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3947 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3948 CORINFO_CLASS_HANDLE elemTypeEq,
3952 FieldSeqNode* fldSeq);
3954 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3955 // by evaluating the array index expression "tree". Returns the value number resulting from
3956 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3957 // "GT_IND" that does the dereference, and it is given the returned value number.
3958 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3960 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3961 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3963 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3965 // Utility functions for fgValueNumber.
3967 // Perform value-numbering for the trees in "blk".
3968 void fgValueNumberBlock(BasicBlock* blk);
3970 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3971 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3972 // assumed for the memoryKind at the start "entryBlk".
3973 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3975 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3976 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3977 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3979 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3981 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3983 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3984 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3985 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3987 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3988 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3990 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3991 // value in that SSA #.
3992 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3994 // The input 'tree' is a leaf node that is a constant
3995 // Assign the proper value number to the tree
3996 void fgValueNumberTreeConst(GenTreePtr tree);
3998 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3999 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4001 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4003 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
4005 // Does value-numbering for a block assignment.
4006 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4008 // Does value-numbering for a cast tree.
4009 void fgValueNumberCastTree(GenTreePtr tree);
4011 // Does value-numbering for an intrinsic tree.
4012 void fgValueNumberIntrinsic(GenTreePtr tree);
4014 // Does value-numbering for a call. We interpret some helper calls.
4015 void fgValueNumberCall(GenTreeCall* call);
4017 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4018 void fgUpdateArgListVNs(GenTreeArgList* args);
4020 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4021 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4023 // Requires "helpCall" to be a helper call. Assigns it a value number;
4024 // we understand the semantics of some of the calls. Returns "true" if
4025 // the call may modify the heap (we assume arbitrary memory side effects if so).
4026 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4028 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4029 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4031 // These are the current value number for the memory implicit variables while
4032 // doing value numbering. These are the value numbers under the "liberal" interpretation
4033 // of memory values; the "conservative" interpretation needs no VN, since every access of
4034 // memory yields an unknown value.
4035 ValueNum fgCurMemoryVN[MemoryKindCount];
4037 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4038 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4039 // is 1, and the rest is an encoding of "elemTyp".
4040 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4042 if (elemStructType != nullptr)
4044 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4045 varTypeIsIntegral(elemTyp));
4046 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4047 return elemStructType;
4051 elemTyp = varTypeUnsignedToSigned(elemTyp);
4052 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4055 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4056 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4057 // the struct type of the element).
4058 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4060 size_t clsHndVal = size_t(clsHnd);
4061 if (clsHndVal & 0x1)
4063 return var_types(clsHndVal >> 1);
4071 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4072 var_types getJitGCType(BYTE gcType);
4074 enum structPassingKind
4076 SPK_Unknown, // Invalid value, never returned
4077 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4078 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4079 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4080 // parameters registers are used, then the stack will be used)
4081 // for X86 passed on the stack, for ARM32 passed in registers
4082 // or the stack or split between registers and the stack.
4083 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4085 }; // The struct is passed/returned by reference to a copy/buffer.
4087 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4088 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4089 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4090 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4092 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4094 // Get the type that is used to pass values of the given struct type.
4095 // If you have already retrieved the struct size then pass it as the optional third argument
4097 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4098 structPassingKind* wbPassStruct,
4099 unsigned structSize = 0);
4101 // Get the type that is used to return values of the given struct type.
4102 // If you have already retrieved the struct size then pass it as the optional third argument
4104 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4105 structPassingKind* wbPassStruct = nullptr,
4106 unsigned structSize = 0);
4109 // Print a representation of "vnp" or "vn" on standard output.
4110 // If "level" is non-zero, we also print out a partial expansion of the value.
4111 void vnpPrint(ValueNumPair vnp, unsigned level);
4112 void vnPrint(ValueNum vn, unsigned level);
4115 // Dominator computation member functions
4116 // Not exposed outside Compiler
4118 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4120 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4122 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4123 // flow graph. We first assume the fields bbIDom on each
4124 // basic block are invalid. This computation is needed later
4125 // by fgBuildDomTree to build the dominance tree structure.
4126 // Based on: A Simple, Fast Dominance Algorithm
4127 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4129 void fgCompDominatedByExceptionalEntryBlocks();
4131 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4132 // Note: this is relatively slow compared to calling fgDominate(),
4133 // especially if dealing with a single block versus block check.
4135 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4137 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4139 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4141 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4143 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4145 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4146 // processed in topological sort, this function takes care of that.
4148 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4150 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4151 // Returns this as a set.
4153 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4154 // root nodes. Returns this as a set.
4157 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4160 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4161 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4164 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4165 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4166 // && postOrder(A) >= postOrder(B) making the computation O(1).
4167 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4169 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4171 void fgUpdateChangedFlowGraph();
4174 // Compute the predecessors of the blocks in the control flow graph.
4175 void fgComputePreds();
4177 // Remove all predecessor information.
4178 void fgRemovePreds();
4180 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4181 // before the full predecessors lists are computed.
4182 void fgComputeCheapPreds();
4185 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4187 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4197 // Initialize the per-block variable sets (used for liveness analysis).
4198 void fgInitBlockVarSets();
4200 // true if we've gone through and created GC Poll calls.
4201 bool fgGCPollsCreated;
4202 void fgMarkGCPollBlocks();
4203 void fgCreateGCPolls();
4204 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4206 // Requires that "block" is a block that returns from
4207 // a finally. Returns the number of successors (jump targets of
4208 // of blocks in the covered "try" that did a "LEAVE".)
4209 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4211 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4212 // a finally. Returns its "i"th successor (jump targets of
4213 // of blocks in the covered "try" that did a "LEAVE".)
4214 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4215 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4218 // Factor out common portions of the impls of the methods above.
4219 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4222 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4223 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4224 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4225 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4226 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4227 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4228 // we leave the entry associated with the block, but it will no longer be accessed.)
4229 struct SwitchUniqueSuccSet
4231 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4232 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4235 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4236 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4237 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4238 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4241 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4242 BlockToSwitchDescMap;
4245 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4246 // iteration over only the distinct successors.
4247 BlockToSwitchDescMap* m_switchDescMap;
4250 BlockToSwitchDescMap* GetSwitchDescMap()
4252 if (m_switchDescMap == nullptr)
4254 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4256 return m_switchDescMap;
4259 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4260 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4261 // we don't accidentally look up and return the wrong switch data.
4262 void InvalidateUniqueSwitchSuccMap()
4264 m_switchDescMap = nullptr;
4267 // Requires "switchBlock" to be a block that ends in a switch. Returns
4268 // the corresponding SwitchUniqueSuccSet.
4269 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4271 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4272 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4273 // remove it from "this", and ensure that "to" is a member.
4274 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4276 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4277 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4279 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4281 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4283 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4285 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4287 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4289 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4291 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4293 void fgRemoveBlockAsPred(BasicBlock* block);
4295 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4297 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4299 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4301 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4303 flowList* fgAddRefPred(BasicBlock* block,
4304 BasicBlock* blockPred,
4305 flowList* oldEdge = nullptr,
4306 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4309 void fgFindBasicBlocks();
4311 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4313 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4315 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4316 bool putInTryRegion,
4317 BasicBlock* startBlk,
4319 BasicBlock* nearBlk,
4320 BasicBlock* jumpBlk,
4323 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4325 void fgRemoveEmptyBlocks();
4327 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4329 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4331 void fgCreateLoopPreHeader(unsigned lnum);
4333 void fgUnreachableBlock(BasicBlock* block);
4335 void fgRemoveConditionalJump(BasicBlock* block);
4337 BasicBlock* fgLastBBInMainFunction();
4339 BasicBlock* fgEndBBAfterMainFunction();
4341 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4343 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4345 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4347 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4349 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4351 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4353 bool fgRenumberBlocks();
4355 bool fgExpandRarelyRunBlocks();
4357 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4359 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4361 enum FG_RELOCATE_TYPE
4363 FG_RELOCATE_TRY, // relocate the 'try' region
4364 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4366 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4368 #if FEATURE_EH_FUNCLETS
4369 #if defined(_TARGET_ARM_)
4370 void fgClearFinallyTargetBit(BasicBlock* block);
4371 #endif // defined(_TARGET_ARM_)
4372 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4373 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4374 void fgInsertFuncletPrologBlock(BasicBlock* block);
4375 void fgCreateFuncletPrologBlocks();
4376 void fgCreateFunclets();
4377 #else // !FEATURE_EH_FUNCLETS
4378 bool fgRelocateEHRegions();
4379 #endif // !FEATURE_EH_FUNCLETS
4381 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4383 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4385 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4387 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4389 bool fgOptimizeEmptyBlock(BasicBlock* block);
4391 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4393 bool fgOptimizeBranch(BasicBlock* bJump);
4395 bool fgOptimizeSwitchBranches(BasicBlock* block);
4397 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4399 bool fgOptimizeSwitchJumps();
4401 void fgPrintEdgeWeights();
4403 void fgComputeEdgeWeights();
4405 void fgReorderBlocks();
4407 void fgDetermineFirstColdBlock();
4409 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4411 bool fgUpdateFlowGraph(bool doTailDup = false);
4413 void fgFindOperOrder();
4415 // method that returns if you should split here
4416 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4418 void fgSetBlockOrder();
4420 void fgRemoveReturnBlock(BasicBlock* block);
4422 /* Helper code that has been factored out */
4423 inline void fgConvertBBToThrowBB(BasicBlock* block);
4425 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4426 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4427 GenTreePtr fgMakeTmpArgNode(
4428 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4430 // The following check for loops that don't execute calls
4431 bool fgLoopCallMarked;
4433 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4434 void fgLoopCallMark();
4436 void fgMarkLoopHead(BasicBlock* block);
4438 unsigned fgGetCodeEstimate(BasicBlock* block);
4441 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4442 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4443 bool fgDumpFlowGraph(Phases phase);
4445 #endif // DUMP_FLOWGRAPHS
4450 void fgDispBBLiveness(BasicBlock* block);
4451 void fgDispBBLiveness();
4452 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4453 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4454 void fgDispBasicBlocks(bool dumpTrees = false);
4455 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4456 void fgDumpBlock(BasicBlock* block);
4457 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4459 static fgWalkPreFn fgStress64RsltMulCB;
4460 void fgStress64RsltMul();
4461 void fgDebugCheckUpdate();
4462 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4463 void fgDebugCheckBlockLinks();
4464 void fgDebugCheckLinks(bool morphTrees = false);
4465 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4466 void fgDebugCheckFlags(GenTreePtr tree);
4467 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4468 void fgDebugCheckTryFinallyExits();
4471 #ifdef LEGACY_BACKEND
4472 static void fgOrderBlockOps(GenTreePtr tree,
4476 GenTreePtr* opsPtr, // OUT
4477 regMaskTP* regsPtr); // OUT
4478 #endif // LEGACY_BACKEND
4480 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4481 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4482 void fgTraverseRPO();
4484 //--------------------- Walking the trees in the IR -----------------------
4489 fgWalkPreFn* wtprVisitorFn;
4490 fgWalkPostFn* wtpoVisitorFn;
4491 void* pCallbackData; // user-provided data
4492 bool wtprLclsOnly; // whether to only visit lclvar nodes
4493 GenTreePtr parent; // parent of current node, provided to callback
4494 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4496 bool printModified; // callback can use this
4500 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4501 fgWalkPreFn* visitor,
4502 void* pCallBackData = nullptr,
4503 bool lclVarsOnly = false,
4504 bool computeStack = false);
4506 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4507 fgWalkPreFn* preVisitor,
4508 fgWalkPostFn* postVisitor,
4509 void* pCallBackData = nullptr);
4511 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4515 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4516 fgWalkPostFn* visitor,
4517 void* pCallBackData = nullptr,
4518 bool computeStack = false);
4520 // An fgWalkPreFn that looks for expressions that have inline throws in
4521 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4522 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4523 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4524 // properly propagated to parent trees). It returns WALK_CONTINUE
4526 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4527 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4528 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4530 /**************************************************************************
4532 *************************************************************************/
4535 friend class SsaBuilder;
4536 friend struct ValueNumberState;
4538 //--------------------- Detect the basic blocks ---------------------------
4540 BasicBlock** fgBBs; // Table of pointers to the BBs
4542 void fgInitBBLookup();
4543 BasicBlock* fgLookupBB(unsigned addr);
4545 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4547 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4549 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4551 void fgLinkBasicBlocks();
4553 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4555 void fgCheckBasicBlockControlFlow();
4557 void fgControlFlowPermitted(BasicBlock* blkSrc,
4558 BasicBlock* blkDest,
4559 BOOL IsLeave = false /* is the src a leave block */);
4561 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4563 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4565 void fgAdjustForAddressExposedOrWrittenThis();
4567 bool fgProfileData_ILSizeMismatch;
4568 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4569 ULONG fgProfileBufferCount;
4570 ULONG fgNumProfileRuns;
4572 unsigned fgStressBBProf()
4575 unsigned result = JitConfig.JitStressBBProf();
4578 if (compStressCompile(STRESS_BB_PROFILE, 15))
4589 bool fgHaveProfileData();
4590 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4591 void fgInstrumentMethod();
4594 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4595 // or if we have some fake profile data for the stress mode
4596 bool fgIsUsingProfileWeights()
4598 return (fgHaveProfileData() || fgStressBBProf());
4601 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4602 // or BB_UNITY_WEIGHT when we aren't using profile data.
4603 unsigned fgProfileRunsCount()
4605 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4608 //-------- Insert a statement at the start or end of a basic block --------
4612 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4616 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4618 public: // Used by linear scan register allocation
4619 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4622 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4623 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4625 public: // Used by linear scan register allocation
4626 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4629 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4631 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4633 // Create a new temporary variable to hold the result of *ppTree,
4634 // and transform the graph accordingly.
4635 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4636 GenTree* fgMakeMultiUse(GenTree** ppTree);
4639 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4640 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4641 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4643 //-------- Determine the order in which the trees will be evaluated -------
4645 unsigned fgTreeSeqNum;
4646 GenTree* fgTreeSeqLst;
4647 GenTree* fgTreeSeqBeg;
4649 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4650 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4651 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4652 void fgSetStmtSeq(GenTree* tree);
4653 void fgSetBlockOrder(BasicBlock* block);
4655 //------------------------- Morphing --------------------------------------
4657 unsigned fgPtrArgCntCur;
4658 unsigned fgPtrArgCntMax;
4659 hashBv* fgOutgoingArgTemps;
4660 hashBv* fgCurrentlyInUseArgTemps;
4662 bool compCanEncodePtrArgCntMax();
4664 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4667 void fgMoveOpsLeft(GenTreePtr tree);
4670 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4672 bool fgIsThrow(GenTreePtr tree);
4674 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4675 bool fgIsBlockCold(BasicBlock* block);
4677 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4679 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4681 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4683 bool fgMorphRelopToQmark(GenTreePtr tree);
4685 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4686 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4687 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4688 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4689 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4690 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4691 // small; hence the other fields of MorphAddrContext.
4692 enum MorphAddrContextKind
4697 struct MorphAddrContext
4699 MorphAddrContextKind m_kind;
4700 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4701 // top-level indirection and here have been constants.
4702 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4703 // In that case, is the sum of those constant offsets.
