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 const size_t lvArgStackSize() const;
727 unsigned lvSlotNum; // original slot # (if remapped)
729 typeInfo lvVerTypeInfo; // type info needed for verification
731 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
733 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
735 BYTE* lvGcLayout; // GC layout info for structs
738 BlockSet lvRefBlks; // Set of blocks that contain refs
739 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
740 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
742 var_types TypeGet() const
744 return (var_types)lvType;
746 bool lvStackAligned() const
748 assert(lvIsStructField);
749 return ((lvFldOffset % sizeof(void*)) == 0);
751 bool lvNormalizeOnLoad() const
753 return varTypeIsSmall(TypeGet()) &&
754 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
755 (lvIsParam || lvAddrExposed || lvIsStructField);
758 bool lvNormalizeOnStore()
760 return varTypeIsSmall(TypeGet()) &&
761 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
762 !(lvIsParam || lvAddrExposed || lvIsStructField);
765 void lvaResetSortAgainFlag(Compiler* pComp);
766 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
767 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
768 void setPrefReg(regNumber regNum, Compiler* pComp);
769 void addPrefReg(regMaskTP regMask, Compiler* pComp);
770 bool IsFloatRegType() const
772 return isFloatRegType(lvType) || lvIsHfaRegArg();
774 var_types GetHfaType() const
776 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
778 void SetHfaType(var_types type)
780 assert(varTypeIsFloating(type));
781 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
784 #ifndef LEGACY_BACKEND
785 var_types lvaArgType();
788 PerSsaArray lvPerSsaData;
791 // Keep track of the # of SsaNames, for a bounds check.
792 unsigned lvNumSsaNames;
795 // Returns the address of the per-Ssa data for the given ssaNum (which is required
796 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
797 // not an SSA variable).
798 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
800 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
801 assert(SsaConfig::RESERVED_SSA_NUM == 0);
802 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
803 assert(zeroBased < lvNumSsaNames);
804 return &lvPerSsaData.GetRef(zeroBased);
809 void PrintVarReg() const
811 if (isRegPairType(TypeGet()))
813 printf("%s:%s", getRegName(lvOtherReg), // hi32
814 getRegName(lvRegNum)); // lo32
818 printf("%s", getRegName(lvRegNum));
823 }; // class LclVarDsc
826 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
827 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
831 XX The temporary lclVars allocated by the compiler for code generation XX
833 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
834 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
837 /*****************************************************************************
839 * The following keeps track of temporaries allocated in the stack frame
840 * during code-generation (after register allocation). These spill-temps are
841 * only used if we run out of registers while evaluating a tree.
843 * These are different from the more common temps allocated by lvaGrabTemp().
854 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
862 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
866 0); // temps must have a negative number (so they have a different number from all local variables)
867 tdOffs = BAD_TEMP_OFFSET;
871 IMPL_LIMITATION("too many spill temps");
876 bool tdLegalOffset() const
878 return tdOffs != BAD_TEMP_OFFSET;
882 int tdTempOffs() const
884 assert(tdLegalOffset());
887 void tdSetTempOffs(int offs)
890 assert(tdLegalOffset());
892 void tdAdjustTempOffs(int offs)
895 assert(tdLegalOffset());
898 int tdTempNum() const
903 unsigned tdTempSize() const
907 var_types tdTempType() const
913 // interface to hide linearscan implementation from rest of compiler
914 class LinearScanInterface
917 virtual void doLinearScan() = 0;
918 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
919 virtual bool willEnregisterLocalVars() const = 0;
922 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
924 // Information about arrays: their element type and size, and the offset of the first element.
925 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
926 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
927 // for example, in value numbering of array index expressions.
930 var_types m_elemType;
931 CORINFO_CLASS_HANDLE m_elemStructType;
933 unsigned m_elemOffset;
935 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
939 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
940 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
945 // This enumeration names the phases into which we divide compilation. The phases should completely
946 // partition a compilation.
949 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
950 #include "compphases.h"
954 extern const char* PhaseNames[];
955 extern const char* PhaseEnums[];
956 extern const LPCWSTR PhaseShortNames[];
958 // The following enum provides a simple 1:1 mapping to CLR API's
959 enum API_ICorJitInfo_Names
961 #define DEF_CLR_API(name) API_##name,
962 #include "ICorJitInfo_API_names.h"
966 //---------------------------------------------------------------
970 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
971 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
972 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
973 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
974 // by "m_timerFailure" being true.
975 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
978 #ifdef FEATURE_JIT_METHOD_PERF
979 // The string names of the phases.
980 static const char* PhaseNames[];
982 static bool PhaseHasChildren[];
983 static int PhaseParent[];
984 static bool PhaseReportsIRSize[];
986 unsigned m_byteCodeBytes;
987 unsigned __int64 m_totalCycles;
988 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
989 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
990 #if MEASURE_CLRAPI_CALLS
991 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
992 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
995 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
997 // For better documentation, we call EndPhase on
998 // non-leaf phases. We should also call EndPhase on the
999 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
1000 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
1001 // We add all such "redundant end phase" intervals to this variable below; we print
1002 // it out in a report, so we can verify that it is, indeed, very small. If it ever
1003 // isn't, this means that we're doing something significant between the end of the last
1004 // declared subphase and the end of its parent.
1005 unsigned __int64 m_parentPhaseEndSlop;
1006 bool m_timerFailure;
1008 #if MEASURE_CLRAPI_CALLS
1009 // The following measures the time spent inside each individual CLR API call.
1010 unsigned m_allClrAPIcalls;
1011 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1012 unsigned __int64 m_allClrAPIcycles;
1013 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1014 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1015 #endif // MEASURE_CLRAPI_CALLS
1017 CompTimeInfo(unsigned byteCodeBytes);
1021 #ifdef FEATURE_JIT_METHOD_PERF
1023 #if MEASURE_CLRAPI_CALLS
1024 struct WrapICorJitInfo;
1027 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1028 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1029 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1030 // threads, so it is protected by a lock.
1031 // This class is intended to be used as a singleton type, with only a single instance.
1032 class CompTimeSummaryInfo
1034 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1035 static CritSecObject s_compTimeSummaryLock;
1039 CompTimeInfo m_total;
1040 CompTimeInfo m_maximum;
1042 int m_numFilteredMethods;
1043 CompTimeInfo m_filtered;
1045 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1046 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1047 // If any OS interaction fails, returns 0.0.
1048 double CyclesPerSecond();
1050 // This can use what ever data you want to determine if the value to be added
1051 // belongs in the filtered section (it's always included in the unfiltered section)
1052 bool IncludedInFilteredData(CompTimeInfo& info);
1055 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1056 static CompTimeSummaryInfo s_compTimeSummary;
1058 CompTimeSummaryInfo()
1059 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1063 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1064 // This is thread safe.
1065 void AddInfo(CompTimeInfo& info, bool includePhases);
1067 // Print the summary information to "f".
1068 // This is not thread-safe; assumed to be called by only one thread.
1069 void Print(FILE* f);
1072 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1073 // and when the current phase started. This is intended to be part of a Compilation object. This is
1074 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1078 unsigned __int64 m_start; // Start of the compilation.
1079 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1080 #if MEASURE_CLRAPI_CALLS
1081 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1082 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1083 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1084 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1085 static double s_cyclesPerSec; // Cached for speedier measurements
1088 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1090 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1092 static CritSecObject s_csvLock; // Lock to protect the time log file.
1093 void PrintCsvMethodStats(Compiler* comp);
1096 void* operator new(size_t);
1097 void* operator new[](size_t);
1098 void operator delete(void*);
1099 void operator delete[](void*);
1102 // Initialized the timer instance
1103 JitTimer(unsigned byteCodeSize);
1105 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1107 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1110 static void PrintCsvHeader();
1112 // Ends the current phase (argument is for a redundant check).
1113 void EndPhase(Compiler* compiler, Phases phase);
1115 #if MEASURE_CLRAPI_CALLS
1116 // Start and end a timed CLR API call.
1117 void CLRApiCallEnter(unsigned apix);
1118 void CLRApiCallLeave(unsigned apix);
1119 #endif // MEASURE_CLRAPI_CALLS
1121 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1122 // and adds it to "sum".
1123 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1125 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1126 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1127 // "m_info" to true.