4705 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4710 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4711 static MorphAddrContext s_CopyBlockMAC;
4714 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4715 var_types* baseTypeOut,
4717 unsigned* simdSizeOut,
4718 bool ignoreUsedInSIMDIntrinsic = false);
4719 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4720 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4721 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4722 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4724 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4725 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4726 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4728 #endif // FEATURE_SIMD
4729 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4730 GenTreePtr fgMorphCast(GenTreePtr tree);
4731 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4732 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4734 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4737 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4738 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4740 void fgFixupStructReturn(GenTreePtr call);
4741 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4742 bool fgAddrCouldBeNull(GenTreePtr addr);
4743 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4744 bool fgCanFastTailCall(GenTreeCall* call);
4745 void fgMorphTailCall(GenTreeCall* call);
4746 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4747 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4748 fgArgTabEntryPtr argTabEntry,
4750 IL_OFFSETX callILOffset,
4751 GenTreePtr tmpAssignmentInsertionPoint,
4752 GenTreePtr paramAssignmentInsertionPoint);
4753 static int fgEstimateCallStackSize(GenTreeCall* call);
4754 GenTreePtr fgMorphCall(GenTreeCall* call);
4755 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4756 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4758 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4759 static fgWalkPreFn fgFindNonInlineCandidate;
4761 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4762 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4763 CORINFO_RESOLVED_TOKEN* ldftnToken);
4764 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4765 void fgAssignSetVarDef(GenTreePtr tree);
4766 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4767 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4768 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4769 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4770 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4771 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4772 void fgMorphUnsafeBlk(GenTreeObj* obj);
4773 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4774 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4775 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4776 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4777 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4778 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4779 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4781 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4782 GenTreePtr fgMorphConst(GenTreePtr tree);
4785 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4788 #if LOCAL_ASSERTION_PROP
4789 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4790 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4792 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4794 GenTreeStmt* fgMorphStmt;
4796 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4797 // used when morphing big offset.
4799 //----------------------- Liveness analysis -------------------------------
4801 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4802 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4804 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4805 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4806 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4808 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4810 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4812 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4813 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4815 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4816 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4818 void fgExtendDbgScopes();
4819 void fgExtendDbgLifetimes();
4822 void fgDispDebugScopes();
4825 //-------------------------------------------------------------------------
4827 // The following keeps track of any code we've added for things like array
4828 // range checking or explicit calls to enable GC, and so on.
4833 AddCodeDsc* acdNext;
4834 BasicBlock* acdDstBlk; // block to which we jump
4836 SpecialCodeKind acdKind; // what kind of a special block is this?
4837 unsigned short acdStkLvl;
4841 static unsigned acdHelper(SpecialCodeKind codeKind);
4843 AddCodeDsc* fgAddCodeList;
4845 bool fgRngChkThrowAdded;
4846 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4848 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4850 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4853 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4856 bool fgIsCodeAdded();
4858 bool fgIsThrowHlpBlk(BasicBlock* block);
4859 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4861 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4863 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4864 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4865 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4866 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4867 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4869 #if FEATURE_MULTIREG_RET
4870 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4871 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4872 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4873 #endif // FEATURE_MULTIREG_RET
4875 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4878 static fgWalkPreFn fgDebugCheckInlineCandidates;
4880 void CheckNoFatPointerCandidatesLeft();
4881 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4884 void fgPromoteStructs();
4885 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4886 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4888 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4889 void fgMarkImplicitByRefArgs();
4891 // Change implicit byrefs' types from struct to pointer, and for any that were
4892 // promoted, create new promoted struct temps.
4893 void fgRetypeImplicitByRefArgs();
4895 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4896 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4897 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4899 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4900 void fgMarkDemotedImplicitByRefArgs();
4902 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4903 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4904 void fgMarkAddressExposedLocals();
4905 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4907 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4909 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4911 // The given local variable, required to be a struct variable, is being assigned via
4912 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4913 // the variable is not enregistered, and is therefore not promoted independently.
4914 void fgLclFldAssign(unsigned lclNum);
4916 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4917 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4918 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4919 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4922 bool fgPrintInlinedMethods;
4925 bool fgIsBigOffset(size_t offset);
4927 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4928 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4929 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4930 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4931 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4934 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4935 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4939 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4940 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4947 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4950 void optRemoveRangeCheck(
4951 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4952 bool optIsRangeCheckRemovable(GenTreePtr tree);
4955 static fgWalkPreFn optValidRangeCheckIndex;
4956 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4959 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4961 /**************************************************************************
4963 *************************************************************************/
4966 // Do hoisting for all loops.
4967 void optHoistLoopCode();
4969 // To represent sets of VN's that have already been hoisted in outer loops.
4970 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4971 typedef VNToBoolMap VNSet;
4973 struct LoopHoistContext
4976 // The set of variables hoisted in the current loop (or nullptr if there are none).
4977 VNSet* m_pHoistedInCurLoop;
4980 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4981 VNSet m_hoistedInParentLoops;
4982 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4983 // Previous decisions on loop-invariance of value numbers in the current loop.
4984 VNToBoolMap m_curLoopVnInvariantCache;
4986 VNSet* GetHoistedInCurLoop(Compiler* comp)
4988 if (m_pHoistedInCurLoop == nullptr)
4990 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4992 return m_pHoistedInCurLoop;
4995 VNSet* ExtractHoistedInCurLoop()
4997 VNSet* res = m_pHoistedInCurLoop;
4998 m_pHoistedInCurLoop = nullptr;
5002 LoopHoistContext(Compiler* comp)
5003 : m_pHoistedInCurLoop(nullptr)
5004 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5005 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5010 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5011 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5012 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5013 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5015 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5016 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5017 // "m_hoistedInParentLoops".
5019 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5021 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5022 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5023 // expressions to "hoistInLoop".
5024 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5026 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5027 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5029 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5030 // that are invariant in loop "lnum" (an index into the optLoopTable)
5031 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5032 // expressions to "hoistInLoop".
5033 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5034 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5035 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5036 bool optHoistLoopExprsForTree(GenTreePtr tree,
5038 LoopHoistContext* hoistCtxt,
5039 bool* firstBlockAndBeforeSideEffect,
5041 bool* pCctorDependent);
5043 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5044 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5046 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5047 // Constants and init values are always loop invariant.
5048 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5049 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5051 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5052 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5053 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5054 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5055 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5057 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5058 // in the loop table.
5059 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5061 // Records the set of "side effects" of all loops: fields (object instance and static)
5062 // written to, and SZ-array element type equivalence classes updated.
5063 void optComputeLoopSideEffects();
5066 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5067 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5068 // static) written to, and SZ-array element type equivalence classes updated.
5069 void optComputeLoopNestSideEffects(unsigned lnum);
5071 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5072 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5074 // Hoist the expression "expr" out of loop "lnum".
5075 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5078 void optOptimizeBools();
5081 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5083 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5086 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5088 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5089 // the loop into a "do-while" loop
5090 // Also finds all natural loops and records them in the loop table
5092 // Optionally clone loops in the loop table.
5093 void optCloneLoops();
5095 // Clone loop "loopInd" in the loop table.
5096 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5098 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5099 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5100 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5102 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5104 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5107 // This enumeration describes what is killed by a call.
5111 CALLINT_NONE, // no interference (most helpers)
5112 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5113 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5114 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5115 CALLINT_ALL, // kills everything (normal method call)
5119 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5120 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5121 // in bbNext order; we use comparisons on the bbNum to decide order.)
5122 // The blocks that define the body are
5123 // first <= top <= entry <= bottom .
5124 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5125 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5126 // Compiler::optFindNaturalLoops().
5129 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5130 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5131 // loop, but not the outer loop.)
5132 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5134 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5135 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5136 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5138 callInterf lpAsgCall; // "callInterf" for calls in the loop
5139 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5140 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5142 unsigned short lpFlags; // Mask of the LPFLG_* constants
5144 unsigned char lpExitCnt; // number of exits from the loop
5146 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5147 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5148 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5149 // (Actually, an "immediately" nested loop --
5150 // no other child of this loop is a parent of lpChild.)
5151 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5152 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5153 // by following "lpChild" then "lpSibling" links.
5155 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5156 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5158 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5159 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5160 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5162 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5163 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5165 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5166 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5167 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5168 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5170 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5171 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5172 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5174 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5175 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5176 // type are assigned to.
5178 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5179 // memory side effects. If this is set, the fields below
5180 // may not be accurate (since they become irrelevant.)
5181 bool lpContainsCall; // True if executing the loop body *may* execute a call
5183 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5184 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5186 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5188 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5189 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5191 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5193 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5194 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5196 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5197 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5199 JitSimplerHashBehavior>
5201 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5202 // instance fields modified
5205 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5206 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5208 JitSimplerHashBehavior>
5210 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5211 // arrays of that type are modified
5214 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5215 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5217 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5218 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5219 // (shifted left, with a low-order bit set to distinguish.)
5220 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5221 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5223 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5225 GenTreePtr lpIterTree; // The "i <op>= const" tree
5226 unsigned lpIterVar(); // iterator variable #
5227 int lpIterConst(); // the constant with which the iterator is incremented
5228 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5229 void VERIFY_lpIterTree();
5231 var_types lpIterOperType(); // For overflow instructions
5234 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5235 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5239 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5241 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5242 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5243 void VERIFY_lpTestTree();
5245 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5246 GenTreePtr lpIterator(); // the iterator node in the loop test
5247 GenTreePtr lpLimit(); // the limit node in the loop test
5249 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5250 // LPFLG_CONST_LIMIT
5251 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5253 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5254 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5255 // LPFLG_ARRLEN_LIMIT
5257 // Returns "true" iff "*this" contains the blk.
5258 bool lpContains(BasicBlock* blk)
5260 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5262 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5263 // to be equal, but requiring bottoms to be different.)
5264 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5266 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5269 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5270 // bottoms to be different.)
5271 bool lpContains(const LoopDsc& lp2)
5273 return lpContains(lp2.lpFirst, lp2.lpBottom);
5276 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5277 // (allowing firsts to be equal, but requiring bottoms to be different.)
5278 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5280 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5283 // Returns "true" iff "*this" is (properly) contained by "lp2"
5284 // (allowing firsts to be equal, but requiring bottoms to be different.)
5285 bool lpContainedBy(const LoopDsc& lp2)
5287 return lpContains(lp2.lpFirst, lp2.lpBottom);
5290 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5291 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5293 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5295 // Returns "true" iff "*this" is disjoint from "lp2".
5296 bool lpDisjoint(const LoopDsc& lp2)
5298 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5300 // Returns "true" iff the loop is well-formed (see code for defn).
5303 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5304 lpEntry->bbNum <= lpBottom->bbNum &&
5305 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5310 bool fgMightHaveLoop(); // returns true if there are any backedges
5311 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5314 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5315 unsigned char optLoopCount; // number of tracked loops
5318 unsigned optCallCount; // number of calls made in the method
5319 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5320 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5321 unsigned optLoopsCloned; // number of loops cloned in the current method.
5324 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5325 void optPrintLoopInfo(unsigned loopNum,
5327 BasicBlock* lpFirst,
5329 BasicBlock* lpEntry,
5330 BasicBlock* lpBottom,
5331 unsigned char lpExitCnt,
5333 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5334 void optPrintLoopInfo(unsigned lnum);
5335 void optPrintLoopRecording(unsigned lnum);
5337 void optCheckPreds();
5340 void optSetBlockWeights();
5342 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5344 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5346 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5348 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5349 unsigned optIsLoopIncrTree(GenTreePtr incr);
5350 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5351 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5352 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5353 bool optExtractInitTestIncr(BasicBlock* head,
5358 GenTreePtr* ppIncr);
5360 void optRecordLoop(BasicBlock* head,
5366 unsigned char exitCnt);
5368 void optFindNaturalLoops();
5370 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5371 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5372 bool optCanonicalizeLoopNest(unsigned char loopInd);
5374 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5375 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5376 bool optCanonicalizeLoop(unsigned char loopInd);
5378 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5379 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5380 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5381 bool optLoopContains(unsigned l1, unsigned l2);
5383 // Requires "loopInd" to be a valid index into the loop table.
5384 // Updates the loop table by changing loop "loopInd", whose head is required
5385 // to be "from", to be "to". Also performs this transformation for any
5386 // loop nested in "loopInd" that shares the same head as "loopInd".
5387 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5389 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5390 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5391 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5393 // Marks the containsCall information to "lnum" and any parent loops.
5394 void AddContainsCallAllContainingLoops(unsigned lnum);
5395 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5396 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5397 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5398 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5399 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5400 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5402 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5403 // of "from".) Copies the jump destination from "from" to "to".
5404 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5406 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5407 unsigned optLoopDepth(unsigned lnum)
5409 unsigned par = optLoopTable[lnum].lpParent;
5410 if (par == BasicBlock::NOT_IN_LOOP)
5416 return 1 + optLoopDepth(par);
5420 void fgOptWhileLoop(BasicBlock* block);
5422 bool optComputeLoopRep(int constInit,
5425 genTreeOps iterOper,
5427 genTreeOps testOper,
5430 unsigned* iterCount);
5431 #if FEATURE_STACK_FP_X87
5434 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5435 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5436 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5437 #endif // FEATURE_STACK_FP_X87
5440 static fgWalkPreFn optIsVarAssgCB;
5443 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5445 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5447 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5449 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5451 /**************************************************************************
5452 * Optimization conditions
5453 *************************************************************************/
5455 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5456 bool optPentium4(void);
5457 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5458 bool optAvoidIntMult(void);
5463 // The following is the upper limit on how many expressions we'll keep track
5464 // of for the CSE analysis.
5466 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5468 static const int MIN_CSE_COST = 2;
5470 // Keeps tracked cse indices
5471 BitVecTraits* cseTraits;
5474 /* Generic list of nodes - used by the CSE logic */
5482 typedef struct treeLst* treeLstPtr;
5486 treeStmtLst* tslNext;
5487 GenTreePtr tslTree; // tree node
5488 GenTreePtr tslStmt; // statement containing the tree
5489 BasicBlock* tslBlock; // block containing the statement
5492 typedef struct treeStmtLst* treeStmtLstPtr;
5494 // The following logic keeps track of expressions via a simple hash table.
5498 CSEdsc* csdNextInBucket; // used by the hash table
5500 unsigned csdHashValue; // the orginal hashkey
5502 unsigned csdIndex; // 1..optCSECandidateCount
5503 char csdLiveAcrossCall; // 0 or 1
5505 unsigned short csdDefCount; // definition count
5506 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5508 unsigned csdDefWtCnt; // weighted def count
5509 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5511 GenTreePtr csdTree; // treenode containing the 1st occurance
5512 GenTreePtr csdStmt; // stmt containing the 1st occurance
5513 BasicBlock* csdBlock; // block containing the 1st occurance
5515 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5516 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5518 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5519 // number, this will reflect it; otherwise, NoVN.
5522 static const size_t s_optCSEhashSize;
5523 CSEdsc** optCSEhash;
5526 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5528 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5529 // re-numbered with the bound to improve range check elimination
5531 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5532 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5536 CSEdsc* optCSEfindDsc(unsigned index);
5537 void optUnmarkCSE(GenTreePtr tree);
5539 // user defined callback data for the tree walk function optCSE_MaskHelper()
5540 struct optCSE_MaskData
5542 EXPSET_TP CSE_defMask;
5543 EXPSET_TP CSE_useMask;
5546 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5547 static fgWalkPreFn optCSE_MaskHelper;
5549 // This function walks all the node for an given tree
5550 // and return the mask of CSE definitions and uses for the tree
5552 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5554 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5555 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5556 bool optCSE_canSwap(GenTree* tree);
5558 static fgWalkPostFn optPropagateNonCSE;
5559 static fgWalkPreFn optHasNonCSEChild;
5561 static fgWalkPreFn optUnmarkCSEs;
5563 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5564 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5566 void optCleanupCSEs();
5569 void optEnsureClearCSEInfo();
5572 #endif // FEATURE_ANYCSE
5574 #if FEATURE_VALNUM_CSE
5575 /**************************************************************************
5576 * Value Number based CSEs
5577 *************************************************************************/
5580 void optOptimizeValnumCSEs();
5583 void optValnumCSE_Init();
5584 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5585 unsigned optValnumCSE_Locate();
5586 void optValnumCSE_InitDataFlow();
5587 void optValnumCSE_DataFlow();
5588 void optValnumCSE_Availablity();
5589 void optValnumCSE_Heuristic();
5590 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5592 #endif // FEATURE_VALNUM_CSE
5595 bool optDoCSE; // True when we have found a duplicate CSE tree
5596 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5597 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5598 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5599 unsigned optCSEstart; // The first local variable number that is a CSE
5600 unsigned optCSEcount; // The total count of CSE's introduced.