1128 bool GetThreadCycles(unsigned __int64* cycles)
1130 bool res = CycleTimer::GetThreadCyclesS(cycles);
1133 m_info.m_timerFailure = true;
1138 #endif // FEATURE_JIT_METHOD_PERF
1140 //------------------- Function/Funclet info -------------------------------
1141 enum FuncKind : BYTE
1143 FUNC_ROOT, // The main/root function (always id==0)
1144 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1145 FUNC_FILTER, // a funclet associated with an EH filter
1154 BYTE funFlags; // Currently unused, just here for padding
1155 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1156 // funclet. It is only valid if funKind field indicates this is a
1157 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1159 #if defined(_TARGET_AMD64_)
1161 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1162 emitLocation* startLoc;
1163 emitLocation* endLoc;
1164 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1165 emitLocation* coldEndLoc;
1166 UNWIND_INFO unwindHeader;
1167 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1168 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1169 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1170 unsigned unwindCodeSlot;
1172 #ifdef UNIX_AMD64_ABI
1173 jitstd::vector<CFI_CODE>* cfiCodes;
1174 #endif // UNIX_AMD64_ABI
1176 #elif defined(_TARGET_ARMARCH_)
1178 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1179 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1180 // Note: we only have a pointer here instead of the actual object,
1181 // to save memory in the JIT case (compared to the NGEN case),
1182 // where we don't have any cold section.
1183 // Note 2: we currently don't support hot/cold splitting in functions
1184 // with EH, so uwiCold will be NULL for all funclets.
1186 #endif // _TARGET_ARMARCH_
1188 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1189 // that isn't shared between the main function body and funclets.
1192 struct fgArgTabEntry
1195 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1198 otherRegNum = REG_NA;
1199 isStruct = false; // is this a struct arg
1201 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1203 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1205 // it will point at the actual argument in the gtCallLateArgs list.
1206 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1208 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1210 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1212 unsigned numRegs; // Count of number of registers that this argument uses
1214 // A slot is a pointer sized region in the OutArg area.
1215 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1216 unsigned numSlots; // Count of number of slots that this argument uses
1218 unsigned alignment; // 1 or 2 (slots/registers)
1219 unsigned lateArgInx; // index into gtCallLateArgs list
1220 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1222 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1223 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1224 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1225 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1226 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1227 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1228 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1229 // previous arguments.
1230 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1231 // to be on the stack despite its arg list position.
1233 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1234 bool isStruct : 1; // True if this is a struct arg
1236 regNumber otherRegNum; // The (second) register to use when passing this argument.
1238 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1239 #elif !defined(_TARGET_64BIT_)
1240 __declspec(property(get = getIsStruct)) bool isStruct;
1243 return varTypeIsStruct(node);
1245 #endif // !_TARGET_64BIT_
1248 void SetIsHfaRegArg(bool hfaRegArg)
1250 isHfaRegArg = hfaRegArg;
1253 void SetIsBackFilled(bool backFilled)
1255 isBackFilled = backFilled;
1258 bool IsBackFilled() const
1260 return isBackFilled;
1262 #else // !_TARGET_ARM_
1263 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1265 void SetIsHfaRegArg(bool hfaRegArg)
1269 void SetIsBackFilled(bool backFilled)
1273 bool IsBackFilled() const
1277 #endif // !_TARGET_ARM_
1283 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1285 //-------------------------------------------------------------------------
1287 // The class fgArgInfo is used to handle the arguments
1288 // when morphing a GT_CALL node.
1293 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1294 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1295 unsigned argCount; // Updatable arg count value
1296 unsigned nextSlotNum; // Updatable slot count value
1297 unsigned stkLevel; // Stack depth when we make this call (for x86)
1299 #if defined(UNIX_X86_ABI)
1300 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1301 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1302 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1303 // Computed dynamically during codegen, based on stkSizeBytes and the current
1304 // stack level (genStackLevel) when the first stack adjustment is made for
1308 #if FEATURE_FIXED_OUT_ARGS
1309 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1312 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1313 bool hasRegArgs; // true if we have one or more register arguments
1314 bool hasStackArgs; // true if we have one or more stack arguments
1315 bool argsComplete; // marker for state
1316 bool argsSorted; // marker for state
1317 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1320 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1323 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1324 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1326 fgArgTabEntryPtr AddRegArg(
1327 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1329 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1330 fgArgTabEntryPtr AddRegArg(
1337 const bool isStruct,
1338 const regNumber otherRegNum = REG_NA,
1339 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1340 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1342 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1346 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1348 void RemorphReset();
1349 fgArgTabEntryPtr RemorphRegArg(
1350 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1352 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1354 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1356 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1358 void ArgsComplete();
1362 void EvalArgsToTemps();
1364 void RecordStkLevel(unsigned stkLvl);
1365 unsigned RetrieveStkLevel();
1371 fgArgTabEntryPtr* ArgTable()
1375 unsigned GetNextSlotNum()
1385 return hasStackArgs;
1387 bool AreArgsComplete() const
1389 return argsComplete;
1391 #if FEATURE_FIXED_OUT_ARGS
1392 unsigned GetOutArgSize() const
1396 void SetOutArgSize(unsigned newVal)
1398 outArgSize = newVal;
1400 #endif // FEATURE_FIXED_OUT_ARGS
1402 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1404 #if defined(UNIX_X86_ABI)
1405 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1406 #endif // defined(UNIX_X86_ABI)
1409 void SetStkSizeBytes(unsigned newStkSizeBytes)
1411 #if defined(UNIX_X86_ABI)
1412 stkSizeBytes = newStkSizeBytes;
1413 #endif // defined(UNIX_X86_ABI)
1416 #if defined(UNIX_X86_ABI)
1417 unsigned GetStkAlign()
1421 unsigned GetStkSizeBytes() const
1423 return stkSizeBytes;
1425 bool IsStkAlignmentDone() const
1427 return alignmentDone;
1429 void SetStkAlignmentDone()
1431 alignmentDone = true;
1433 #endif // defined(UNIX_X86_ABI)
1435 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1436 GenTreePtr GetLateArg(unsigned argIndex);
1440 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1441 // We have the ability to mark source expressions with "Test Labels."
1442 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1443 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1445 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1448 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1449 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1450 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1451 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1452 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1455 struct TestLabelAndNum
1460 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1465 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1467 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1470 // This class implements the "IAllocator" interface, so that we can use
1471 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1473 class CompAllocator : public IAllocator
1476 #if MEASURE_MEM_ALLOC
1480 CompAllocator(Compiler* comp, CompMemKind cmk)
1482 #if MEASURE_MEM_ALLOC
1488 inline void* Alloc(size_t sz);
1490 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1492 // For the compiler's no-release allocator, free operations are no-ops.
1499 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1500 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1502 XX The big guy. The sections are currently organized as : XX
1504 XX o GenTree and BasicBlock XX
1516 XX o PrologScopeInfo XX
1517 XX o CodeGenerator XX
1522 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1523 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1528 friend class emitter;
1529 friend class UnwindInfo;
1530 friend class UnwindFragmentInfo;
1531 friend class UnwindEpilogInfo;
1532 friend class JitTimer;
1533 friend class LinearScan;
1534 friend class fgArgInfo;
1535 friend class Rationalizer;
1537 friend class Lowering;
1538 friend class CSE_DataFlow;
1539 friend class CSE_Heuristic;
1540 friend class CodeGenInterface;
1541 friend class CodeGen;
1542 friend class LclVarDsc;
1543 friend class TempDsc;
1545 friend class ObjectAllocator;
1547 #ifndef _TARGET_64BIT_
1548 friend class DecomposeLongs;
1549 #endif // !_TARGET_64BIT_
1552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1553 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1555 XX Misc structs definitions XX
1557 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1558 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1562 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1581 bool dumpIRDataflow;
1582 bool dumpIRBlockHeaders;
1584 LPCWSTR dumpIRPhase;
1585 LPCWSTR dumpIRFormat;
1587 bool shouldUseVerboseTrees();
1588 bool asciiTrees; // If true, dump trees using only ASCII characters
1589 bool shouldDumpASCIITrees();
1590 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1591 bool shouldUseVerboseSsa();
1592 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1593 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1595 const char* VarNameToStr(VarName name)
1600 DWORD expensiveDebugCheckLevel;
1603 #if FEATURE_MULTIREG_RET
1604 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1605 #endif // FEATURE_MULTIREG_RET
1608 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1609 #endif // ARM_SOFTFP
1611 //-------------------------------------------------------------------------
1612 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1613 // HFAs are one to four element structs where each element is the same
1614 // type, either all float or all double. They are treated specially
1615 // in the ARM Procedure Call Standard, specifically, they are passed in
1616 // floating-point registers instead of the general purpose registers.
1619 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1620 bool IsHfa(GenTreePtr tree);
1622 var_types GetHfaType(GenTreePtr tree);
1623 unsigned GetHfaCount(GenTreePtr tree);
1625 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1626 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1628 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1629 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1631 //-------------------------------------------------------------------------
1632 // The following is used for validating format of EH table
1636 typedef struct EHNodeDsc* pEHNodeDsc;
1638 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1639 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1652 EHBlockType ehnBlockType; // kind of EH block
1653 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1654 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1655 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1657 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1658 pEHNodeDsc ehnChild; // leftmost nested block
1660 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1661 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1663 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1664 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1666 inline void ehnSetTryNodeType()
1668 ehnBlockType = TryNode;
1670 inline void ehnSetFilterNodeType()
1672 ehnBlockType = FilterNode;
1674 inline void ehnSetHandlerNodeType()
1676 ehnBlockType = HandlerNode;
1678 inline void ehnSetFinallyNodeType()
1680 ehnBlockType = FinallyNode;
1682 inline void ehnSetFaultNodeType()
1684 ehnBlockType = FaultNode;
1687 inline BOOL ehnIsTryBlock()
1689 return ehnBlockType == TryNode;
1691 inline BOOL ehnIsFilterBlock()
1693 return ehnBlockType == FilterNode;
1695 inline BOOL ehnIsHandlerBlock()
1697 return ehnBlockType == HandlerNode;
1699 inline BOOL ehnIsFinallyBlock()
1701 return ehnBlockType == FinallyNode;
1703 inline BOOL ehnIsFaultBlock()
1705 return ehnBlockType == FaultNode;
1708 // returns true if there is any overlap between the two nodes
1709 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1711 if (node1->ehnStartOffset < node2->ehnStartOffset)
1713 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1717 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1721 // fails with BADCODE if inner is not completely nested inside outer
1722 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1724 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1728 //-------------------------------------------------------------------------
1729 // Exception handling functions
1732 #if !FEATURE_EH_FUNCLETS
1734 bool ehNeedsShadowSPslots()
1736 return (info.compXcptnsCount || opts.compDbgEnC);
1739 // 0 for methods with no EH
1740 // 1 for methods with non-nested EH, or where only the try blocks are nested
1741 // 2 for a method with a catch within a catch
1743 unsigned ehMaxHndNestingCount;
1745 #endif // !FEATURE_EH_FUNCLETS
1747 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1748 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1750 bool bbInCatchHandlerILRange(BasicBlock* blk);
1751 bool bbInFilterILRange(BasicBlock* blk);
1752 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1753 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1754 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1755 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1756 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1758 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1759 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1761 // Returns true if "block" is the start of a try region.
1762 bool bbIsTryBeg(BasicBlock* block);
1764 // Returns true if "block" is the start of a handler or filter region.
1765 bool bbIsHandlerBeg(BasicBlock* block);
1767 // Returns true iff "block" is where control flows if an exception is raised in the
1768 // try region, and sets "*regionIndex" to the index of the try for the handler.
1769 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1770 // block of the filter, but not for the filter's handler.
1771 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1773 bool ehHasCallableHandlers();
1775 // Return the EH descriptor for the given region index.
1776 EHblkDsc* ehGetDsc(unsigned regionIndex);
1778 // Return the EH index given a region descriptor.
1779 unsigned ehGetIndex(EHblkDsc* ehDsc);
1781 // Return the EH descriptor index of the enclosing try, for the given region index.
1782 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1784 // Return the EH descriptor index of the enclosing handler, for the given region index.
1785 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1787 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1788 // block is not in a 'try' region).
1789 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1791 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1792 // if this block is not in a filter or handler region).
1793 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1795 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1796 // nullptr if this block's exceptions propagate to caller).
1797 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1799 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1800 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1801 bool ehIsBlockEHLast(BasicBlock* block);
1803 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1805 // Return the region index of the most nested EH region this block is in.
1806 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1808 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1809 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1811 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1812 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1813 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1814 // (It can never be a filter.)
1815 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1817 // A block has been deleted. Update the EH table appropriately.
1818 void ehUpdateForDeletedBlock(BasicBlock* block);
1820 // Determine whether a block can be deleted while preserving the EH normalization rules.
1821 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1823 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1824 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1826 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1827 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1828 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1829 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1830 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1831 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1832 // lives in a filter.)
1833 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1835 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1836 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1837 // (nullptr if the last block is the last block in the program).
1838 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1839 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1842 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1843 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1844 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1847 #if FEATURE_EH_FUNCLETS
1848 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1849 // if there is a filter that protects a region with a nested EH clause (such as a
1850 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1851 // genFuncletProlog() for more details. However, the VM seems to use it for more
1852 // purposes, maybe including debugging. Until we are sure otherwise, always create
1853 // a PSPSym for functions with any EH.
1854 bool ehNeedsPSPSym() const
1858 #else // _TARGET_X86_
1859 return compHndBBtabCount > 0;
1860 #endif // _TARGET_X86_
1863 bool ehAnyFunclets(); // Are there any funclets in this function?
1864 unsigned ehFuncletCount(); // Return the count of funclets in the function
1866 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1867 #else // !FEATURE_EH_FUNCLETS
1868 bool ehAnyFunclets()
1872 unsigned ehFuncletCount()
1877 unsigned bbThrowIndex(BasicBlock* blk)
1879 return blk->bbTryIndex;
1880 } // Get the index to use as the cache key for sharing throw blocks
1881 #endif // !FEATURE_EH_FUNCLETS
1883 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1884 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1885 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1886 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1887 // convenient to also consider it a predecessor.)
1888 flowList* BlockPredsWithEH(BasicBlock* blk);
1890 // This table is useful for memoization of the method above.
1891 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1893 BlockToFlowListMap* m_blockToEHPreds;
1894 BlockToFlowListMap* GetBlockToEHPreds()
1896 if (m_blockToEHPreds == nullptr)
1898 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1900 return m_blockToEHPreds;
1903 void* ehEmitCookie(BasicBlock* block);
1904 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1906 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1908 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1910 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1912 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1914 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1916 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1918 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1920 void fgAllocEHTable();
1922 void fgRemoveEHTableEntry(unsigned XTnum);
1924 #if FEATURE_EH_FUNCLETS
1926 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1928 #endif // FEATURE_EH_FUNCLETS
1932 #endif // !FEATURE_EH
1934 void fgSortEHTable();
1936 // Causes the EH table to obey some well-formedness conditions, by inserting
1937 // empty BB's when necessary:
1938 // * No block is both the first block of a handler and the first block of a try.
1939 // * No block is the first block of multiple 'try' regions.
1940 // * No block is the last block of multiple EH regions.
1941 void fgNormalizeEH();
1942 bool fgNormalizeEHCase1();
1943 bool fgNormalizeEHCase2();
1944 bool fgNormalizeEHCase3();
1947 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1948 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1949 void fgVerifyHandlerTab();
1950 void fgDispHandlerTab();
1953 bool fgNeedToSortEHTable;
1955 void verInitEHTree(unsigned numEHClauses);
1956 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1957 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1958 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1959 void verCheckNestingLevel(EHNodeDsc* initRoot);
1962 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1963 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1965 XX GenTree and BasicBlock XX
1967 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1969 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1970 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1973 // Functions to create nodes
1974 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1977 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1979 // For binary opers.