5601 unsigned optCSEweight; // The weight of the current block when we are
5602 // scanning for CSE expressions
5604 bool optIsCSEcandidate(GenTreePtr tree);
5606 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5608 bool lclNumIsTrueCSE(unsigned lclNum) const
5610 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5613 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5615 bool lclNumIsCSE(unsigned lclNum) const
5617 return lvaTable[lclNum].lvIsCSE;
5621 bool optConfigDisableCSE();
5622 bool optConfigDisableCSE2();
5624 void optOptimizeCSEs();
5626 #endif // FEATURE_ANYCSE
5634 unsigned ivaVar; // Variable we are interested in, or -1
5635 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5636 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5637 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5638 callInterf ivaMaskCall; // What kind of calls are there?
5641 static callInterf optCallInterf(GenTreeCall* call);
5644 // VN based copy propagation.
5645 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5646 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5647 LclNumToGenTreePtrStack;
5649 // Kill set to track variables with intervening definitions.
5650 VARSET_TP optCopyPropKillSet;
5652 // Copy propagation functions.
5653 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5654 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5655 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5656 bool optIsSsaLocal(GenTreePtr tree);
5657 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5658 void optVnCopyProp();
5660 /**************************************************************************
5661 * Early value propagation
5662 *************************************************************************/
5668 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5672 static unsigned GetHashCode(SSAName ssaNm)
5674 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5677 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5679 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5683 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5684 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5685 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5686 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5687 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5688 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5690 bool doesMethodHaveFatPointer()
5692 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5695 void setMethodHasFatPointer()
5697 optMethodFlags |= OMF_HAS_FATPOINTER;
5700 void clearMethodHasFatPointer()
5702 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5705 void addFatPointerCandidate(GenTreeCall* call)
5707 setMethodHasFatPointer();
5708 call->SetFatPointerCandidate();
5711 unsigned optMethodFlags;
5713 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5714 // No throughput diff was found with backward walk bound between 3-8.
5715 static const int optEarlyPropRecurBound = 5;
5717 enum class optPropKind
5725 bool gtIsVtableRef(GenTreePtr tree);
5726 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5727 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5728 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5729 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5730 bool optEarlyPropRewriteTree(GenTreePtr tree);
5731 bool optDoEarlyPropForBlock(BasicBlock* block);
5732 bool optDoEarlyPropForFunc();
5733 void optEarlyProp();
5734 void optFoldNullCheck(GenTreePtr tree);
5735 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5738 /**************************************************************************
5739 * Value/Assertion propagation
5740 *************************************************************************/
5742 // Data structures for assertion prop
5743 BitVecTraits* apTraits;
5746 enum optAssertionKind
5763 O1K_CONSTANT_LOOP_BND,
5784 optAssertionKind assertionKind;
5787 unsigned lclNum; // assigned to or property of this local var number
5795 struct AssertionDscOp1
5797 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5804 struct AssertionDscOp2
5806 optOp2Kind kind; // a const or copy assignment
5810 ssize_t iconVal; // integer
5811 unsigned iconFlags; // gtFlags
5813 struct Range // integer subrange
5827 bool IsCheckedBoundArithBound()
5829 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5831 bool IsCheckedBoundBound()
5833 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5835 bool IsConstantBound()
5837 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5838 op1.kind == O1K_CONSTANT_LOOP_BND);
5840 bool IsBoundsCheckNoThrow()
5842 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5845 bool IsCopyAssertion()
5847 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5850 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5852 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5853 a1->op2.kind == a2->op2.kind;
5856 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5858 if (kind == OAK_EQUAL)
5860 return kind2 == OAK_NOT_EQUAL;
5862 else if (kind == OAK_NOT_EQUAL)
5864 return kind2 == OAK_EQUAL;
5869 static ssize_t GetLowerBoundForIntegralType(var_types type)
5889 static ssize_t GetUpperBoundForIntegralType(var_types type)
5913 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5915 if (op1.kind != that->op1.kind)
5919 else if (op1.kind == O1K_ARR_BND)
5922 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5926 return ((vnBased && (op1.vn == that->op1.vn)) ||
5927 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5931 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5933 if (op2.kind != that->op2.kind)
5939 case O2K_IND_CNS_INT:
5941 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5943 case O2K_CONST_LONG:
5944 return (op2.lconVal == that->op2.lconVal);
5946 case O2K_CONST_DOUBLE:
5947 // exact match because of positive and negative zero.
5948 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5950 case O2K_LCLVAR_COPY:
5952 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5953 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5956 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5959 // we will return false
5963 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5969 bool Complementary(AssertionDsc* that, bool vnBased)
5971 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5972 HasSameOp2(that, vnBased);
5975 bool Equals(AssertionDsc* that, bool vnBased)
5977 if (assertionKind != that->assertionKind)
5981 else if (assertionKind == OAK_NO_THROW)
5983 assert(op2.kind == O2K_INVALID);
5984 return HasSameOp1(that, vnBased);
5988 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5994 static fgWalkPreFn optAddCopiesCallback;
5995 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5996 unsigned optAddCopyLclNum;
5997 GenTreePtr optAddCopyAsgnNode;
5999 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6000 bool optAssertionPropagated; // set to true if we modified the trees
6001 bool optAssertionPropagatedCurrentStmt;
6003 GenTreePtr optAssertionPropCurrentTree;
6005 AssertionIndex* optComplementaryAssertionMap;
6006 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6007 // using the value of a local var) for each local var
6008 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6009 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6010 AssertionIndex optMaxAssertionCount;
6013 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6014 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6015 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6016 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6017 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6018 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6020 AssertionIndex GetAssertionCount()
6022 return optAssertionCount;
6024 ASSERT_TP* bbJtrueAssertionOut;
6025 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6026 ValueNumToAssertsMap;
6027 ValueNumToAssertsMap* optValueNumToAsserts;
6029 // Assertion prop helpers.
6030 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6031 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6032 void optAssertionInit(bool isLocalProp);
6033 void optAssertionTraitsInit(AssertionIndex assertionCount);
6034 #if LOCAL_ASSERTION_PROP
6035 void optAssertionReset(AssertionIndex limit);
6036 void optAssertionRemove(AssertionIndex index);
6039 // Assertion prop data flow functions.
6040 void optAssertionPropMain();
6041 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6042 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6043 ASSERT_TP* optInitAssertionDataflowFlags();
6044 ASSERT_TP* optComputeAssertionGen();
6046 // Assertion Gen functions.
6047 void optAssertionGen(GenTreePtr tree);
6048 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6049 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6050 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6051 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6052 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6053 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6055 // Assertion creation functions.
6056 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6057 AssertionIndex optCreateAssertion(GenTreePtr op1,
6059 optAssertionKind assertionKind,
6060 AssertionDsc* assertion);
6061 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6063 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6064 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6065 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6067 void optPrintVnAssertionMapping();
6069 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6071 // Used for respective assertion propagations.
6072 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6073 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6074 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6075 bool optAssertionIsNonNull(GenTreePtr op,
6076 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6078 // Used for Relop propagation.
6079 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6080 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6081 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6083 // Assertion prop for lcl var functions.
6084 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6085 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6087 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6088 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6089 const GenTreePtr tree,
6090 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6091 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6093 // Assertion propagation functions.
6094 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6095 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6096 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6097 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6098 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6099 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6100 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6101 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6102 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6103 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6104 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6105 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6107 // Implied assertion functions.
6108 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6109 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6110 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6111 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6114 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6115 void optDebugCheckAssertion(AssertionDsc* assertion);
6116 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6118 void optAddCopies();
6119 #endif // ASSERTION_PROP
6121 /**************************************************************************
6123 *************************************************************************/
6126 struct LoopCloneVisitorInfo
6128 LoopCloneContext* context;
6131 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6132 : context(context), loopNum(loopNum), stmt(nullptr)
6137 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6138 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6139 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6140 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6141 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6142 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6143 void optObtainLoopCloningOpts(LoopCloneContext* context);
6144 bool optIsLoopClonable(unsigned loopInd);
6146 bool optCanCloneLoops();
6149 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6151 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6152 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6153 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6154 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6158 void optInsertLoopCloningStress(BasicBlock* head);
6160 #if COUNT_RANGECHECKS
6161 static unsigned optRangeChkRmv;
6162 static unsigned optRangeChkAll;
6171 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6176 RngChkDsc* rcdNextInBucket; // used by the hash table
6178 unsigned short rcdHashValue; // to make matching faster
6179 unsigned short rcdIndex; // 0..optRngChkCount-1
6181 GenTreePtr rcdTree; // the array index tree
6184 unsigned optRngChkCount;
6185 static const size_t optRngChkHashSize;
6187 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6188 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6190 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6193 bool optLoopsMarked;
6196 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6197 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6201 XX Does the register allocation and puts the remaining lclVars on the stack XX
6203 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6204 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6208 #ifndef LEGACY_BACKEND
6213 #else // LEGACY_BACKEND
6218 #endif // LEGACY_BACKEND
6220 #ifdef LEGACY_BACKEND
6222 void raAssignVars(); // register allocation
6223 #endif // LEGACY_BACKEND
6225 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6227 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6229 void raMarkStkVars();
6232 // Some things are used by both LSRA and regpredict allocators.
6234 FrameType rpFrameType;
6235 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6237 #ifdef LEGACY_BACKEND
6238 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6240 #endif // LEGACY_BACKEND
6242 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6244 #if FEATURE_FP_REGALLOC
6245 enum enumConfigRegisterFP
6247 CONFIG_REGISTER_FP_NONE = 0x0,
6248 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6249 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6250 CONFIG_REGISTER_FP_FULL = 0x3,
6252 enumConfigRegisterFP raConfigRegisterFP();
6253 #endif // FEATURE_FP_REGALLOC
6256 regMaskTP raConfigRestrictMaskFP();
6259 #ifndef LEGACY_BACKEND
6260 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6261 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6262 #else // LEGACY_BACKEND
6263 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6264 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6265 bool raNewBlocks; // True is we added killing blocks for FPU registers
6266 unsigned rpPasses; // Number of passes made by the register predicter
6267 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6268 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6269 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6270 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6271 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6272 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6273 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6274 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6275 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6276 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6277 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6278 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6280 bool rpRegAllocDone; // Set to true after we have completed register allocation
6282 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6284 void raSetupArgMasks(RegState* r);
6286 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6288 void raDumpVarIntf(); // Dump the variable to variable interference graph
6289 void raDumpRegIntf(); // Dump the variable to register interference graph
6291 void raAdjustVarIntf();
6293 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6295 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6297 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6298 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6300 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6302 static fgWalkPreFn rpMarkRegIntf;
6304 regMaskTP rpPredictAddressMode(
6305 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6307 void rpPredictRefAssign(unsigned lclNum);
6309 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6311 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6313 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6315 void rpPredictRegUse(); // Entry point
6317 unsigned raPredictTreeRegUse(GenTreePtr tree);
6318 unsigned raPredictListRegUse(GenTreePtr list);
6320 void raSetRegVarOrder(var_types regType,
6321 regNumber* customVarOrder,
6322 unsigned* customVarOrderSize,
6324 regMaskTP avoidReg);
6326 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6327 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6328 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6329 void raAddToStkPredict(unsigned val)
6331 unsigned newStkPredict = rpStkPredict + val;
6332 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6333 rpStkPredict = UINT_MAX - 1;
6335 rpStkPredict = newStkPredict;
6339 #if !FEATURE_FP_REGALLOC
6340 void raDispFPlifeInfo();
6344 regMaskTP genReturnRegForTree(GenTreePtr tree);
6345 #endif // LEGACY_BACKEND
6347 /* raIsVarargsStackArg is called by raMaskStkVars and by
6348 lvaSortByRefCount. It identifies the special case
6349 where a varargs function has a parameter passed on the
6350 stack, other than the special varargs handle. Such parameters
6351 require special treatment, because they cannot be tracked
6352 by the GC (their offsets in the stack are not known
6356 bool raIsVarargsStackArg(unsigned lclNum)
6360 LclVarDsc* varDsc = &lvaTable[lclNum];
6362 assert(varDsc->lvIsParam);
6364 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6366 #else // _TARGET_X86_
6370 #endif // _TARGET_X86_
6373 #ifdef LEGACY_BACKEND
6374 // Records the current prediction, if it's better than any previous recorded prediction.
6375 void rpRecordPrediction();
6376 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6377 void rpUseRecordedPredictionIfBetter();
6379 // Data members used in the methods above.
6380 unsigned rpBestRecordedStkPredict;
6381 struct VarRegPrediction
6383 bool m_isEnregistered;
6384 regNumberSmall m_regNum;
6385 regNumberSmall m_otherReg;
6387 VarRegPrediction* rpBestRecordedPrediction;
6388 #endif // LEGACY_BACKEND
6391 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6392 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6396 XX Get to the class and method info from the Execution Engine given XX
6397 XX tokens for the class and method XX
6399 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6400 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6404 /* These are the different addressing modes used to access a local var.
6405 * The JIT has to report the location of the locals back to the EE
6406 * for debugging purposes.
6412 VLT_REG_BYREF, // this type is currently only used for value types on X64
6415 VLT_STK_BYREF, // this type is currently only used for value types on X64
6429 siVarLocType vlType;
6432 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6434 // VLT_REG_BYREF -- the specified register contains the address of the variable
6442 // VLT_STK -- Any 32 bit value which is on the stack
6443 // eg. [ESP+0x20], or [EBP-0x28]
6444 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6445 // eg. mov EAX, [ESP+0x20]; [EAX]
6449 regNumber vlsBaseReg;
6450 NATIVE_OFFSET vlsOffset;
6453 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6462 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6463 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6471 regNumber vlrssBaseReg;
6472 NATIVE_OFFSET vlrssOffset;
6476 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6477 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6483 regNumber vlsrsBaseReg;
6484 NATIVE_OFFSET vlsrsOffset;
6490 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6491 // eg 2 DWords at [ESP+0x10]
6495 regNumber vls2BaseReg;
6496 NATIVE_OFFSET vls2Offset;
6499 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6500 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6507 // VLT_FIXED_VA -- fixed argument of a varargs function.
6508 // The argument location depends on the size of the variable
6509 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6510 // location of the first arg. This argument can then be accessed
6511 // relative to the position of the first arg
6515 unsigned vlfvOffset;
6522 void* rpValue; // pointer to the in-process
6523 // location of the value.
6529 bool vlIsInReg(regNumber reg);
6530 bool vlIsOnStk(regNumber reg, signed offset);
6533 /*************************************************************************/
6538 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6539 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6540 CORINFO_CALLINFO_FLAGS flags,
6541 CORINFO_CALL_INFO* pResult);
6542 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6544 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6545 CORINFO_ACCESS_FLAGS flags,
6546 CORINFO_FIELD_INFO* pResult);
6550 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6552 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6554 bool IsSuperPMIException(unsigned code)
6556 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6558 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6559 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6560 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6561 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6562 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6563 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6564 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6565 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6569 case EXCEPTIONCODE_DebugBreakorAV:
6570 case EXCEPTIONCODE_MC:
6571 case EXCEPTIONCODE_LWM:
6572 case EXCEPTIONCODE_SASM:
6573 case EXCEPTIONCODE_SSYM:
6574 case EXCEPTIONCODE_CALLUTILS:
6575 case EXCEPTIONCODE_TYPEUTILS:
6576 case EXCEPTIONCODE_ASSERT:
6583 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6584 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6586 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6587 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6590 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6591 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6592 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6594 // VOM info, method sigs
6596 void eeGetSig(unsigned sigTok,
6597 CORINFO_MODULE_HANDLE scope,
6598 CORINFO_CONTEXT_HANDLE context,
6599 CORINFO_SIG_INFO* retSig);
6601 void eeGetCallSiteSig(unsigned sigTok,
6602 CORINFO_MODULE_HANDLE scope,
6603 CORINFO_CONTEXT_HANDLE context,
6604 CORINFO_SIG_INFO* retSig);
6606 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6608 // Method entry-points, instrs
6610 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6612 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6614 CORINFO_EE_INFO eeInfo;
6615 bool eeInfoInitialized;
6617 CORINFO_EE_INFO* eeGetEEInfo();
6619 // Gets the offset of a SDArray's first element
6620 unsigned eeGetArrayDataOffset(var_types type);
6621 // Gets the offset of a MDArray's first element
6622 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6624 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6626 // Returns the page size for the target machine as reported by the EE.
6627 inline size_t eeGetPageSize()
6629 return eeGetEEInfo()->osPageSize;
6632 // Returns the frame size at which we will generate a loop to probe the stack.
6633 inline size_t getVeryLargeFrameSize()
6636 // The looping probe code is 40 bytes, whereas the straight-line probing for
6637 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6638 // or greater, to generate smaller code.
6639 return 2 * eeGetPageSize();
6641 return 3 * eeGetPageSize();
6645 //------------------------------------------------------------------------
6646 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6648 // It represents Abi and target specific registers for the parameter.
6650 class VirtualStubParamInfo
6653 VirtualStubParamInfo(bool isCoreRTABI)
6655 #if defined(_TARGET_X86_)
6658 #elif defined(_TARGET_AMD64_)
6669 #elif defined(_TARGET_ARM_)
6672 #elif defined(_TARGET_ARM64_)
6676 #error Unsupported or unset target architecture
6679 #ifdef LEGACY_BACKEND
6680 #if defined(_TARGET_X86_)
6681 predict = PREDICT_REG_EAX;
6682 #elif defined(_TARGET_ARM_)
6683 predict = PREDICT_REG_R4;
6685 #error Unsupported or unset target architecture
6687 #endif // LEGACY_BACKEND
6690 regNumber GetReg() const
6695 _regMask_enum GetRegMask() const
6700 #ifdef LEGACY_BACKEND
6701 rpPredictReg GetPredict() const
6709 _regMask_enum regMask;
6711 #ifdef LEGACY_BACKEND
6712 rpPredictReg predict;
6716 VirtualStubParamInfo* virtualStubParamInfo;
6718 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6720 return eeGetEEInfo()->targetAbi == abi;
6723 inline bool generateCFIUnwindCodes()
6725 #ifdef UNIX_AMD64_ABI
6726 return IsTargetAbi(CORINFO_CORERT_ABI);
6734 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6736 // Debugging support - Line number info
6738 void eeGetStmtOffsets();
6740 unsigned eeBoundariesCount;
6742 struct boundariesDsc
6744 UNATIVE_OFFSET nativeIP;
6746 unsigned sourceReason;
6747 } * eeBoundaries; // Boundaries to report to EE
6748 void eeSetLIcount(unsigned count);
6749 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6753 static void eeDispILOffs(IL_OFFSET offs);
6754 static void eeDispLineInfo(const boundariesDsc* line);
6755 void eeDispLineInfos();
6758 // Debugging support - Local var info
6762 unsigned eeVarsCount;
6764 struct VarResultInfo
6766 UNATIVE_OFFSET startOffset;
6767 UNATIVE_OFFSET endOffset;
6771 void eeSetLVcount(unsigned count);
6772 void eeSetLVinfo(unsigned which,
6773 UNATIVE_OFFSET startOffs,
6774 UNATIVE_OFFSET length,
6779 const siVarLoc& loc);
6783 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6784 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6787 // ICorJitInfo wrappers
6789 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6791 void eeAllocUnwindInfo(BYTE* pHotCode,
6797 CorJitFuncKind funcKind);
6799 void eeSetEHcount(unsigned cEH);
6801 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6803 WORD eeGetRelocTypeHint(void* target);
6805 // ICorStaticInfo wrapper functions
6807 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6809 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6811 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6814 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6815 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6816 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6817 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6819 template <typename ParamType>
6820 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6822 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6825 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6827 // Utility functions
6829 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6832 const wchar_t* eeGetCPString(size_t stringHandle);
6835 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6837 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6838 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6840 static fgWalkPreFn CountSharedStaticHelper;
6841 static bool IsSharedStaticHelper(GenTreePtr tree);
6842 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6843 static bool IsGcSafePoint(GenTreePtr tree);
6845 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6846 // returns true/false if 'field' is a Jit Data offset
6847 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6848 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6849 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6851 /*****************************************************************************/
6856 enum TEMP_USAGE_TYPE
6862 static var_types tmpNormalizeType(var_types type);
6863 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6864 void tmpRlsTemp(TempDsc* temp);
6865 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6868 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6869 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6873 bool tmpAllFree() const;
6876 #ifndef LEGACY_BACKEND
6877 void tmpPreAllocateTemps(var_types type, unsigned count);
6878 #endif // !LEGACY_BACKEND
6881 #ifdef LEGACY_BACKEND
6882 unsigned tmpIntSpillMax; // number of int-sized spill temps
6883 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6884 #endif // LEGACY_BACKEND
6886 unsigned tmpCount; // Number of temps
6887 unsigned tmpSize; // Size of all the temps
6890 // Used by RegSet::rsSpillChk()
6891 unsigned tmpGetCount; // Temps which haven't been released yet
6894 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6896 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6897 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6900 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6901 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6905 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6906 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6910 CodeGenInterface* codeGen;
6912 // The following holds information about instr offsets in terms of generated code.
6916 IPmappingDsc* ipmdNext; // next line# record
6917 IL_OFFSETX ipmdILoffsx; // the instr offset
6918 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6919 bool ipmdIsLabel; // Can this code be a branch label?
6922 // Record the instr offset mapping to the generated code
6924 IPmappingDsc* genIPmappingList;
6925 IPmappingDsc* genIPmappingLast;
6927 // Managed RetVal - A side hash table meant to record the mapping from a
6928 // GT_CALL node to its IL offset. This info is used to emit sequence points
6929 // that can be used by debugger to determine the native offset at which the
6930 // managed RetVal will be available.
6932 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6933 // favor of a side table for two reasons: 1) We need IL offset for only those
6934 // GT_CALL nodes (created during importation) that correspond to an IL call and
6935 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6936 // structure and IL offset is needed only when generating debuggable code. Therefore
6937 // it is desirable to avoid memory size penalty in retail scenarios.
6938 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6939 CallSiteILOffsetTable;
6940 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6942 unsigned genReturnLocal; // Local number for the return value when applicable.
6943 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6945 // The following properties are part of CodeGenContext. Getters are provided here for
6946 // convenience and backward compatibility, but the properties can only be set by invoking
6947 // the setter on CodeGenContext directly.
6949 __declspec(property(get = getEmitter)) emitter* genEmitter;
6950 emitter* getEmitter()
6952 return codeGen->getEmitter();
6955 const bool isFramePointerUsed()
6957 return codeGen->isFramePointerUsed();
6960 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6961 bool getInterruptible()
6963 return codeGen->genInterruptible;
6965 void setInterruptible(bool value)
6967 codeGen->setInterruptible(value);
6971 const bool genDoubleAlign()
6973 return codeGen->doDoubleAlign();
6975 DWORD getCanDoubleAlign();
6976 bool shouldDoubleAlign(unsigned refCntStk,
6978 unsigned refCntWtdReg,
6979 unsigned refCntStkParam,
6980 unsigned refCntWtdStkDbl);
6981 #endif // DOUBLE_ALIGN
6983 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6984 bool getFullPtrRegMap()
6986 return codeGen->genFullPtrRegMap;
6988 void setFullPtrRegMap(bool value)
6990 codeGen->setFullPtrRegMap(value);
6993 // Things that MAY belong either in CodeGen or CodeGenContext
6995 #if FEATURE_EH_FUNCLETS
6996 FuncInfoDsc* compFuncInfos;
6997 unsigned short compCurrFuncIdx;
6998 unsigned short compFuncInfoCount;
7000 unsigned short compFuncCount()
7002 assert(fgFuncletsCreated);
7003 return compFuncInfoCount;
7006 #else // !FEATURE_EH_FUNCLETS
7008 // This is a no-op when there are no funclets!
7009 void genUpdateCurrentFunclet(BasicBlock* block)
7014 FuncInfoDsc compFuncInfoRoot;
7016 static const unsigned compCurrFuncIdx = 0;
7018 unsigned short compFuncCount()
7023 #endif // !FEATURE_EH_FUNCLETS
7025 FuncInfoDsc* funCurrentFunc();
7026 void funSetCurrentFunc(unsigned funcIdx);
7027 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7028 unsigned int funGetFuncIdx(BasicBlock* block);
7032 VARSET_TP compCurLife; // current live variables
7033 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7035 template <bool ForCodeGen>
7036 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7038 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7040 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7043 template <bool ForCodeGen>
7044 void compUpdateLife(GenTreePtr tree);
7046 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7047 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7048 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7049 template <bool ForCodeGen>
7050 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7052 template <bool ForCodeGen>
7053 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7055 // Gets a register mask that represent the kill set for a helper call since
7056 // not all JIT Helper calls follow the standard ABI on the target architecture.
7057 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7059 // Gets a register mask that represent the kill set for a NoGC helper call.
7060 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7063 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7064 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7065 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7066 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7067 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7068 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7069 #endif // _TARGET_ARM_
7071 // 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
7073 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7075 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7076 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7077 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7078 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7079 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7080 // for the tracked var indices of the field vars, as in a live var set).
7081 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7083 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7085 if (m_promotedStructDeathVars == nullptr)
7087 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7089 return m_promotedStructDeathVars;
7093 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7094 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7098 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7099 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7102 #if !defined(__GNUC__)
7103 #pragma region Unwind information
7108 // Infrastructure functions: start/stop/reserve/emit.
7111 void unwindBegProlog();
7112 void unwindEndProlog();
7113 void unwindBegEpilog();
7114 void unwindEndEpilog();
7115 void unwindReserve();
7116 void unwindEmit(void* pHotCode, void* pColdCode);
7119 // Specific unwind information functions: called by code generation to indicate a particular
7120 // prolog or epilog unwindable instruction has been generated.
7123 void unwindPush(regNumber reg);
7124 void unwindAllocStack(unsigned size);
7125 void unwindSetFrameReg(regNumber reg, unsigned offset);
7126 void unwindSaveReg(regNumber reg, unsigned offset);
7128 #if defined(_TARGET_ARM_)
7129 void unwindPushMaskInt(regMaskTP mask);
7130 void unwindPushMaskFloat(regMaskTP mask);
7131 void unwindPopMaskInt(regMaskTP mask);
7132 void unwindPopMaskFloat(regMaskTP mask);
7133 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7134 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7135 // called via unwindPadding().
7136 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7137 // instruction and the current location.
7138 #endif // _TARGET_ARM_
7140 #if defined(_TARGET_ARM64_)
7142 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7143 // instruction and the current location.
7144 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7145 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7146 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7147 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7148 void unwindSaveNext(); // unwind code: save_next
7149 void unwindReturn(regNumber reg); // ret lr
7150 #endif // defined(_TARGET_ARM64_)
7153 // Private "helper" functions for the unwind implementation.
7157 #if FEATURE_EH_FUNCLETS
7158 void unwindGetFuncLocations(FuncInfoDsc* func,
7159 bool getHotSectionData,
7160 /* OUT */ emitLocation** ppStartLoc,
7161 /* OUT */ emitLocation** ppEndLoc);
7162 #endif // FEATURE_EH_FUNCLETS
7164 void unwindReserveFunc(FuncInfoDsc* func);
7165 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7167 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7169 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7170 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7172 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7174 #if defined(_TARGET_AMD64_)
7176 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7178 void unwindBegPrologWindows();
7179 void unwindPushWindows(regNumber reg);
7180 void unwindAllocStackWindows(unsigned size);
7181 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7182 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7184 #ifdef UNIX_AMD64_ABI
7185 void unwindBegPrologCFI();
7186 void unwindPushCFI(regNumber reg);
7187 void unwindAllocStackCFI(unsigned size);
7188 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7189 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7190 int mapRegNumToDwarfReg(regNumber reg);
7191 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7192 #endif // UNIX_AMD64_ABI
7193 #elif defined(_TARGET_ARM_)
7195 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7196 void unwindPushPopMaskFloat(regMaskTP mask);
7197 void unwindSplit(FuncInfoDsc* func);
7199 #endif // _TARGET_ARM_
7201 #if !defined(__GNUC__)
7202 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7206 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7207 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7211 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7212 XX that contains the distinguished, well-known SIMD type definitions). XX
7214 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7215 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7218 // Get highest available instruction set for floating point codegen
7219 InstructionSet getFloatingPointInstructionSet()
7221 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7224 return InstructionSet_AVX;
7229 return InstructionSet_SSE3_4;
7233 assert(canUseSSE2());
7234 return InstructionSet_SSE2;
7236 assert(!"getFPInstructionSet() is not implemented for target arch");
7238 return InstructionSet_NONE;
7242 // Get highest available instruction set for SIMD codegen
7243 InstructionSet getSIMDInstructionSet()
7245 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7246 return getFloatingPointInstructionSet();
7248 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7250 return InstructionSet_NONE;
7256 // Should we support SIMD intrinsics?
7259 // Have we identified any SIMD types?
7260 // This is currently used by struct promotion to avoid getting type information for a struct
7261 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7263 bool _usesSIMDTypes;
7264 bool usesSIMDTypes()
7266 return _usesSIMDTypes;
7268 void setUsesSIMDTypes(bool value)
7270 _usesSIMDTypes = value;
7273 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7274 // that require indexed access to the individual fields of the vector, which is not well supported
7275 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7276 unsigned lvaSIMDInitTempVarNum;
7279 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7280 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7281 CORINFO_CLASS_HANDLE SIMDIntHandle;
7282 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7283 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7284 CORINFO_CLASS_HANDLE SIMDShortHandle;
7285 CORINFO_CLASS_HANDLE SIMDByteHandle;
7286 CORINFO_CLASS_HANDLE SIMDLongHandle;
7287 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7288 CORINFO_CLASS_HANDLE SIMDULongHandle;
7289 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7290 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7291 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7292 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7294 // Get the handle for a SIMD type.
7295 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7297 if (simdBaseType == TYP_FLOAT)
7302 return SIMDVector2Handle;
7304 return SIMDVector3Handle;
7306 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7308 return SIMDVector4Handle;
7317 assert(simdType == getSIMDVectorType());
7318 switch (simdBaseType)
7321 return SIMDFloatHandle;
7323 return SIMDDoubleHandle;
7325 return SIMDIntHandle;
7327 return SIMDUShortHandle;
7329 return SIMDUShortHandle;
7331 return SIMDUByteHandle;
7333 return SIMDShortHandle;
7335 return SIMDByteHandle;
7337 return SIMDLongHandle;
7339 return SIMDUIntHandle;
7341 return SIMDULongHandle;
7343 assert(!"Didn't find a class handle for simdType");
7345 return NO_CLASS_HANDLE;
7349 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7350 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7351 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7353 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7354 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7355 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7356 bool isSIMDTypeLocal(GenTree* tree)
7358 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7361 // Returns true if the type of the tree is a byref of TYP_SIMD
7362 bool isAddrOfSIMDType(GenTree* tree)
7364 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7366 switch (tree->OperGet())
7369 return varTypeIsSIMD(tree->gtGetOp1());
7371 case GT_LCL_VAR_ADDR:
7372 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7375 return isSIMDTypeLocal(tree);
7382 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7384 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7385 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7386 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7389 // Returns base type of a TYP_SIMD local.