1980 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1982 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1984 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1985 var_types type = TYP_I_IMPL,
1986 GenTreePtr op1 = nullptr,
1987 GenTreePtr op2 = nullptr);
1989 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1991 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1993 GenTreePtr gtNewJmpTableNode();
1994 GenTreePtr gtNewIconHandleNode(
1995 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1997 unsigned gtTokenToIconFlags(unsigned token);
1999 GenTreePtr gtNewIconEmbHndNode(void* value,
2002 unsigned handle1 = 0,
2003 void* handle2 = nullptr,
2004 void* compileTimeHandle = nullptr);
2006 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2007 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2008 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2009 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
2011 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2013 GenTreePtr gtNewLconNode(__int64 value);
2015 GenTreePtr gtNewDconNode(double value);
2017 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2019 GenTreePtr gtNewZeroConNode(var_types type);
2021 GenTreePtr gtNewOneConNode(var_types type);
2024 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2025 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2028 GenTreeBlk* gtNewBlkOpNode(
2029 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
2031 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2033 GenTree* gtNewPutArgReg(var_types type, GenTreePtr arg, regNumber argReg);
2036 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
2039 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2040 void gtSetObjGcInfo(GenTreeObj* objNode);
2041 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
2042 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
2044 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2046 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
2048 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2049 CORINFO_METHOD_HANDLE handle,
2051 GenTreeArgList* args,
2052 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2054 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
2056 GenTreeArgList* args,
2057 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2059 GenTreeCall* gtNewHelperCallNode(unsigned helper, var_types type, GenTreeArgList* args = nullptr);
2061 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2064 GenTreeSIMD* gtNewSIMDNode(
2065 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2066 GenTreeSIMD* gtNewSIMDNode(var_types type,
2069 SIMDIntrinsicID simdIntrinsicID,
2072 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2075 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2076 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2077 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2079 GenTreePtr gtNewCodeRef(BasicBlock* block);
2081 GenTreePtr gtNewFieldRef(
2082 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2084 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2086 GenTreeArrLen* gtNewArrLen(var_types typ, GenTree* arrayOp, int lenOffset);
2088 GenTree* gtNewIndir(var_types typ, GenTree* addr);
2090 GenTreeArgList* gtNewArgList(GenTreePtr op);
2091 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2092 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2094 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2095 static fgArgTabEntryPtr gtArgEntryByNode(GenTreeCall* call, GenTreePtr node);
2096 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2097 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2099 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2101 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2103 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2104 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2105 CORINFO_ACCESS_FLAGS access,
2106 CORINFO_FIELD_INFO* pFieldInfo,
2108 CORINFO_CLASS_HANDLE structType,
2111 GenTreePtr gtNewNothingNode();
2113 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2115 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2117 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2119 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2121 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2123 //------------------------------------------------------------------------
2124 // Other GenTree functions
2126 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2128 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2129 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2130 // IntCnses with value `deepVarVal`.
2131 GenTreePtr gtCloneExpr(
2132 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2134 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2135 // `varNum` to int constants with value `varVal`.
2136 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2138 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2141 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2143 void gtUpdateSideEffects(GenTree* stmt, GenTree* tree);
2145 void gtUpdateTreeAncestorsSideEffects(GenTree* tree);
2147 void gtUpdateStmtSideEffects(GenTree* stmt);
2149 void gtResetNodeSideEffects(GenTree* tree);
2151 // Returns "true" iff the complexity (not formally defined, but first interpretation
2152 // is #of nodes in subtree) of "tree" is greater than "limit".
2153 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2154 // before they have been set.)
2155 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2157 bool gtCompareTree(GenTree* op1, GenTree* op2);
2159 GenTreePtr gtReverseCond(GenTree* tree);
2161 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2163 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2165 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2167 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2170 unsigned gtHashValue(GenTree* tree);
2172 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2175 void gtPrepareCost(GenTree* tree);
2176 bool gtIsLikelyRegVar(GenTree* tree);
2178 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2180 // Returns true iff the secondNode can be swapped with firstNode.
2181 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2183 unsigned gtSetEvalOrder(GenTree* tree);
2185 #if FEATURE_STACK_FP_X87
2187 void gtComputeFPlvls(GenTreePtr tree);
2188 #endif // FEATURE_STACK_FP_X87
2190 void gtSetStmtInfo(GenTree* stmt);
2192 // Returns "true" iff "node" has any of the side effects in "flags".
2193 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2195 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2196 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2198 // Appends 'expr' in front of 'list'
2199 // 'list' will typically start off as 'nullptr'
2200 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2201 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2203 void gtExtractSideEffList(GenTreePtr expr,
2205 unsigned flags = GTF_SIDE_EFFECT,
2206 bool ignoreRoot = false);
2208 GenTreePtr gtGetThisArg(GenTreeCall* call);
2210 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2211 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2212 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2213 // the given "fldHnd", is such an object pointer.
2214 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2216 // Return true if call is a recursive call; return false otherwise.
2217 // Note when inlining, this looks for calls back to the root method.
2218 bool gtIsRecursiveCall(GenTreeCall* call)
2220 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2223 //-------------------------------------------------------------------------
2225 GenTreePtr gtFoldExpr(GenTreePtr tree);
2228 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2229 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2230 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2231 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2232 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2233 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2234 // optimizations for now.
2235 __attribute__((optnone))
2237 gtFoldExprConst(GenTreePtr tree);
2238 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2239 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2240 bool gtTryRemoveBoxUpstreamEffects(GenTreePtr tree);
2242 //-------------------------------------------------------------------------
2243 // Get the handle, if any.
2244 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2245 // Get the handle, and assert if not found.
2246 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2247 // Get the handle for a ref type.
2248 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTreePtr tree, bool* isExact, bool* isNonNull);
2250 //-------------------------------------------------------------------------
2251 // Functions to display the trees
2254 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2256 void gtDispVN(GenTreePtr tree);
2257 void gtDispConst(GenTreePtr tree);
2258 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2259 void gtDispNodeName(GenTreePtr tree);
2260 void gtDispRegVal(GenTreePtr tree);
2272 void gtDispChild(GenTreePtr child,
2273 IndentStack* indentStack,
2275 __in_opt const char* msg = nullptr,
2276 bool topOnly = false);
2277 void gtDispTree(GenTreePtr tree,
2278 IndentStack* indentStack = nullptr,
2279 __in_opt const char* msg = nullptr,
2280 bool topOnly = false,
2281 bool isLIR = false);
2282 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2283 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2284 char* gtGetLclVarName(unsigned lclNum);
2285 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2286 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2287 void gtGetArgMsg(GenTreeCall* call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2288 void gtGetLateArgMsg(GenTreeCall* call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2289 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2290 void gtDispFieldSeq(FieldSeqNode* pfsn);
2292 void gtDispRange(LIR::ReadOnlyRange const& range);
2294 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2296 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2308 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2309 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2312 static fgWalkPreFn gtAssertColonCond;
2314 static fgWalkPreFn gtMarkColonCond;
2315 static fgWalkPreFn gtClearColonCond;
2317 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2318 bool gtHasCatchArg(GenTreePtr tree);
2319 bool gtHasUnmanagedCall(GenTreePtr tree);
2321 typedef ArrayStack<GenTree*> GenTreeStack;
2323 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2324 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2326 //=========================================================================
2327 // BasicBlock functions
2329 // This is a debug flag we will use to assert when creating block during codegen
2330 // as this interferes with procedure splitting. If you know what you're doing, set
2331 // it to true before creating the block. (DEBUG only)
2332 bool fgSafeBasicBlockCreation;
2335 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2338 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2339 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2343 XX The variables to be used by the code generator. XX
2345 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2346 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2350 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2351 // be placed in the stack frame and it's fields must be laid out sequentially.
2353 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2354 // a local variable that can be enregistered or placed in the stack frame.
2355 // The fields do not need to be laid out sequentially
2357 enum lvaPromotionType
2359 PROMOTION_TYPE_NONE, // The struct local is not promoted
2360 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2361 // and its field locals are independent of its parent struct local.
2362 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2363 // but its field locals depend on its parent struct local.
2366 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2367 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2369 /*****************************************************************************/
2371 enum FrameLayoutState
2374 INITIAL_FRAME_LAYOUT,
2375 PRE_REGALLOC_FRAME_LAYOUT,
2376 REGALLOC_FRAME_LAYOUT,
2377 TENTATIVE_FRAME_LAYOUT,
2382 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2383 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2384 bool lvaSortAgain; // true: We need to sort the lvaTable
2385 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2386 unsigned lvaCount; // total number of locals
2388 unsigned lvaRefCount; // total number of references to locals
2389 LclVarDsc* lvaTable; // variable descriptor table
2390 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2392 LclVarDsc** lvaRefSorted; // table sorted by refcount
2394 unsigned short lvaTrackedCount; // actual # of locals being tracked
2395 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2397 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2398 // Only for AMD64 System V cache the first caller stack homed argument.