7390 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7391 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7393 if (isSIMDTypeLocal(tree))
7395 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7401 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7403 return info.compCompHnd->isInSIMDModule(clsHnd);
7406 bool isSIMDClass(typeInfo* pTypeInfo)
7408 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7411 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7412 // if it is not a SIMD type or is an unsupported base type.
7413 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7415 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7417 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7420 // Get SIMD Intrinsic info given the method handle.
7421 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7422 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7423 CORINFO_METHOD_HANDLE methodHnd,
7424 CORINFO_SIG_INFO* sig,
7427 var_types* baseType,
7428 unsigned* sizeBytes);
7430 // Pops and returns GenTree node from importers type stack.
7431 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7432 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7434 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7435 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7437 // Creates a GT_SIMD tree for Select operation
7438 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7440 unsigned simdVectorSize,
7445 // Creates a GT_SIMD tree for Min/Max operation
7446 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7447 CORINFO_CLASS_HANDLE typeHnd,
7449 unsigned simdVectorSize,
7453 // Transforms operands and returns the SIMD intrinsic to be applied on
7454 // transformed operands to obtain given relop result.
7455 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7456 CORINFO_CLASS_HANDLE typeHnd,
7457 unsigned simdVectorSize,
7458 var_types* baseType,
7462 // Creates a GT_SIMD tree for Abs intrinsic.
7463 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7465 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7466 // Transforms operands and returns the SIMD intrinsic to be applied on
7467 // transformed operands to obtain == comparison result.
7468 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7469 unsigned simdVectorSize,
7473 // Transforms operands and returns the SIMD intrinsic to be applied on
7474 // transformed operands to obtain > comparison result.
7475 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7476 unsigned simdVectorSize,
7480 // Transforms operands and returns the SIMD intrinsic to be applied on
7481 // transformed operands to obtain >= comparison result.
7482 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7483 unsigned simdVectorSize,
7487 // Transforms operands and returns the SIMD intrinsic to be applied on
7488 // transformed operands to obtain >= comparison result in case of int32
7489 // and small int base type vectors.
7490 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7491 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7492 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7494 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7495 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7496 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7497 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7498 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7500 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7501 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7502 GenTreePtr newobjThis,
7503 CORINFO_CLASS_HANDLE clsHnd,
7504 CORINFO_METHOD_HANDLE method,
7505 CORINFO_SIG_INFO* sig,
7508 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7510 // Whether SIMD vector occupies part of SIMD register.
7511 // SSE2: vector2f/3f are considered sub register SIMD types.
7512 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7513 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7515 unsigned sizeBytes = 0;
7516 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7517 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7520 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7522 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7525 // Get the type for the hardware SIMD vector.
7526 // This is the maximum SIMD type supported for this target.
7527 var_types getSIMDVectorType()
7529 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7536 assert(canUseSSE2());
7540 assert(!"getSIMDVectorType() unimplemented on target arch");
7545 // Get the size of the SIMD type in bytes
7546 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7548 unsigned sizeBytes = 0;
7549 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7553 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7554 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7556 // Get the the number of elements of basetype of SIMD vector given by its type handle
7557 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7559 // Get preferred alignment of SIMD type.
7560 int getSIMDTypeAlignment(var_types simdType);
7562 // Get the number of bytes in a SIMD Vector for the current compilation.
7563 unsigned getSIMDVectorRegisterByteLength()
7565 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7568 return YMM_REGSIZE_BYTES;
7572 assert(canUseSSE2());
7573 return XMM_REGSIZE_BYTES;
7576 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7581 // The minimum and maximum possible number of bytes in a SIMD vector.
7582 unsigned int maxSIMDStructBytes()
7584 return getSIMDVectorRegisterByteLength();
7586 unsigned int minSIMDStructBytes()
7588 return emitTypeSize(TYP_SIMD8);
7591 #ifdef FEATURE_AVX_SUPPORT
7592 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7593 static const unsigned maxPossibleSIMDStructBytes = 32;
7594 #else // !FEATURE_AVX_SUPPORT
7595 static const unsigned maxPossibleSIMDStructBytes = 16;
7596 #endif // !FEATURE_AVX_SUPPORT
7598 // Returns the codegen type for a given SIMD size.
7599 var_types getSIMDTypeForSize(unsigned size)
7601 var_types simdType = TYP_UNDEF;
7604 simdType = TYP_SIMD8;
7606 else if (size == 12)
7608 simdType = TYP_SIMD12;
7610 else if (size == 16)
7612 simdType = TYP_SIMD16;
7614 #ifdef FEATURE_AVX_SUPPORT
7615 else if (size == 32)
7617 simdType = TYP_SIMD32;
7619 #endif // FEATURE_AVX_SUPPORT
7622 noway_assert(!"Unexpected size for SIMD type");
7627 unsigned getSIMDInitTempVarNum()
7629 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7631 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7632 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7634 return lvaSIMDInitTempVarNum;
7637 #endif // FEATURE_SIMD
7640 //------------------------------------------------------------------------
7641 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7643 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7644 // candidate for enregistration.
7646 unsigned largestEnregisterableStructSize()
7649 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7650 if (vectorRegSize > TARGET_POINTER_SIZE)
7652 return vectorRegSize;
7655 #endif // FEATURE_SIMD
7657 return TARGET_POINTER_SIZE;
7662 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7663 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7664 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7666 // Is this var is of type simd struct?
7667 bool lclVarIsSIMDType(unsigned varNum)
7669 LclVarDsc* varDsc = lvaTable + varNum;
7670 return varDsc->lvIsSIMDType();
7673 // Is this Local node a SIMD local?
7674 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7676 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7679 // Returns true if the TYP_SIMD locals on stack are aligned at their
7680 // preferred byte boundary specified by getSIMDTypeAlignment().
7682 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7683 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7684 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7685 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7686 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7687 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7688 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7691 bool isSIMDTypeLocalAligned(unsigned varNum)
7693 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7694 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7697 int off = lvaFrameAddress(varNum, &ebpBased);
7698 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7699 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7700 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7703 #endif // FEATURE_SIMD
7708 // Whether SSE2 is available
7709 bool canUseSSE2() const
7711 #ifdef _TARGET_XARCH_
7712 return opts.compCanUseSSE2;
7718 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7719 bool CanUseSSE3_4() const
7721 #ifdef _TARGET_XARCH_
7722 return opts.compCanUseSSE3_4;
7728 bool canUseAVX() const
7730 #ifdef FEATURE_AVX_SUPPORT
7731 return opts.compCanUseAVX;
7738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7739 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7743 XX Generic info about the compilation and the method being compiled. XX
7744 XX It is responsible for driving the other phases. XX
7745 XX It is also responsible for all the memory management. XX
7747 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7748 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7752 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7754 InlineResult* compInlineResult; // The result of importing the inlinee method.
7756 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7757 bool compJmpOpUsed; // Does the method do a JMP
7758 bool compLongUsed; // Does the method use TYP_LONG
7759 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7760 bool compTailCallUsed; // Does the method do a tailcall
7761 bool compLocallocUsed; // Does the method use localloc.
7762 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7763 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7764 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7766 // NOTE: These values are only reliable after
7767 // the importing is completely finished.
7769 #ifdef LEGACY_BACKEND
7770 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7771 // we can iterate over these efficiently.
7774 #if CPU_USES_BLOCK_MOVE
7775 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7779 // State information - which phases have completed?
7780 // These are kept together for easy discoverability
7782 bool bRangeAllowStress;
7783 bool compCodeGenDone;
7784 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7785 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7786 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7787 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7790 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7791 bool fgLocalVarLivenessChanged;
7793 bool compStackProbePrologDone;
7795 #ifndef LEGACY_BACKEND
7797 #endif // !LEGACY_BACKEND
7798 bool compRationalIRForm;
7800 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7802 bool compGeneratingProlog;
7803 bool compGeneratingEpilog;
7804 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7805 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7806 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7807 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7808 bool getNeedsGSSecurityCookie() const
7810 return compNeedsGSSecurityCookie;
7812 void setNeedsGSSecurityCookie()
7814 compNeedsGSSecurityCookie = true;
7817 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7818 // frame layout calculations, this is the level we are currently
7821 //---------------------------- JITing options -----------------------------
7834 JitFlags* jitFlags; // all flags passed from the EE
7835 unsigned compFlags; // method attributes
7837 codeOptimize compCodeOpt; // what type of code optimizations
7841 #ifdef _TARGET_XARCH_
7842 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7843 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7845 #ifdef FEATURE_AVX_SUPPORT
7846 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7847 #endif // FEATURE_AVX_SUPPORT
7848 #endif // _TARGET_XARCH_
7850 // optimize maximally and/or favor speed over size?
7852 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7853 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7854 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7855 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7856 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7858 // Maximun number of locals before turning off the inlining
7859 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7862 unsigned instrCount;
7863 unsigned lvRefCount;
7864 bool compMinOptsIsSet;
7866 bool compMinOptsIsUsed;
7868 inline bool MinOpts()
7870 assert(compMinOptsIsSet);
7871 compMinOptsIsUsed = true;
7874 inline bool IsMinOptsSet()
7876 return compMinOptsIsSet;
7879 inline bool MinOpts()
7883 inline bool IsMinOptsSet()
7885 return compMinOptsIsSet;
7888 inline void SetMinOpts(bool val)
7890 assert(!compMinOptsIsUsed);
7891 assert(!compMinOptsIsSet || (compMinOpts == val));
7893 compMinOptsIsSet = true;
7896 // true if the CLFLG_* for an optimization is set.
7897 inline bool OptEnabled(unsigned optFlag)
7899 return !!(compFlags & optFlag);
7902 #ifdef FEATURE_READYTORUN_COMPILER
7903 inline bool IsReadyToRun()
7905 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7908 inline bool IsReadyToRun()
7914 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7915 // PInvoke transitions inline (e.g. when targeting CoreRT).
7916 inline bool ShouldUsePInvokeHelpers()
7918 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7921 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7923 inline bool IsReversePInvoke()
7925 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7928 // true if we must generate code compatible with JIT32 quirks
7929 inline bool IsJit32Compat()
7931 #if defined(_TARGET_X86_)
7932 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7938 // true if we must generate code compatible with Jit64 quirks
7939 inline bool IsJit64Compat()
7941 #if defined(_TARGET_AMD64_)
7942 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7943 #elif !defined(FEATURE_CORECLR)
7950 bool compScopeInfo; // Generate the LocalVar info ?
7951 bool compDbgCode; // Generate debugger-friendly code?
7952 bool compDbgInfo; // Gather debugging info?
7955 #ifdef PROFILING_SUPPORTED
7956 bool compNoPInvokeInlineCB;
7958 static const bool compNoPInvokeInlineCB;
7962 bool compGcChecks; // Check arguments and return values to ensure they are sane
7963 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7964 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7968 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7969 // to be allocated on the stack.
7970 // It will be set to true in the following cases:
7971 // 1. When the method being compiled has a declarative security
7972 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7973 // This is also the case when we inject a prolog and epilog in the method.
7975 // 2. When the method being compiled has imperative security (i.e. the method
7976 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7978 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7980 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7981 // which gets reported as a GC root to stackwalker.
7982 // (See also ICodeManager::GetAddrOfSecurityObject.)
7987 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7988 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7992 #ifdef UNIX_AMD64_ABI
7993 // This flag is indicating if there is a need to align the frame.
7994 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7995 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7996 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7997 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7998 // there are calls and making sure the frame alignment logic is executed.
7999 bool compNeedToAlignFrame;
8000 #endif // UNIX_AMD64_ABI
8002 bool compProcedureSplitting; // Separate cold code from hot code
8004 bool genFPorder; // Preserve FP order (operations are non-commutative)
8005 bool genFPopt; // Can we do frame-pointer-omission optimization?
8006 bool altJit; // True if we are an altjit and are compiling this method
8009 bool optRepeat; // Repeat optimizer phases k times
8013 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8014 bool dspCode; // Display native code generated
8015 bool dspEHTable; // Display the EH table reported to the VM
8016 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8017 bool dspEmit; // Display emitter output
8018 bool dspLines; // Display source-code lines intermixed with native code output
8019 bool dmpHex; // Display raw bytes in hex of native code output
8020 bool varNames; // Display variables names in native code output
8021 bool disAsm; // Display native code as it is generated
8022 bool disAsmSpilled; // Display native code when any register spilling occurs
8023 bool disDiffable; // Makes the Disassembly code 'diff-able'
8024 bool disAsm2; // Display native code after it is generated using external disassembler
8025 bool dspOrder; // Display names of each of the methods that we ngen/jit
8026 bool dspUnwind; // Display the unwind info output
8027 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8028 bool compLongAddress; // Force using large pseudo instructions for long address
8029 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8030 bool dspGCtbls; // Display the GC tables
8034 bool doLateDisasm; // Run the late disassembler
8035 #endif // LATE_DISASM
8037 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8038 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8039 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8040 static const bool dspGCtbls = true;
8043 // We need stack probes to guarantee that we won't trigger a stack overflow
8044 // when calling unmanaged code until they get a chance to set up a frame, because
8045 // the EE will have no idea where it is.
8047 // We will only be doing this currently for hosted environments. Unfortunately
8048 // we need to take care of stubs, so potentially, we will have to do the probes
8049 // for any call. We have a plan for not needing for stubs though
8050 bool compNeedStackProbes;
8052 #ifdef PROFILING_SUPPORTED
8053 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8054 // This option helps make the JIT behave as if it is running under a profiler.
8055 bool compJitELTHookEnabled;
8056 #endif // PROFILING_SUPPORTED
8058 #if FEATURE_TAILCALL_OPT
8059 // Whether opportunistic or implicit tail call optimization is enabled.
8060 bool compTailCallOpt;
8061 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8062 bool compTailCallLoopOpt;
8066 static const bool compUseSoftFP = true;
8067 #else // !ARM_SOFTFP
8068 static const bool compUseSoftFP = false;
8071 GCPollType compGCPollType;
8075 static bool s_pAltJitExcludeAssembliesListInitialized;
8076 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8081 template <typename T>
8084 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8087 template <typename T>
8090 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8093 static int dspTreeID(GenTree* tree)
8095 return tree->gtTreeID;
8097 static void printTreeID(GenTree* tree)
8099 if (tree == nullptr)
8105 printf("[%06d]", dspTreeID(tree));
8112 #define STRESS_MODES \
8116 /* "Variations" stress areas which we try to mix up with each other. */ \
8117 /* These should not be exhaustively used as they might */ \
8118 /* hide/trivialize other areas */ \
8121 STRESS_MODE(DBL_ALN) \
8122 STRESS_MODE(LCL_FLDS) \
8123 STRESS_MODE(UNROLL_LOOPS) \
8124 STRESS_MODE(MAKE_CSE) \
8125 STRESS_MODE(LEGACY_INLINE) \
8126 STRESS_MODE(CLONE_EXPR) \
8127 STRESS_MODE(USE_FCOMI) \
8128 STRESS_MODE(USE_CMOV) \
8130 STRESS_MODE(BB_PROFILE) \
8131 STRESS_MODE(OPT_BOOLS_GC) \
8132 STRESS_MODE(REMORPH_TREES) \
8133 STRESS_MODE(64RSLT_MUL) \
8134 STRESS_MODE(DO_WHILE_LOOPS) \
8135 STRESS_MODE(MIN_OPTS) \
8136 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8137 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8138 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8139 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8140 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8141 STRESS_MODE(NULL_OBJECT_CHECK) \
8142 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8143 STRESS_MODE(RANDOM_INLINE) \
8144 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8145 STRESS_MODE(GENERIC_VARN) \
8147 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8149 STRESS_MODE(COUNT_VARN) \
8151 /* "Check" stress areas that can be exhaustively used if we */ \
8152 /* dont care about performance at all */ \
8154 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8155 STRESS_MODE(CHK_FLOW_UPDATE) \
8156 STRESS_MODE(EMITTER) \
8157 STRESS_MODE(CHK_REIMPORT) \
8158 STRESS_MODE(FLATFP) \
8159 STRESS_MODE(GENERIC_CHECK) \
8164 #define STRESS_MODE(mode) STRESS_##mode,
8171 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8172 BYTE compActiveStressModes[STRESS_COUNT];
8175 #define MAX_STRESS_WEIGHT 100
8177 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8181 bool compInlineStress()
8183 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8186 bool compRandomInlineStress()
8188 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8193 bool compTailCallStress()
8196 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8202 codeOptimize compCodeOpt()
8205 // Switching between size & speed has measurable throughput impact
8206 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8207 // DEBUG, but should generate identical code between CHK & RET builds,
8208 // so that's not acceptable.
8209 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8210 // Investigate the cause of the throughput regression.
8212 return opts.compCodeOpt;
8214 return BLENDED_CODE;
8218 //--------------------- Info about the procedure --------------------------
8222 COMP_HANDLE compCompHnd;
8223 CORINFO_MODULE_HANDLE compScopeHnd;
8224 CORINFO_CLASS_HANDLE compClassHnd;
8225 CORINFO_METHOD_HANDLE compMethodHnd;
8226 CORINFO_METHOD_INFO* compMethodInfo;
8228 BOOL hasCircularClassConstraints;
8229 BOOL hasCircularMethodConstraints;
8231 #if defined(DEBUG) || defined(LATE_DISASM)
8232 const char* compMethodName;
8233 const char* compClassName;
8234 const char* compFullName;
8235 #endif // defined(DEBUG) || defined(LATE_DISASM)
8237 #if defined(DEBUG) || defined(INLINE_DATA)
8238 // Method hash is logcally const, but computed
8240 mutable unsigned compMethodHashPrivate;
8241 unsigned compMethodHash() const;
8242 #endif // defined(DEBUG) || defined(INLINE_DATA)
8244 #ifdef PSEUDORANDOM_NOP_INSERTION
8245 // things for pseudorandom nop insertion
8246 unsigned compChecksum;
8250 // The following holds the FLG_xxxx flags for the method we're compiling.
8253 // The following holds the class attributes for the method we're compiling.
8254 unsigned compClassAttr;
8256 const BYTE* compCode;
8257 IL_OFFSET compILCodeSize; // The IL code size
8258 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8259 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8260 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8261 // (2) the code is hot/cold split, and we issued less code than we expected
8262 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8264 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8265 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8266 bool compIsContextful : 1; // contextful method
8267 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8268 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8269 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8270 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8271 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8273 var_types compRetType; // Return type of the method as declared in IL
8274 var_types compRetNativeType; // Normalized return type as per target arch ABI
8275 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8276 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8277 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8278 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8279 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8280 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8281 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8282 unsigned compMaxStack;
8283 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8284 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8286 unsigned compCallUnmanaged; // count of unmanaged calls
8287 unsigned compLvFrameListRoot; // lclNum for the Frame root
8288 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8289 // You should generally use compHndBBtabCount instead: it is the
8290 // current number of EH clauses (after additions like synchronized
8291 // methods and funclets, and removals like unreachable code deletion).
8293 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8294 // and the VM expects that, or the JIT is a "self-host" compiler
8295 // (e.g., x86 hosted targeting x86) and the VM expects that.
8297 /* The following holds IL scope information about local variables.
8300 unsigned compVarScopesCount;
8301 VarScopeDsc* compVarScopes;
8303 /* The following holds information about instr offsets for
8304 * which we need to report IP-mappings
8307 IL_OFFSET* compStmtOffsets; // sorted
8308 unsigned compStmtOffsetsCount;
8309 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8311 #define CPU_X86 0x0100 // The generic X86 CPU
8312 #define CPU_X86_PENTIUM_4 0x0110
8314 #define CPU_X64 0x0200 // The generic x64 CPU
8315 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8316 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8318 #define CPU_ARM 0x0300 // The generic ARM CPU
8320 unsigned genCPU; // What CPU are we running on
8323 // Returns true if the method being compiled returns a non-void and non-struct value.
8324 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8325 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8326 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8327 // Methods returning such structs are considered to return non-struct return value and
8328 // this method returns true in that case.
8329 bool compMethodReturnsNativeScalarType()
8331 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8334 // Returns true if the method being compiled returns RetBuf addr as its return value
8335 bool compMethodReturnsRetBufAddr()
8337 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8338 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8340 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8341 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8342 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8343 // methods with hidden RetBufArg.
8345 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8346 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8347 // returning the address of RetBuf.
8349 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8350 // to be returned in RAX.
8351 CLANG_FORMAT_COMMENT_ANCHOR;
8353 #ifdef _TARGET_AMD64_
8354 return (info.compRetBuffArg != BAD_VAR_NUM);
8355 #else // !_TARGET_AMD64_
8356 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8357 #endif // !_TARGET_AMD64_
8360 // Returns true if the method returns a value in more than one return register
8361 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8362 // TODO-ARM64: Does this apply for ARM64 too?
8363 bool compMethodReturnsMultiRegRetType()
8365 #if FEATURE_MULTIREG_RET
8366 #if defined(_TARGET_X86_)
8367 // On x86 only 64-bit longs are returned in multiple registers
8368 return varTypeIsLong(info.compRetNativeType);
8369 #else // targets: X64-UNIX, ARM64 or ARM32
8370 // On all other targets that support multireg return values:
8371 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8372 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8373 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8374 #endif // TARGET_XXX
8376 #else // not FEATURE_MULTIREG_RET
8378 // For this architecture there are no multireg returns
8381 #endif // FEATURE_MULTIREG_RET
8384 #if FEATURE_MULTIREG_ARGS
8385 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8386 // return the gcPtr layout for the pointers sized fields
8387 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8388 #endif // FEATURE_MULTIREG_ARGS
8390 // Returns true if the method being compiled returns a value
8391 bool compMethodHasRetVal()
8393 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8394 compMethodReturnsMultiRegRetType();
8399 void compDispLocalVars();
8403 //-------------------------- Global Compiler Data ------------------------------------
8406 static unsigned s_compMethodsCount; // to produce unique label names
8407 unsigned compGenTreeID;
8410 BasicBlock* compCurBB; // the current basic block in process
8411 GenTreePtr compCurStmt; // the current statement in process
8413 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8416 // The following is used to create the 'method JIT info' block.
8417 size_t compInfoBlkSize;
8418 BYTE* compInfoBlkAddr;
8420 EHblkDsc* compHndBBtab; // array of EH data
8421 unsigned compHndBBtabCount; // element count of used elements in EH data array
8422 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8424 #if defined(_TARGET_X86_)
8426 //-------------------------------------------------------------------------
8427 // Tracking of region covered by the monitor in synchronized methods
8428 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8429 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8431 #endif // !_TARGET_X86_
8433 Phases previousCompletedPhase; // the most recently completed phase
8435 //-------------------------------------------------------------------------
8436 // The following keeps track of how many bytes of local frame space we've
8437 // grabbed so far in the current function, and how many argument bytes we
8438 // need to pop when we return.
8441 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8443 // Count of callee-saved regs we pushed in the prolog.
8444 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8445 // In case of Amd64 this doesn't include float regs saved on stack.
8446 unsigned compCalleeRegsPushed;
8448 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8449 // Mask of callee saved float regs on stack.
8450 regMaskTP compCalleeFPRegsSavedMask;
8452 #ifdef _TARGET_AMD64_
8453 // Quirk for VS debug-launch scenario to work:
8454 // Bytes of padding between save-reg area and locals.
8455 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8456 unsigned compVSQuirkStackPaddingNeeded;
8457 bool compQuirkForPPPflag;
8460 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8462 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8463 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8464 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8466 //-------------------------------------------------------------------------
8468 static void compStartup(); // One-time initialization
8469 static void compShutdown(); // One-time finalization
8471 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8474 static void compDisplayStaticSizes(FILE* fout);
8476 //------------ Some utility functions --------------
8478 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8479 void** ppIndirection); /* OUT */
8481 // Several JIT/EE interface functions return a CorInfoType, and also return a
8482 // class handle as an out parameter if the type is a value class. Returns the
8483 // size of the type these describe.
8484 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8487 // Components used by the compiler may write unit test suites, and
8488 // have them run within this method. They will be run only once per process, and only
8489 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8490 // These should fail by asserting.
8491 void compDoComponentUnitTestsOnce();
8494 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8495 CORINFO_MODULE_HANDLE classPtr,
8496 COMP_HANDLE compHnd,
8497 CORINFO_METHOD_INFO* methodInfo,
8498 void** methodCodePtr,
8499 ULONG* methodCodeSize,
8500 JitFlags* compileFlags);
8501 void compCompileFinish();
8502 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8503 COMP_HANDLE compHnd,
8504 CORINFO_METHOD_INFO* methodInfo,
8505 void** methodCodePtr,
8506 ULONG* methodCodeSize,
8507 JitFlags* compileFlags,
8508 CorInfoInstantiationVerification instVerInfo);
8510 ArenaAllocator* compGetAllocator();
8512 #if MEASURE_MEM_ALLOC
8514 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8518 unsigned allocCnt; // # of allocs
8519 UINT64 allocSz; // total size of those alloc.
8520 UINT64 allocSzMax; // Maximum single allocation.
8521 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8522 UINT64 nraTotalSizeAlloc;
8523 UINT64 nraTotalSizeUsed;
8525 static const char* s_CompMemKindNames[]; // Names of the kinds.
8527 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8529 for (int i = 0; i < CMK_Count; i++)
8531 allocSzByKind[i] = 0;
8534 MemStats(const MemStats& ms)
8535 : allocCnt(ms.allocCnt)
8536 , allocSz(ms.allocSz)
8537 , allocSzMax(ms.allocSzMax)
8538 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8539 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8541 for (int i = 0; i < CMK_Count; i++)
8543 allocSzByKind[i] = ms.allocSzByKind[i];
8547 // Until we have ubiquitous constructors.
8550 this->MemStats::MemStats();
8553 void AddAlloc(size_t sz, CompMemKind cmk)
8557 if (sz > allocSzMax)
8561 allocSzByKind[cmk] += sz;
8564 void Print(FILE* f); // Print these stats to f.
8565 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8567 MemStats genMemStats;
8569 struct AggregateMemStats : public MemStats
8573 AggregateMemStats() : MemStats(), nMethods(0)
8577 void Add(const MemStats& ms)
8580 allocCnt += ms.allocCnt;
8581 allocSz += ms.allocSz;
8582 allocSzMax = max(allocSzMax, ms.allocSzMax);
8583 for (int i = 0; i < CMK_Count; i++)
8585 allocSzByKind[i] += ms.allocSzByKind[i];
8587 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8588 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8591 void Print(FILE* f); // Print these stats to jitstdout.
8594 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8595 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8596 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8598 #endif // MEASURE_MEM_ALLOC
8600 #if LOOP_HOIST_STATS
8601 unsigned m_loopsConsidered;
8602 bool m_curLoopHasHoistedExpression;
8603 unsigned m_loopsWithHoistedExpressions;
8604 unsigned m_totalHoistedExpressions;
8606 void AddLoopHoistStats();
8607 void PrintPerMethodLoopHoistStats();
8609 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8610 static unsigned s_loopsConsidered;
8611 static unsigned s_loopsWithHoistedExpressions;
8612 static unsigned s_totalHoistedExpressions;
8614 static void PrintAggregateLoopHoistStats(FILE* f);
8615 #endif // LOOP_HOIST_STATS
8617 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8618 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8619 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8620 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8621 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8622 void compFreeMem(void*);
8624 bool compIsForImportOnly();
8625 bool compIsForInlining();
8626 bool compDonotInline();
8629 const char* compLocalVarName(unsigned varNum, unsigned offs);
8630 VarName compVarName(regNumber reg, bool isFloatReg = false);
8631 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8632 const char* compRegPairName(regPairNo regPair);
8633 const char* compRegNameForSize(regNumber reg, size_t size);
8634 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8635 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8636 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8639 //-------------------------------------------------------------------------
8641 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8643 struct VarScopeMapInfo
8645 VarScopeListNode* head;
8646 VarScopeListNode* tail;
8647 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8649 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8656 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8657 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8659 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8660 VarNumToScopeDscMap;
8662 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8663 VarNumToScopeDscMap* compVarScopeMap;
8665 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8667 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8669 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8671 void compInitVarScopeMap();
8673 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8674 // enter scope, sorted by instr offset
8675 unsigned compNextEnterScope;
8677 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8678 // go out of scope, sorted by instr offset
8679 unsigned compNextExitScope;
8681 void compInitScopeLists();
8683 void compResetScopeLists();
8685 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8687 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8689 void compProcessScopesUntil(unsigned offset,
8691 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8692 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8695 void compDispScopeLists();
8698 bool compIsProfilerHookNeeded();
8700 //-------------------------------------------------------------------------
8701 /* Statistical Data Gathering */
8703 void compJitStats(); // call this function and enable
8704 // various ifdef's below for statistical data
8707 void compCallArgStats();
8708 static void compDispCallArgStats(FILE* fout);
8711 //-------------------------------------------------------------------------
8718 ArenaAllocator* compAllocator;
8721 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8722 // suitable for use by utilcode collection types.
8723 IAllocator* compAsIAllocator;
8725 #if MEASURE_MEM_ALLOC
8726 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8727 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8728 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8730 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8732 #endif // MEASURE_MEM_ALLOC
8734 void compFunctionTraceStart();
8735 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8738 size_t compMaxUncheckedOffsetForNullObject;
8740 void compInitOptions(JitFlags* compileFlags);
8742 void compSetProcessor();
8743 void compInitDebuggingInfo();
8744 void compSetOptimizationLevel();
8745 #ifdef _TARGET_ARMARCH_
8746 bool compRsvdRegCheck(FrameLayoutState curState);
8748 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8750 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8751 void ResetOptAnnotations();
8753 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8754 void RecomputeLoopInfo();
8756 #ifdef PROFILING_SUPPORTED
8757 // Data required for generating profiler Enter/Leave/TailCall hooks
8759 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8760 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8761 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8764 #ifdef _TARGET_AMD64_
8765 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8768 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8769 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8771 IAllocator* getAllocator()
8773 return compAsIAllocator;
8776 #if MEASURE_MEM_ALLOC
8777 IAllocator* getAllocatorBitset()
8779 return compAsIAllocatorBitset;
8781 IAllocator* getAllocatorGC()
8783 return compAsIAllocatorGC;
8785 IAllocator* getAllocatorLoopHoist()
8787 return compAsIAllocatorLoopHoist;
8789 #else // !MEASURE_MEM_ALLOC
8790 IAllocator* getAllocatorBitset()
8792 return compAsIAllocator;
8794 IAllocator* getAllocatorGC()
8796 return compAsIAllocator;
8798 IAllocator* getAllocatorLoopHoist()
8800 return compAsIAllocator;
8802 #endif // !MEASURE_MEM_ALLOC
8805 IAllocator* getAllocatorDebugOnly()
8807 #if MEASURE_MEM_ALLOC
8808 return compAsIAllocatorDebugOnly;
8809 #else // !MEASURE_MEM_ALLOC
8810 return compAsIAllocator;
8811 #endif // !MEASURE_MEM_ALLOC
8816 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8817 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8821 XX Checks for type compatibility and merges types XX
8823 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8824 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8828 // Set to TRUE if verification cannot be skipped for this method
8829 // If we detect unverifiable code, we will lazily check
8830 // canSkipMethodVerification() to see if verification is REALLY needed.