2399 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2400 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2403 VARSET_TP lvaTrackedVars; // set of tracked variables
2405 #ifndef _TARGET_64BIT_
2406 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2408 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2410 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2411 // It that changes, this changes. VarSets from different epochs
2412 // cannot be meaningfully combined.
2414 unsigned GetCurLVEpoch()
2419 // reverse map of tracked number to var number
2420 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2422 #ifdef LEGACY_BACKEND
2423 // variable interference graph
2424 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2427 // variable preference graph
2428 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2432 // # of procs compiled a with double-aligned stack
2433 static unsigned s_lvaDoubleAlignedProcsCount;
2437 // Getters and setters for address-exposed and do-not-enregister local var properties.
2438 bool lvaVarAddrExposed(unsigned varNum);
2439 void lvaSetVarAddrExposed(unsigned varNum);
2440 bool lvaVarDoNotEnregister(unsigned varNum);
2442 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2443 enum DoNotEnregisterReason
2448 DNER_VMNeedsStackAddr,
2449 DNER_LiveInOutOfHandler,
2450 DNER_LiveAcrossUnmanagedCall,
2451 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2452 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2453 DNER_DepField, // It is a field of a dependently promoted struct
2454 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2455 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2456 #if !defined(LEGACY_BACKEND) && !defined(_TARGET_64BIT_)
2457 DNER_LongParamField, // It is a decomposed field of a long parameter.
2459 #ifdef JIT32_GCENCODER
2464 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2466 unsigned lvaVarargsHandleArg;
2468 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2470 #endif // _TARGET_X86_
2472 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2473 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2474 #if FEATURE_FIXED_OUT_ARGS
2475 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2477 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2478 // that tracks whether the lock has been taken
2480 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2481 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2482 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2484 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2485 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2487 #if FEATURE_FIXED_OUT_ARGS
2488 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2489 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2490 #endif // FEATURE_FIXED_OUT_ARGS
2493 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2494 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2495 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2496 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2497 // this variable to be this scratch word whenever struct promotion occurs.
2498 unsigned lvaPromotedStructAssemblyScratchVar;
2499 #endif // _TARGET_ARM_
2502 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2503 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2506 unsigned lvaGenericsContextUseCount;
2508 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2509 // CORINFO_GENERICS_CTXT_FROM_THIS?
2510 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2512 //-------------------------------------------------------------------------
2513 // All these frame offsets are inter-related and must be kept in sync
2515 #if !FEATURE_EH_FUNCLETS
2516 // This is used for the callable handlers
2517 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2518 #endif // FEATURE_EH_FUNCLETS
2520 unsigned lvaCachedGenericContextArgOffs;
2521 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2524 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2526 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2528 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2529 // after the reg predict we will use a computed maxTmpSize
2530 // which is based upon the number of spill temps predicted by reg predict
2531 // All this is necessary because if we under-estimate the size of the spill
2532 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2534 // Pre codegen max spill temp size.
2535 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2537 //-------------------------------------------------------------------------
2539 unsigned lvaGetMaxSpillTempSize();
2541 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2542 #endif // _TARGET_ARM_
2543 void lvaAssignFrameOffsets(FrameLayoutState curState);
2544 void lvaFixVirtualFrameOffsets();
2546 #ifndef LEGACY_BACKEND
2547 void lvaUpdateArgsWithInitialReg();
2548 #endif // !LEGACY_BACKEND
2550 void lvaAssignVirtualFrameOffsetsToArgs();
2551 #ifdef UNIX_AMD64_ABI
2552 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2553 #else // !UNIX_AMD64_ABI
2554 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2555 #endif // !UNIX_AMD64_ABI
2556 void lvaAssignVirtualFrameOffsetsToLocals();
2557 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2558 #ifdef _TARGET_AMD64_
2559 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2560 bool lvaIsCalleeSavedIntRegCountEven();
2562 void lvaAlignFrame();
2563 void lvaAssignFrameOffsetsToPromotedStructs();
2564 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2567 void lvaDumpRegLocation(unsigned lclNum);
2568 void lvaDumpFrameLocation(unsigned lclNum);
2569 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2570 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2571 // layout state defined by lvaDoneFrameLayout
2574 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2575 // to avoid bugs from borderline cases.
2576 #define MAX_FrameSize 0x3FFFFFFF
2577 void lvaIncrementFrameSize(unsigned size);
2579 unsigned lvaFrameSize(FrameLayoutState curState);
2581 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2582 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2584 // Returns the caller-SP-relative offset for the local variable "varNum."
2585 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2587 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2588 int lvaGetSPRelativeOffset(unsigned varNum);
2590 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2591 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2593 //------------------------ For splitting types ----------------------------
2595 void lvaInitTypeRef();
2597 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2598 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2599 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2600 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2601 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2602 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2604 void lvaInitVarDsc(LclVarDsc* varDsc,
2606 CorInfoType corInfoType,
2607 CORINFO_CLASS_HANDLE typeHnd,
2608 CORINFO_ARG_LIST_HANDLE varList,
2609 CORINFO_SIG_INFO* varSig);
2611 static unsigned lvaTypeRefMask(var_types type);
2613 var_types lvaGetActualType(unsigned lclNum);
2614 var_types lvaGetRealType(unsigned lclNum);
2616 //-------------------------------------------------------------------------
2620 unsigned lvaLclSize(unsigned varNum);
2621 unsigned lvaLclExactSize(unsigned varNum);
2623 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2625 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2626 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2627 // the return result.
2628 bool lvaLclVarRefsAccum(
2629 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2631 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2632 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2633 // and (destructively) unions "trkedVars" into "*result".
2634 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2636 ALLVARSET_VALARG_TP allVars,
2637 VARSET_VALARG_TP trkdVars);
2639 bool lvaHaveManyLocals() const;
2641 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2642 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2643 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2646 void lvaSortByRefCount();
2647 void lvaDumpRefCounts();
2649 void lvaMarkLocalVars(BasicBlock* block);
2651 void lvaMarkLocalVars(); // Local variable ref-counting
2653 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
2655 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2657 void lvaIncRefCnts(GenTreePtr tree);
2658 void lvaDecRefCnts(GenTreePtr tree);
2660 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2661 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2662 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2665 struct lvaStressLclFldArgs
2667 Compiler* m_pCompiler;
2671 static fgWalkPreFn lvaStressLclFldCB;
2672 void lvaStressLclFld();
2674 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2675 void lvaDispVarSet(VARSET_VALARG_TP set);
2680 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2682 int lvaFrameAddress(int varNum, bool* pFPbased);
2685 bool lvaIsParameter(unsigned varNum);
2686 bool lvaIsRegArgument(unsigned varNum);
2687 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2688 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2689 // that writes to arg0
2691 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2692 // (this is an overload of lvIsTemp because there are no temp parameters).
2693 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2694 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2695 bool lvaIsImplicitByRefLocal(unsigned varNum)
2697 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2698 LclVarDsc* varDsc = &(lvaTable[varNum]);
2699 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2701 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
2704 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2708 // Returns true if this local var is a multireg struct
2709 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2711 // If the local is a TYP_STRUCT, get/set a class handle describing it
2712 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2713 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2715 // If the local is TYP_REF, set or update the associated class information.
2716 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2717 void lvaSetClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2718 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
2719 void lvaUpdateClass(unsigned varNum, GenTreePtr tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
2721 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2723 // Info about struct fields
2724 struct lvaStructFieldInfo
2726 CORINFO_FIELD_HANDLE fldHnd;
2727 unsigned char fldOffset;
2728 unsigned char fldOrdinal;
2731 CORINFO_CLASS_HANDLE fldTypeHnd;
2734 // Info about struct to be promoted.