8831 BOOL tiVerificationNeeded;
8833 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8834 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8835 BOOL tiIsVerifiableCode;
8837 // Set to TRUE if runtime callout is needed for this method
8838 BOOL tiRuntimeCalloutNeeded;
8840 // Set to TRUE if security prolog/epilog callout is needed for this method
8841 // Note: This flag is different than compNeedSecurityCheck.
8842 // compNeedSecurityCheck means whether or not a security object needs
8843 // to be allocated on the stack, which is currently true for EnC as well.
8844 // tiSecurityCalloutNeeded means whether or not security callouts need
8845 // to be inserted in the jitted code.
8846 BOOL tiSecurityCalloutNeeded;
8848 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8849 // This support is necessary to suport attributes that are not described in
8850 // for example, signatures. For example, the permanent home byref (byref that
8851 // points to the gc heap), isn't a property of method signatures, therefore,
8852 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8853 // but when deciding if we need to reimport a block, we need to take these
8855 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8857 // Returns TRUE if child is equal to or a subtype of parent.
8858 // normalisedForStack indicates that both types are normalised for the stack
8859 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8861 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8862 // *pDest is modified to represent the merged type. Sets "*changed" to true
8863 // if this changes "*pDest".
8864 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8866 // Set pDest from the primitive value type.
8867 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8869 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8872 // <BUGNUM> VSW 471305
8873 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8874 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8875 // We use a "short" as we need to push/pop this scope.
8877 short compRegSetCheckLevel;
8881 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8882 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8884 XX IL verification stuff XX
8887 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8888 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8892 // The following is used to track liveness of local variables, initialization
8893 // of valueclass constructors, and type safe use of IL instructions.
8895 // dynamic state info needed for verification
8896 EntryState verCurrentState;
8898 // this ptr of object type .ctors are considered intited only after
8899 // the base class ctor is called, or an alternate ctor is called.
8900 // An uninited this ptr can be used to access fields, but cannot
8901 // be used to call a member function.
8902 BOOL verTrackObjCtorInitState;
8904 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8906 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8907 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8908 void verInitCurrentState();
8909 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8911 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8912 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8913 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8915 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8916 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8917 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8918 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8919 typeInfo verMakeTypeInfo(CorInfoType ciType,
8920 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8921 BOOL verIsSDArray(typeInfo ti);
8922 typeInfo verGetArrayElemType(typeInfo ti);
8924 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8925 BOOL verNeedsVerification();
8926 BOOL verIsByRefLike(const typeInfo& ti);
8927 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8929 // generic type variables range over types that satisfy IsBoxable
8930 BOOL verIsBoxable(const typeInfo& ti);
8932 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8933 DEBUGARG(unsigned line));
8934 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8935 DEBUGARG(unsigned line));
8936 bool verCheckTailCallConstraint(OPCODE opcode,
8937 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8938 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8939 // on a type parameter?
8940 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8941 // return false to the caller.
8942 // If false, it will throw.
8944 bool verIsBoxedValueType(typeInfo ti);
8946 void verVerifyCall(OPCODE opcode,
8947 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8948 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8950 bool readonlyCall, // is this a "readonly." call?
8951 const BYTE* delegateCreateStart,
8952 const BYTE* codeAddr,
8953 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8955 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8957 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8958 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8959 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8960 const CORINFO_FIELD_INFO& fieldInfo,
8961 const typeInfo* tiThis,
8963 BOOL allowPlainStructAsThis = FALSE);
8964 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8965 void verVerifyThisPtrInitialised();
8966 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8968 // Register allocator
8969 void raInitStackFP();
8970 void raEnregisterVarsPrePassStackFP();
8971 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8972 void raEnregisterVarsPostPassStackFP();
8973 void raGenerateFPRefCounts();
8974 void raEnregisterVarsStackFP();
8975 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8977 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8978 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8980 // returns true if enregistering v1 would save more mem accesses than v2
8981 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8984 void raDumpHeightsStackFP();
8985 void raDumpVariableRegIntfFloat();
8988 #if FEATURE_STACK_FP_X87
8990 // Currently, we use FP transition blocks in only 2 situations:
8992 // -conditional jump on longs where FP stack differs with target: it's not strictly
8993 // necessary, but its low frequency and the code would get complicated if we try to
8994 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8995 // minimize the way we generate the jump code.
8996 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8997 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8999 // However, transition blocks have 2 problems
9001 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9002 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9003 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9004 // in the right place without preordering them), this causes us to have to generate the transition
9005 // blocks in the cold area if we want procedure splitting.
9008 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9009 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9010 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9011 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9012 // a big change in the exception.
9014 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9015 // optimizations. For these 2 cases:
9017 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9018 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9019 // a switch statement.
9021 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9022 // current procedure splitting and exception code have.
9023 bool compMayHaveTransitionBlocks;
9025 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9027 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9029 unsigned raCntStkStackFP;
9030 unsigned raCntWtdStkDblStackFP;
9031 unsigned raCntStkParamDblStackFP;
9033 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9034 // TODO: Do we want to put this in LclVarDsc?
9035 unsigned raPayloadStackFP[lclMAX_TRACKED];
9036 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9038 // Useful for debugging
9039 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9041 #endif // FEATURE_STACK_FP_X87
9044 // One line log function. Default level is 0. Increasing it gives you
9045 // more log information
9047 // levels are currently unused: #define JITDUMP(level,...) ();
9048 void JitLogEE(unsigned level, const char* fmt, ...);
9050 bool compDebugBreak;
9052 bool compJitHaltMethod();
9057 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9058 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9060 XX GS Security checks for unsafe buffers XX
9062 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9063 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9066 struct ShadowParamVarInfo
9068 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9069 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9071 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9073 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9074 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9075 // slots and update all trees to refer to shadow slots is done immediately after
9076 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9077 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9078 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9079 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9080 // creating a shadow slot even though this routine returns true.
9082 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9083 // required. There are two cases under which a reg arg could potentially be used from its
9085 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9086 // b) LSRA spills it
9088 // Possible solution to address case (a)
9089 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9090 // in this routine. Note that live out of exception handler is something we may not be
9091 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9092 // Therefore, for methods with exception handling and need GS cookie check we might have
9093 // to take conservative approach.
9095 // Possible solution to address case (b)
9096 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9097 // create a new spill temp if the method needs GS cookie check.
9098 return varDsc->lvIsParam;
9099 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9100 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9107 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9112 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9113 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9114 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9116 void gsGSChecksInitCookie(); // Grabs cookie variable
9117 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9118 bool gsFindVulnerableParams(); // Shadow param analysis code
9119 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9121 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9122 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9124 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9125 // This can be overwritten by setting complus_JITInlineSize env variable.
9127 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9130 #ifdef FEATURE_JIT_METHOD_PERF
9131 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9132 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9134 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9135 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9137 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9139 #if MEASURE_CLRAPI_CALLS
9140 // Thin wrappers that call into JitTimer (if present).
9141 inline void CLRApiCallEnter(unsigned apix);
9142 inline void CLRApiCallLeave(unsigned apix);
9145 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9146 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9151 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9152 // These variables are associated with maintaining SQM data about compile time.
9153 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9154 // in the current compilation.
9155 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9156 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9157 // the inlining phase in the current compilation.
9158 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9160 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9161 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9162 // type-loading and class initialization).
9163 void RecordStateAtEndOfInlining();
9164 // Assumes being called at the end of compilation. Update the SQM state.
9165 void RecordStateAtEndOfCompilation();
9167 #ifdef FEATURE_CLRSQM
9168 // Does anything SQM related necessary at process shutdown time.
9169 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9170 #endif // FEATURE_CLRSQM
9173 #if FUNC_INFO_LOGGING
9174 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9175 // filename to write it to.
9176 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9177 #endif // FUNC_INFO_LOGGING
9179 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9181 // Is the compilation in a full trust context?
9182 bool compIsFullTrust();
9185 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9186 #endif // MEASURE_NOWAY
9188 #ifndef FEATURE_TRACELOGGING
9189 // Should we actually fire the noway assert body and the exception handler?
9190 bool compShouldThrowOnNoway();
9191 #else // FEATURE_TRACELOGGING
9192 // Should we actually fire the noway assert body and the exception handler?
9193 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9195 // Telemetry instance to use per method compilation.
9196 JitTelemetry compJitTelemetry;
9198 // Get common parameters that have to be logged with most telemetry data.
9199 void compGetTelemetryDefaults(const char** assemblyName,
9200 const char** scopeName,
9201 const char** methodName,
9202 unsigned* methodHash);
9203 #endif // !FEATURE_TRACELOGGING
9207 NodeToTestDataMap* m_nodeTestData;
9209 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9210 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9211 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9212 // Current kept in this.
9214 NodeToTestDataMap* GetNodeTestData()
9216 Compiler* compRoot = impInlineRoot();
9217 if (compRoot->m_nodeTestData == nullptr)
9219 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9221 return compRoot->m_nodeTestData;
9224 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9226 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9227 // currently occur in the AST graph.
9228 NodeToIntMap* FindReachableNodesInNodeTestData();
9230 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9231 // test data, associate that data with "to".
9232 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9234 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9235 // have annotations, attach similar annotations to the corresponding nodes in "to".
9236 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9238 // These are the methods that test that the various conditions implied by the
9239 // test attributes are satisfied.
9240 void JitTestCheckSSA(); // SSA builder tests.
9241 void JitTestCheckVN(); // Value numbering tests.
9244 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9246 FieldSeqStore* m_fieldSeqStore;
9248 FieldSeqStore* GetFieldSeqStore()
9250 Compiler* compRoot = impInlineRoot();
9251 if (compRoot->m_fieldSeqStore == nullptr)
9253 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9254 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9255 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9257 return compRoot->m_fieldSeqStore;
9260 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9262 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9263 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9264 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9265 // attach the field sequence directly to the address node.
9266 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9268 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9270 // Don't need to worry about inlining here
9271 if (m_zeroOffsetFieldMap == nullptr)
9273 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9275 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9276 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9278 return m_zeroOffsetFieldMap;
9281 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9282 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9283 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9284 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9285 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9286 // record the the field sequence using the ZeroOffsetFieldMap described above.
9288 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9289 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9290 // CoreRT. Such case is handled same as the default case.
9291 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9293 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9295 NodeToArrayInfoMap* m_arrayInfoMap;
9297 NodeToArrayInfoMap* GetArrayInfoMap()
9299 Compiler* compRoot = impInlineRoot();
9300 if (compRoot->m_arrayInfoMap == nullptr)
9302 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9303 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9304 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9306 return compRoot->m_arrayInfoMap;
9309 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9311 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9312 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9313 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9314 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9316 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9318 // Use the same map for GCHeap and ByrefExposed when their states match.
9319 memoryKind = ByrefExposed;
9322 assert(memoryKind < MemoryKindCount);
9323 Compiler* compRoot = impInlineRoot();
9324 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9326 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9327 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9328 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9330 return compRoot->m_memorySsaMap[memoryKind];
9333 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9334 CORINFO_CLASS_HANDLE m_refAnyClass;
9335 CORINFO_FIELD_HANDLE GetRefanyDataField()
9337 if (m_refAnyClass == nullptr)
9339 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9341 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9343 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9345 if (m_refAnyClass == nullptr)
9347 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9349 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9353 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9355 #if ALLVARSET_COUNTOPS
9356 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9359 static HelperCallProperties s_helperCallProperties;
9361 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9362 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9363 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9365 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9368 unsigned __int8* offset0,
9369 unsigned __int8* offset1);
9370 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9371 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9373 void fgMorphMultiregStructArgs(GenTreeCall* call);
9374 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9376 }; // end of class Compiler
9378 // Inline methods of CompAllocator.
9379 void* CompAllocator::Alloc(size_t sz)
9381 #if MEASURE_MEM_ALLOC
9382 return m_comp->compGetMem(sz, m_cmk);
9384 return m_comp->compGetMem(sz);
9388 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9390 #if MEASURE_MEM_ALLOC
9391 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9393 return m_comp->compGetMemArray(elems, elemSize);
9397 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9398 inline LclVarDsc::LclVarDsc(Compiler* comp)
9399 : // Initialize the ArgRegs to REG_STK.
9400 // The morph will do the right thing to change
9401 // to the right register if passed in register.
9404 #if FEATURE_MULTIREG_ARGS
9405 _lvOtherArgReg(REG_STK)
9407 #endif // FEATURE_MULTIREG_ARGS
9409 lvRefBlks(BlockSetOps::UninitVal())
9411 #endif // ASSERTION_PROP
9412 lvPerSsaData(comp->getAllocator())
9416 //---------------------------------------------------------------------------------------------------------------------
9417 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9419 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9420 // shown in parentheses):
9422 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9423 // of a misnomer, as the first entry will always be the current node.
9425 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9426 // argument before visiting the node's operands.
9428 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9429 // argument after visiting the node's operands.
9431 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9432 // `DoPreOrder` must be true if this option is true.
9434 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9435 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9436 // visited before the first).