2735 struct lvaStructPromotionInfo
2737 CORINFO_CLASS_HANDLE typeHnd;
2739 bool requiresScratchVar;
2742 unsigned char fieldCnt;
2743 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2745 lvaStructPromotionInfo()
2746 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2751 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2752 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2753 lvaStructPromotionInfo* StructPromotionInfo,
2755 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2756 bool lvaShouldPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* structPromotionInfo);
2757 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2758 #if !defined(_TARGET_64BIT_)
2759 void lvaPromoteLongVars();
2760 #endif // !defined(_TARGET_64BIT_)
2761 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2762 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2763 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2764 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2765 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2766 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2767 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2769 #if defined(FEATURE_SIMD)
2770 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2772 assert(varDsc->lvType == TYP_SIMD12);
2773 assert(varDsc->lvExactSize == 12);
2775 #if defined(_TARGET_64BIT_)
2776 assert(varDsc->lvSize() == 16);
2777 #endif // defined(_TARGET_64BIT_)
2779 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2780 // already does this calculation. However, we also need to prevent mapping types if the var is a
2781 // dependently promoted struct field, which must remain its exact size within its parent struct.
2782 // However, we don't know this until late, so we may have already pretended the field is bigger
2784 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2793 #endif // defined(FEATURE_SIMD)
2795 BYTE* lvaGetGcLayout(unsigned varNum);
2796 bool lvaTypeIsGC(unsigned varNum);
2797 unsigned lvaGSSecurityCookie; // LclVar number
2798 bool lvaTempsHaveLargerOffsetThanVars();
2800 unsigned lvaSecurityObject; // variable representing the security object on the stack
2801 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2803 #if FEATURE_EH_FUNCLETS
2804 unsigned lvaPSPSym; // variable representing the PSPSym
2807 InlineInfo* impInlineInfo;
2808 InlineStrategy* m_inlineStrategy;
2810 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2811 Compiler* impInlineRoot();
2813 #if defined(DEBUG) || defined(INLINE_DATA)
2814 unsigned __int64 getInlineCycleCount()
2816 return m_compCycles;
2818 #endif // defined(DEBUG) || defined(INLINE_DATA)
2820 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2821 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2823 //=========================================================================
2825 //=========================================================================
2828 //---------------- Local variable ref-counting ----------------------------
2831 BasicBlock* lvaMarkRefsCurBlock;
2832 GenTreePtr lvaMarkRefsCurStmt;
2834 BasicBlock::weight_t lvaMarkRefsWeight;
2836 void lvaMarkLclRefs(GenTreePtr tree);
2838 bool IsDominatedByExceptionalEntry(BasicBlock* block);
2839 void SetVolatileHint(LclVarDsc* varDsc);
2841 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2842 PerSsaArray lvMemoryPerSsaData;
2843 unsigned lvMemoryNumSsaNames;
2846 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2847 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2848 // not an SSA variable).
2849 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2851 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2852 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2854 assert(ssaNum < lvMemoryNumSsaNames);
2855 return &lvMemoryPerSsaData.GetRef(ssaNum);
2859 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2860 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2864 XX Imports the given method and converts it to semantic trees XX
2866 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2867 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2873 void impImport(BasicBlock* method);
2875 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2876 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2877 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2878 CORINFO_CLASS_HANDLE impGetStringClass();
2879 CORINFO_CLASS_HANDLE impGetObjectClass();
2881 //=========================================================================
2883 //=========================================================================
2886 //-------------------- Stack manipulation ---------------------------------
2888 unsigned impStkSize; // Size of the full stack
2890 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2892 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2894 struct SavedStack // used to save/restore stack contents.
2896 unsigned ssDepth; // number of values on stack
2897 StackEntry* ssTrees; // saved tree values
2900 bool impIsPrimitive(CorInfoType type);
2901 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2903 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2905 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2906 void impPushNullObjRefOnStack();
2907 StackEntry impPopStack();
2908 StackEntry& impStackTop(unsigned n = 0);
2909 unsigned impStackHeight();
2911 void impSaveStackState(SavedStack* savePtr, bool copy);
2912 void impRestoreStackState(SavedStack* savePtr);
2914 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2915 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2916 CORINFO_CALL_INFO* pCallInfo);
2918 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2920 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2922 bool impCanPInvokeInline();
2923 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2924 void impCheckForPInvokeCall(
2925 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2926 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2927 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2929 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2930 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2931 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2933 var_types impImportCall(OPCODE opcode,
2934 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2935 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2937 GenTreePtr newobjThis,
2939 CORINFO_CALL_INFO* callInfo,
2940 IL_OFFSET rawILOffset);
2942 void impDevirtualizeCall(GenTreeCall* call,
2944 CORINFO_METHOD_HANDLE* method,
2945 unsigned* methodFlags,
2946 CORINFO_CONTEXT_HANDLE* contextHandle,
2947 CORINFO_CONTEXT_HANDLE* exactContextHandle);
2949 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2951 GenTreePtr impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
2953 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2956 var_types impImportJitTestLabelMark(int numArgs);
2959 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2961 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2963 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2964 CORINFO_ACCESS_FLAGS access,
2965 CORINFO_FIELD_INFO* pFieldInfo,
2968 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2970 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2972 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2974 void impImportLeave(BasicBlock* block);
2975 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2976 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2977 CORINFO_CLASS_HANDLE clsHnd,
2978 CORINFO_METHOD_HANDLE method,
2979 CORINFO_SIG_INFO* sig,
2983 bool isJitIntrinsic,
2984 CorInfoIntrinsics* pIntrinsicID);
2985 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2986 CORINFO_SIG_INFO* sig,
2989 CorInfoIntrinsics intrinsicID);
2990 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2992 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2994 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2995 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2996 CORINFO_THIS_TRANSFORM transform);
2998 //----------------- Manipulating the trees and stmts ----------------------
3000 GenTreePtr impTreeList; // Trees for the BB being imported
3001 GenTreePtr impTreeLast; // The last tree for the current BB
3005 CHECK_SPILL_ALL = -1,
3006 CHECK_SPILL_NONE = -2
3010 void impBeginTreeList();
3011 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
3012 void impEndTreeList(BasicBlock* block);
3013 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
3014 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
3015 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
3016 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
3017 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
3018 void impAssignTempGen(unsigned tmp,
3021 GenTreePtr* pAfterStmt = nullptr,
3022 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3023 BasicBlock* block = nullptr);
3024 void impAssignTempGen(unsigned tmpNum,
3026 CORINFO_CLASS_HANDLE structHnd,
3028 GenTreePtr* pAfterStmt = nullptr,
3029 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3030 BasicBlock* block = nullptr);
3031 GenTreePtr impCloneExpr(GenTreePtr tree,
3033 CORINFO_CLASS_HANDLE structHnd,
3035 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
3036 GenTreePtr impAssignStruct(GenTreePtr dest,
3038 CORINFO_CLASS_HANDLE structHnd,
3040 GenTreePtr* pAfterStmt = nullptr,
3041 BasicBlock* block = nullptr);
3042 GenTreePtr impAssignStructPtr(GenTreePtr dest,
3044 CORINFO_CLASS_HANDLE structHnd,
3046 GenTreePtr* pAfterStmt = nullptr,
3047 BasicBlock* block = nullptr);
3049 GenTreePtr impGetStructAddr(GenTreePtr structVal,
3050 CORINFO_CLASS_HANDLE structHnd,
3054 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3055 BYTE* gcLayout = nullptr,
3056 unsigned* numGCVars = nullptr,
3057 var_types* simdBaseType = nullptr);
3059 GenTreePtr impNormStructVal(GenTreePtr structVal,
3060 CORINFO_CLASS_HANDLE structHnd,
3062 bool forceNormalization = false);
3064 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3065 BOOL* pRuntimeLookup = nullptr,
3066 BOOL mustRestoreHandle = FALSE,
3067 BOOL importParent = FALSE);
3069 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3070 BOOL* pRuntimeLookup = nullptr,
3071 BOOL mustRestoreHandle = FALSE)
3073 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3076 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3077 CORINFO_LOOKUP* pLookup,
3079 void* compileTimeHandle);
3081 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3083 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3084 CORINFO_LOOKUP* pLookup,
3085 void* compileTimeHandle);
3087 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3089 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3090 CorInfoHelpFunc helper,
3092 GenTreeArgList* arg = nullptr,
3093 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3095 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3097 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3100 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3101 CORINFO_CLASS_HANDLE typeClass,
3105 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3106 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3107 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3108 static bool IsMathIntrinsic(GenTreePtr tree);
3111 //----------------- Importing the method ----------------------------------
3113 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3116 unsigned impCurOpcOffs;
3117 const char* impCurOpcName;
3118 bool impNestedStackSpill;
3120 // For displaying instrs with generated native code (-n:B)
3121 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3122 void impNoteLastILoffs();
3125 /* IL offset of the stmt currently being imported. It gets set to
3126 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3127 updated at IL offsets for which we have to report mapping info.