9438 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9440 // A simple pre-order visitor might look something like the following:
9442 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9447 // DoPreOrder = true
9450 // unsigned m_count;
9452 // CountingVisitor(Compiler* compiler)
9453 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9457 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9463 // This visitor would then be used like so:
9465 // CountingVisitor countingVisitor(compiler);
9466 // countingVisitor.WalkTree(root);
9468 template <typename TVisitor>
9469 class GenTreeVisitor
9472 typedef Compiler::fgWalkResult fgWalkResult;
9476 ComputeStack = false,
9478 DoPostOrder = false,
9479 DoLclVarsOnly = false,
9480 UseExecutionOrder = false,
9483 Compiler* m_compiler;
9484 ArrayStack<GenTree*> m_ancestors;
9486 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9488 assert(compiler != nullptr);
9490 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9491 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9494 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9496 return fgWalkResult::WALK_CONTINUE;
9499 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9501 return fgWalkResult::WALK_CONTINUE;
9505 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9507 assert(use != nullptr);
9509 GenTree* node = *use;
9511 if (TVisitor::ComputeStack)
9513 m_ancestors.Push(node);
9516 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9517 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9519 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9520 if (result == fgWalkResult::WALK_ABORT)
9526 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9532 switch (node->OperGet())
9537 case GT_LCL_VAR_ADDR:
9538 case GT_LCL_FLD_ADDR:
9539 if (TVisitor::DoLclVarsOnly)
9541 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9542 if (result == fgWalkResult::WALK_ABORT)
9558 case GT_MEMORYBARRIER:
9563 case GT_START_NONGC:
9565 #if !FEATURE_EH_FUNCLETS
9567 #endif // !FEATURE_EH_FUNCLETS
9569 #ifndef LEGACY_BACKEND
9571 #endif // LEGACY_BACKEND
9574 case GT_CLS_VAR_ADDR:
9578 case GT_PINVOKE_PROLOG:
9579 case GT_PINVOKE_EPILOG:
9583 // Lclvar unary operators
9584 case GT_STORE_LCL_VAR:
9585 case GT_STORE_LCL_FLD:
9586 if (TVisitor::DoLclVarsOnly)
9588 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9589 if (result == fgWalkResult::WALK_ABORT)
9596 // Standard unary operators
9624 GenTreeUnOp* const unOp = node->AsUnOp();
9625 if (unOp->gtOp1 != nullptr)
9627 result = WalkTree(&unOp->gtOp1, unOp);
9628 if (result == fgWalkResult::WALK_ABORT)
9639 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9641 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9642 if (result == fgWalkResult::WALK_ABORT)
9646 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9647 if (result == fgWalkResult::WALK_ABORT)
9651 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9652 if (result == fgWalkResult::WALK_ABORT)
9659 case GT_ARR_BOUNDS_CHECK:
9662 #endif // FEATURE_SIMD
9664 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9666 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9667 if (result == fgWalkResult::WALK_ABORT)
9671 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9672 if (result == fgWalkResult::WALK_ABORT)
9681 GenTreeField* const field = node->AsField();
9683 if (field->gtFldObj != nullptr)
9685 result = WalkTree(&field->gtFldObj, field);
9686 if (result == fgWalkResult::WALK_ABORT)
9696 GenTreeArrElem* const arrElem = node->AsArrElem();
9698 result = WalkTree(&arrElem->gtArrObj, arrElem);
9699 if (result == fgWalkResult::WALK_ABORT)
9704 const unsigned rank = arrElem->gtArrRank;
9705 for (unsigned dim = 0; dim < rank; dim++)
9707 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9708 if (result == fgWalkResult::WALK_ABORT)
9718 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9720 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9721 if (result == fgWalkResult::WALK_ABORT)
9725 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9726 if (result == fgWalkResult::WALK_ABORT)
9730 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9731 if (result == fgWalkResult::WALK_ABORT)
9740 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9742 GenTree** op1Use = &dynBlock->gtOp1;
9743 GenTree** op2Use = &dynBlock->gtDynamicSize;
9745 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9747 std::swap(op1Use, op2Use);
9750 result = WalkTree(op1Use, dynBlock);
9751 if (result == fgWalkResult::WALK_ABORT)
9755 result = WalkTree(op2Use, dynBlock);
9756 if (result == fgWalkResult::WALK_ABORT)
9763 case GT_STORE_DYN_BLK:
9765 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9767 GenTree** op1Use = &dynBlock->gtOp1;
9768 GenTree** op2Use = &dynBlock->gtOp2;
9769 GenTree** op3Use = &dynBlock->gtDynamicSize;
9771 if (TVisitor::UseExecutionOrder)
9773 if (dynBlock->IsReverseOp())
9775 std::swap(op1Use, op2Use);
9777 if (dynBlock->gtEvalSizeFirst)
9779 std::swap(op3Use, op2Use);
9780 std::swap(op2Use, op1Use);
9784 result = WalkTree(op1Use, dynBlock);
9785 if (result == fgWalkResult::WALK_ABORT)
9789 result = WalkTree(op2Use, dynBlock);
9790 if (result == fgWalkResult::WALK_ABORT)
9794 result = WalkTree(op3Use, dynBlock);
9795 if (result == fgWalkResult::WALK_ABORT)
9804 GenTreeCall* const call = node->AsCall();
9806 if (call->gtCallObjp != nullptr)
9808 result = WalkTree(&call->gtCallObjp, call);
9809 if (result == fgWalkResult::WALK_ABORT)
9815 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9817 result = WalkTree(args->pCurrent(), call);
9818 if (result == fgWalkResult::WALK_ABORT)
9824 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9826 result = WalkTree(args->pCurrent(), call);
9827 if (result == fgWalkResult::WALK_ABORT)
9833 if (call->gtCallType == CT_INDIRECT)
9835 if (call->gtCallCookie != nullptr)
9837 result = WalkTree(&call->gtCallCookie, call);
9838 if (result == fgWalkResult::WALK_ABORT)
9844 result = WalkTree(&call->gtCallAddr, call);
9845 if (result == fgWalkResult::WALK_ABORT)
9851 if (call->gtControlExpr != nullptr)
9853 result = WalkTree(&call->gtControlExpr, call);
9854 if (result == fgWalkResult::WALK_ABORT)
9866 assert(node->OperIsBinary());
9868 GenTreeOp* const op = node->AsOp();
9870 GenTree** op1Use = &op->gtOp1;
9871 GenTree** op2Use = &op->gtOp2;
9873 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9875 std::swap(op1Use, op2Use);
9878 if (*op1Use != nullptr)
9880 result = WalkTree(op1Use, op);
9881 if (result == fgWalkResult::WALK_ABORT)
9887 if (*op2Use != nullptr)
9889 result = WalkTree(op2Use, op);
9890 if (result == fgWalkResult::WALK_ABORT)
9900 // Finally, visit the current node
9901 if (TVisitor::DoPostOrder)
9903 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9906 if (TVisitor::ComputeStack)
9915 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9916 class GenericTreeWalker final
9917 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
9922 ComputeStack = computeStack,
9923 DoPreOrder = doPreOrder,
9924 DoPostOrder = doPostOrder,
9925 DoLclVarsOnly = doLclVarsOnly,
9926 UseExecutionOrder = useExecutionOrder,
9930 Compiler::fgWalkData* m_walkData;
9933 GenericTreeWalker(Compiler::fgWalkData* walkData)
9934 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
9936 , m_walkData(walkData)
9938 assert(walkData != nullptr);
9942 walkData->parentStack = &this->m_ancestors;
9946 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9948 m_walkData->parent = user;
9949 return m_walkData->wtprVisitorFn(use, m_walkData);
9952 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9954 m_walkData->parent = user;
9955 return m_walkData->wtpoVisitorFn(use, m_walkData);
9959 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
9965 DoLclVarsOnly = true
9968 IncLclVarRefCountsVisitor(Compiler* compiler);
9969 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9971 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9974 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
9980 DoLclVarsOnly = true
9983 DecLclVarRefCountsVisitor(Compiler* compiler);
9984 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
9986 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
9990 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9991 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9993 XX Miscellaneous Compiler stuff XX
9995 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9996 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9999 // Values used to mark the types a stack slot is used for
10001 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10002 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10003 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10004 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10005 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10006 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10007 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10008 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10010 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10012 /*****************************************************************************
10014 * Variables to keep track of total code amounts.
10019 extern size_t grossVMsize;
10020 extern size_t grossNCsize;
10021 extern size_t totalNCsize;
10023 extern unsigned genMethodICnt;
10024 extern unsigned genMethodNCnt;
10025 extern size_t gcHeaderISize;
10026 extern size_t gcPtrMapISize;
10027 extern size_t gcHeaderNSize;
10028 extern size_t gcPtrMapNSize;
10030 #endif // DISPLAY_SIZES
10032 /*****************************************************************************
10034 * Variables to keep track of basic block counts (more data on 1 BB methods)
10037 #if COUNT_BASIC_BLOCKS
10038 extern Histogram bbCntTable;
10039 extern Histogram bbOneBBSizeTable;
10042 /*****************************************************************************
10044 * Used by optFindNaturalLoops to gather statistical information such as
10045 * - total number of natural loops
10046 * - number of loops with 1, 2, ... exit conditions
10047 * - number of loops that have an iterator (for like)
10048 * - number of loops that have a constant iterator
10053 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10054 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10055 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10056 extern unsigned totalLoopCount; // counts the total number of natural loops
10057 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10058 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10059 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10060 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10062 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10063 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10064 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10065 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10066 extern Histogram loopCountTable; // Histogram of loop counts
10067 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10069 #endif // COUNT_LOOPS
10071 /*****************************************************************************
10072 * variables to keep track of how many iterations we go in a dataflow pass
10077 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10078 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10080 #endif // DATAFLOW_ITER
10082 #if MEASURE_BLOCK_SIZE
10083 extern size_t genFlowNodeSize;
10084 extern size_t genFlowNodeCnt;
10085 #endif // MEASURE_BLOCK_SIZE
10087 #if MEASURE_NODE_SIZE
10088 struct NodeSizeStats
10092 genTreeNodeCnt = 0;
10093 genTreeNodeSize = 0;
10094 genTreeNodeActualSize = 0;
10097 // Count of tree nodes allocated.
10098 unsigned __int64 genTreeNodeCnt;
10100 // The size we allocate.
10101 unsigned __int64 genTreeNodeSize;
10103 // The actual size of the node. Note that the actual size will likely be smaller
10104 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10105 // a smaller node to a larger one. TODO-Cleanup: add stats on
10106 // SetOper()/ChangeOper() usage to quantify this.
10107 unsigned __int64 genTreeNodeActualSize;
10109 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10110 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10111 extern Histogram genTreeNcntHist;
10112 extern Histogram genTreeNsizHist;
10113 #endif // MEASURE_NODE_SIZE
10115 /*****************************************************************************
10116 * Count fatal errors (including noway_asserts).
10120 extern unsigned fatal_badCode;
10121 extern unsigned fatal_noWay;
10122 extern unsigned fatal_NOMEM;
10123 extern unsigned fatal_noWayAssertBody;
10125 extern unsigned fatal_noWayAssertBodyArgs;
10127 extern unsigned fatal_NYI;
10128 #endif // MEASURE_FATAL
10130 /*****************************************************************************
10134 #ifdef _TARGET_XARCH_
10136 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10137 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10138 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10140 const instruction INS_AND = INS_and;
10141 const instruction INS_OR = INS_or;
10142 const instruction INS_XOR = INS_xor;
10143 const instruction INS_NEG = INS_neg;
10144 const instruction INS_TEST = INS_test;
10145 const instruction INS_MUL = INS_imul;
10146 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10147 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10148 const instruction INS_BREAKPOINT = INS_int3;
10149 const instruction INS_ADDC = INS_adc;
10150 const instruction INS_SUBC = INS_sbb;
10151 const instruction INS_NOT = INS_not;
10155 #ifdef _TARGET_ARM_
10157 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10158 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10159 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10161 const instruction INS_AND = INS_and;
10162 const instruction INS_OR = INS_orr;
10163 const instruction INS_XOR = INS_eor;
10164 const instruction INS_NEG = INS_rsb;
10165 const instruction INS_TEST = INS_tst;
10166 const instruction INS_MUL = INS_mul;
10167 const instruction INS_MULADD = INS_mla;
10168 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10169 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10170 const instruction INS_BREAKPOINT = INS_bkpt;
10171 const instruction INS_ADDC = INS_adc;
10172 const instruction INS_SUBC = INS_sbc;
10173 const instruction INS_NOT = INS_mvn;
10175 const instruction INS_ABS = INS_vabs;
10176 const instruction INS_ROUND = INS_invalid;
10177 const instruction INS_SQRT = INS_vsqrt;
10181 #ifdef _TARGET_ARM64_
10183 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10184 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10185 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10187 const instruction INS_AND = INS_and;
10188 const instruction INS_OR = INS_orr;
10189 const instruction INS_XOR = INS_eor;
10190 const instruction INS_NEG = INS_neg;
10191 const instruction INS_TEST = INS_tst;
10192 const instruction INS_MUL = INS_mul;
10193 const instruction INS_MULADD = INS_madd;
10194 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10195 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10196 const instruction INS_BREAKPOINT = INS_bkpt;
10197 const instruction INS_ADDC = INS_adc;
10198 const instruction INS_SUBC = INS_sbc;
10199 const instruction INS_NOT = INS_mvn;
10201 const instruction INS_ABS = INS_fabs;
10202 const instruction INS_ROUND = INS_frintn;
10203 const instruction INS_SQRT = INS_fsqrt;
10207 /*****************************************************************************/
10209 extern const BYTE genTypeSizes[];
10210 extern const BYTE genTypeAlignments[];
10211 extern const BYTE genTypeStSzs[];
10212 extern const BYTE genActualTypes[];
10214 /*****************************************************************************/
10216 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10217 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10219 #ifdef _TARGET_ARM_
10220 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10221 #elif defined(_TARGET_ARM64_)
10222 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10225 /*****************************************************************************/
10227 #define REG_CORRUPT regNumber(REG_NA + 1)
10228 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10229 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10231 /*****************************************************************************/
10233 extern BasicBlock dummyBB;
10235 /*****************************************************************************/
10236 /*****************************************************************************/
10238 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10239 // nodes of a statement in execution order.
10240 // __stmt: a GT_STMT type GenTree*
10241 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10243 #define foreach_treenode_execution_order(__node, __stmt) \
10244 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10246 // foreach_block: An iterator over all blocks in the function.
10247 // __compiler: the Compiler* object
10248 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10250 #define foreach_block(__compiler, __block) \
10251 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10253 /*****************************************************************************/
10254 /*****************************************************************************/
10258 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10260 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10261 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10263 XX Debugging helpers XX
10265 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10266 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10269 /*****************************************************************************/
10270 /* The following functions are intended to be called from the debugger, to dump
10271 * various data structures. The can be used in the debugger Watch or Quick Watch
10272 * windows. They are designed to be short to type and take as few arguments as
10273 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10274 * See the function definition comment for more details.
10277 void cBlock(Compiler* comp, BasicBlock* block);
10278 void cBlocks(Compiler* comp);
10279 void cBlocksV(Compiler* comp);
10280 void cTree(Compiler* comp, GenTree* tree);
10281 void cTrees(Compiler* comp);
10282 void cEH(Compiler* comp);
10283 void cVar(Compiler* comp, unsigned lclNum);
10284 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10285 void cVars(Compiler* comp);
10286 void cVarsFinal(Compiler* comp);
10287 void cBlockPreds(Compiler* comp, BasicBlock* block);
10288 void cReach(Compiler* comp);
10289 void cDoms(Compiler* comp);
10290 void cLiveness(Compiler* comp);
10291 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10293 void cFuncIR(Compiler* comp);
10294 void cBlockIR(Compiler* comp, BasicBlock* block);
10295 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10296 void cTreeIR(Compiler* comp, GenTree* tree);
10297 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10298 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10299 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10300 int cOperandIR(Compiler* comp, GenTree* operand);
10301 int cLeafIR(Compiler* comp, GenTree* tree);
10302 int cIndirIR(Compiler* comp, GenTree* tree);
10303 int cListIR(Compiler* comp, GenTree* list);
10304 int cSsaNumIR(Compiler* comp, GenTree* tree);
10305 int cValNumIR(Compiler* comp, GenTree* tree);
10306 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10308 void dBlock(BasicBlock* block);
10311 void dTree(GenTree* tree);
10314 void dVar(unsigned lclNum);
10315 void dVarDsc(LclVarDsc* varDsc);
10318 void dBlockPreds(BasicBlock* block);
10322 void dCVarSet(VARSET_VALARG_TP vars);
10324 void dVarSet(VARSET_VALARG_TP vars);
10325 void dRegMask(regMaskTP mask);
10328 void dBlockIR(BasicBlock* block);
10329 void dTreeIR(GenTree* tree);
10330 void dLoopIR(Compiler::LoopDsc* loop);
10331 void dLoopNumIR(unsigned loopNum);
10332 int dTabStopIR(int curr, int tabstop);
10333 int dTreeTypeIR(GenTree* tree);
10334 int dTreeKindsIR(GenTree* tree);
10335 int dTreeFlagsIR(GenTree* tree);
10336 int dOperandIR(GenTree* operand);
10337 int dLeafIR(GenTree* tree);
10338 int dIndirIR(GenTree* tree);
10339 int dListIR(GenTree* list);
10340 int dSsaNumIR(GenTree* tree);
10341 int dValNumIR(GenTree* tree);
10342 int dDependsIR(GenTree* comma);
10345 GenTree* dFindTree(GenTree* tree, unsigned id);
10346 GenTree* dFindTree(unsigned id);
10347 GenTreeStmt* dFindStmt(unsigned id);
10348 BasicBlock* dFindBlock(unsigned bbNum);
10352 #include "compiler.hpp" // All the shared inline functions
10354 /*****************************************************************************/
10355 #endif //_COMPILER_H_
10356 /*****************************************************************************/