3128 It also includes flag bits, so use jitGetILoffs()
3129 to get the actual IL offset value.
3132 IL_OFFSETX impCurStmtOffs;
3133 void impCurStmtOffsSet(IL_OFFSET offs);
3135 void impNoteBranchOffs();
3137 unsigned impInitBlockLineInfo();
3139 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3140 bool impIsThis(GenTreePtr obj);
3141 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3142 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3143 bool impIsAnySTLOC(OPCODE opcode)
3145 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3146 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3149 GenTreeArgList* impPopList(unsigned count, CORINFO_SIG_INFO* sig, GenTreeArgList* prefixTree = nullptr);
3151 GenTreeArgList* impPopRevList(unsigned count, CORINFO_SIG_INFO* sig, unsigned skipReverseCount = 0);
3154 * Get current IL offset with stack-empty info incoporated
3156 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3158 //---------------- Spilling the importer stack ----------------------------
3160 // The maximum number of bytes of IL processed without clean stack state.
3161 // It allows to limit the maximum tree size and depth.
3162 static const unsigned MAX_TREE_SIZE = 200;
3163 bool impCanSpillNow(OPCODE prevOpcode);
3169 SavedStack pdSavedStack;
3170 ThisInitState pdThisPtrInit;
3173 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3174 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3176 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3177 ExpandArray<BYTE> impPendingBlockMembers;
3179 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3180 // Operates on the map in the top-level ancestor.
3181 BYTE impGetPendingBlockMember(BasicBlock* blk)
3183 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3186 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3187 // Operates on the map in the top-level ancestor.
3188 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3190 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3193 bool impCanReimport;
3195 bool impSpillStackEntry(unsigned level,
3199 bool bAssertOnRecursion,
3204 void impSpillStackEnsure(bool spillLeaves = false);
3205 void impEvalSideEffects();
3206 void impSpillSpecialSideEff();
3207 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3208 void impSpillValueClasses();
3209 void impSpillEvalStack();
3210 static fgWalkPreFn impFindValueClasses;
3211 void impSpillLclRefs(ssize_t lclNum);
3213 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3215 void impImportBlockCode(BasicBlock* block);
3217 void impReimportMarkBlock(BasicBlock* block);
3218 void impReimportMarkSuccessors(BasicBlock* block);
3220 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3222 void impImportBlockPending(BasicBlock* block);
3224 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3225 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3226 // for the block, but instead, just re-uses the block's existing EntryState.
3227 void impReimportBlockPending(BasicBlock* block);
3229 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3231 void impImportBlock(BasicBlock* block);
3233 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3234 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3235 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3236 // the variables that will be used -- and for all the predecessors of those successors, and the
3237 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3238 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3239 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3240 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3241 // of local variable numbers, so we represent them with the base local variable number), returns that.
3242 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3243 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3244 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3245 // on which kind of member of the clique the block is).
3246 unsigned impGetSpillTmpBase(BasicBlock* block);
3248 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3249 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3250 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3251 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3252 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3253 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3254 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3255 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3256 // successors receive a native int. Similarly float and double are unified to double.
3257 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3258 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3259 // predecessors, so they insert an upcast if needed).
3260 void impReimportSpillClique(BasicBlock* block);
3262 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3263 // block, and represent the predecessor and successor members of the clique currently being computed.
3264 // *** Access to these will need to be locked in a parallel compiler.
3265 ExpandArray<BYTE> impSpillCliquePredMembers;
3266 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3274 // Abstract class for receiving a callback while walking a spill clique
3275 class SpillCliqueWalker
3278 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3281 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3282 class SetSpillTempsBase : public SpillCliqueWalker
3287 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3290 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3293 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3294 class ReimportSpillClique : public SpillCliqueWalker
3299 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3302 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3305 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3306 // predecessor or successor within the spill clique
3307 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3309 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3310 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3311 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3312 void impRetypeEntryStateTemps(BasicBlock* blk);
3314 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3315 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3317 void impPushVar(GenTree* op, typeInfo tiRetVal);
3318 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3319 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3321 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3323 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3324 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3325 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3328 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3331 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3332 struct BlockListNode
3335 BlockListNode* m_next;
3336 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3339 void* operator new(size_t sz, Compiler* comp);
3341 BlockListNode* impBlockListNodeFreeList;
3343 BlockListNode* AllocBlockListNode();
3344 void FreeBlockListNode(BlockListNode* node);
3346 bool impIsValueType(typeInfo* pTypeInfo);
3347 var_types mangleVarArgsType(var_types type);
3350 regNumber getCallArgIntRegister(regNumber floatReg);
3351 regNumber getCallArgFloatRegister(regNumber intReg);
3352 #endif // FEATURE_VARARG
3355 static unsigned jitTotalMethodCompiled;
3359 static LONG jitNestingLevel;
3362 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3364 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3366 // STATIC inlining decision based on the IL code.
3367 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3368 CORINFO_METHOD_INFO* methInfo,
3370 InlineResult* inlineResult);
3372 void impCheckCanInline(GenTreePtr call,
3373 CORINFO_METHOD_HANDLE fncHandle,
3375 CORINFO_CONTEXT_HANDLE exactContextHnd,
3376 InlineCandidateInfo** ppInlineCandidateInfo,
3377 InlineResult* inlineResult);
3379 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3380 GenTreePtr curArgVal,
3382 InlineResult* inlineResult);
3384 void impInlineInitVars(InlineInfo* pInlineInfo);
3386 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3388 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3390 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3392 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3393 GenTreePtr variableBeingDereferenced,
3394 InlArgInfo* inlArgInfo);
3396 void impMarkInlineCandidate(GenTreePtr call,
3397 CORINFO_CONTEXT_HANDLE exactContextHnd,
3398 bool exactContextNeedsRuntimeLookup,
3399 CORINFO_CALL_INFO* callInfo);
3401 bool impTailCallRetTypeCompatible(var_types callerRetType,
3402 CORINFO_CLASS_HANDLE callerRetTypeClass,
3403 var_types calleeRetType,
3404 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3406 bool impIsTailCallILPattern(bool tailPrefixed,
3408 const BYTE* codeAddrOfNextOpcode,
3409 const BYTE* codeEnd,
3411 bool* IsCallPopRet = nullptr);
3413 bool impIsImplicitTailCallCandidate(
3414 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3416 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3419 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3420 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3424 XX Info about the basic-blocks, their contents and the flow analysis XX
3426 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3427 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3431 BasicBlock* fgFirstBB; // Beginning of the basic block list
3432 BasicBlock* fgLastBB; // End of the basic block list
3433 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3434 #if FEATURE_EH_FUNCLETS
3435 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3437 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3439 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3440 unsigned fgEdgeCount; // # of control flow edges between the BBs
3441 unsigned fgBBcount; // # of BBs in the method
3443 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3445 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3446 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3447 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3448 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3450 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3451 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3452 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3453 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3454 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3455 // index). The arrays are of size fgBBNumMax + 1.
3456 unsigned* fgDomTreePreOrder;
3457 unsigned* fgDomTreePostOrder;
3459 bool fgBBVarSetsInited;
3461 // Allocate array like T* a = new T[fgBBNumMax + 1];
3462 // Using helper so we don't keep forgetting +1.
3463 template <typename T>
3464 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3466 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3469 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3470 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3471 // cannot be meaningfully combined. Note that new blocks can be created with higher
3472 // block numbers without changing the basic block epoch. These blocks *cannot*
3473 // participate in a block set until the blocks are all renumbered, causing the epoch
3474 // to change. This is useful if continuing to use previous block sets is valuable.
3475 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3476 unsigned fgCurBBEpoch;
3478 unsigned GetCurBasicBlockEpoch()
3480 return fgCurBBEpoch;
3483 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3484 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3485 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3486 unsigned fgCurBBEpochSize;
3488 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3489 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3490 unsigned fgBBSetCountInSizeTUnits;
3492 void NewBasicBlockEpoch()
3494 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3496 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3498 fgCurBBEpochSize = fgBBNumMax + 1;
3499 fgBBSetCountInSizeTUnits =
3500 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3503 // All BlockSet objects are now invalid!
3504 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3505 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3509 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3510 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3511 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3512 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3514 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3515 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3516 // array of size_t bitsets), then print that out.
3517 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3524 void EnsureBasicBlockEpoch()
3526 if (fgCurBBEpochSize != fgBBNumMax + 1)
3528 NewBasicBlockEpoch();
3532 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3533 void fgEnsureFirstBBisScratch();
3534 bool fgFirstBBisScratch();
3535 bool fgBBisScratch(BasicBlock* block);
3537 void fgExtendEHRegionBefore(BasicBlock* block);
3538 void fgExtendEHRegionAfter(BasicBlock* block);
3540 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3542 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3544 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3547 BasicBlock* nearBlk,
3548 bool putInFilter = false,
3549 bool runRarely = false,
3550 bool insertAtEnd = false);
3552 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3554 bool runRarely = false,
3555 bool insertAtEnd = false);
3557 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3559 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3560 BasicBlock* afterBlk,
3561 unsigned xcptnIndex,
3562 bool putInTryRegion);
3564 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3565 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3566 void fgUnlinkBlock(BasicBlock* block);
3568 unsigned fgMeasureIR();
3570 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3571 bool fgMultipleNots;
3574 bool fgModified; // True if the flow graph has been modified recently
3575 bool fgComputePredsDone; // Have we computed the bbPreds list
3576 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3577 bool fgDomsComputed; // Have we computed the dominator sets?
3578 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3580 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3581 bool fgHasPostfix; // any postfix ++/-- found?
3582 unsigned fgIncrCount; // number of increment nodes found
3584 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3588 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3589 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3592 bool fgRemoveRestOfBlock; // true if we know that we will throw
3593 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3595 // There are two modes for ordering of the trees.
3596 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3597 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3598 // by traversing the tree according to the order of the operands.
3599 // - In FGOrderLinear, the dominant ordering is the linear order.
3606 FlowGraphOrder fgOrder;
3608 // The following are boolean flags that keep track of the state of internal data structures
3610 bool fgStmtListThreaded;
3611 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3612 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3613 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3614 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3615 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3616 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3617 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
3618 // This is derived from the profile data
3619 // or is BB_UNITY_WEIGHT when we don't have profile data
3621 #if FEATURE_EH_FUNCLETS
3622 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3623 #endif // FEATURE_EH_FUNCLETS
3625 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3626 // since fgMorphTree can be called from several places
3628 bool impBoxTempInUse; // the temp below is valid and available
3629 unsigned impBoxTemp; // a temporary that is used for boxing
3632 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3633 // and we are trying to compile again in a "safer", minopts mode?
3637 unsigned impInlinedCodeSize;
3640 //-------------------------------------------------------------------------
3646 void fgTransformFatCalli();
3650 void fgRemoveEmptyTry();
3652 void fgRemoveEmptyFinally();
3654 void fgMergeFinallyChains();
3656 void fgCloneFinally();
3658 void fgCleanupContinuation(BasicBlock* continuation);
3660 void fgUpdateFinallyTargetFlags();
3662 void fgClearAllFinallyTargetBits();
3664 void fgAddFinallyTargetFlags();
3666 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3667 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
3668 // when this is necessary.
3669 bool fgNeedToAddFinallyTargetBits;
3670 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
3672 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
3673 BasicBlock* handler,
3674 BlockToBlockMap& continuationMap);
3676 GenTreePtr fgGetCritSectOfStaticMethod();
3678 #if FEATURE_EH_FUNCLETS
3680 void fgAddSyncMethodEnterExit();
3682 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3684 void fgConvertSyncReturnToLeave(BasicBlock* block);
3686 #endif // FEATURE_EH_FUNCLETS
3688 void fgAddReversePInvokeEnterExit();
3690 bool fgMoreThanOneReturnBlock();
3692 // The number of separate return points in the method.
3693 unsigned fgReturnCount;
3695 void fgAddInternal();
3697 bool fgFoldConditional(BasicBlock* block);
3699 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3700 void fgMorphBlocks();
3702 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3704 void fgCheckArgCnt();
3705 void fgSetOptions();
3708 static fgWalkPreFn fgAssertNoQmark;
3709 void fgPreExpandQmarkChecks(GenTreePtr expr);
3710 void fgPostExpandQmarkChecks();
3711 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3714 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3716 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3717 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3718 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3719 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3720 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3722 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3723 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3724 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3725 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3727 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3728 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3729 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3730 void fgExpandQmarkNodes();
3734 // Do "simple lowering." This functionality is (conceptually) part of "general"
3735 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3736 void fgSimpleLowering();
3738 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3740 GenTreePtr fgInitThisClass();
3742 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3744 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3746 inline bool backendRequiresLocalVarLifetimes()
3748 #if defined(LEGACY_BACKEND)
3751 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
3755 void fgLocalVarLiveness();
3757 void fgLocalVarLivenessInit();
3759 #ifdef LEGACY_BACKEND
3760 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3762 void fgPerNodeLocalVarLiveness(GenTree* node);
3764 void fgPerBlockLocalVarLiveness();
3766 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3768 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3770 // This is used in the liveness computation, as a temporary. When we use the
3771 // arbitrary-length VarSet representation, it is better not to allocate a new one
3773 VARSET_TP fgMarkIntfUnionVS;
3775 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3777 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3779 bool fgMarkIntf(VARSET_VALARG_TP varSet1, unsigned varIndex);
3781 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3783 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3785 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3787 void fgComputeLifeTrackedLocalUse(VARSET_TP& life, LclVarDsc& varDsc, GenTreeLclVarCommon* node);
3788 bool fgComputeLifeTrackedLocalDef(VARSET_TP& life,
3789 VARSET_VALARG_TP keepAliveVars,
3791 GenTreeLclVarCommon* node);
3792 void fgComputeLifeUntrackedLocal(VARSET_TP& life,
3793 VARSET_VALARG_TP keepAliveVars,
3795 GenTreeLclVarCommon* lclVarNode,
3797 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode, GenTree* node);
3799 void fgComputeLife(VARSET_TP& life,
3800 GenTreePtr startNode,
3802 VARSET_VALARG_TP volatileVars,
3803 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3805 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3807 bool fgRemoveDeadStore(GenTree** pTree,
3809 VARSET_VALARG_TP life,
3811 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3813 // For updating liveset during traversal AFTER fgComputeLife has completed
3814 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3815 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3817 // Returns the set of live variables after endTree,
3818 // assuming that liveSet is the set of live variables BEFORE tree.
3819 // Requires that fgComputeLife has completed, and that tree is in the same
3820 // statement as endTree, and that it comes before endTree in execution order
3822 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3824 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
3825 while (tree != nullptr && tree != endTree->gtNext)
3827 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3828 tree = tree->gtNext;
3830 assert(tree == endTree->gtNext);
3834 void fgInterBlockLocalVarLiveness();
3836 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3837 // "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
3838 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3839 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3840 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3841 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3842 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3844 if (m_opAsgnVarDefSsaNums == nullptr)
3846 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3848 return m_opAsgnVarDefSsaNums;
3851 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3852 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3853 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3855 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3857 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3858 // Except: assumes that lcl is a def, and if it is
3859 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3860 // rather than the "use" SSA number recorded in the tree "lcl".
3861 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3863 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3864 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3865 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3866 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3867 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3869 // (byref addrS1 = &s1,
3870 // *(addrS1 * offsetof(f0)) = s2f0,
3872 // *(addrS1 * offsetof(fn)) = s2fn)
3874 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3875 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3876 // give it SSA names and value numbers?
3878 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3879 // end with an instance of the structure below, whose fields are described in the declaration.
3880 struct IndirectAssignmentAnnotation
3882 unsigned m_lclNum; // The local num that is being indirectly assigned.
3883 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3884 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3885 // be the singleton field sequence "g". The individual assignments would
3886 // further append the fields of "s.g" to that.
3887 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3888 // structure has a single field).
3889 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3890 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3893 IndirectAssignmentAnnotation(unsigned lclNum,
3894 FieldSeqNode* fldSeq,
3896 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3897 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3898 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3902 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3903 NodeToIndirAssignMap;
3904 NodeToIndirAssignMap* m_indirAssignMap;
3905 NodeToIndirAssignMap* GetIndirAssignMap()
3907 if (m_indirAssignMap == nullptr)
3909 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3910 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3911 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3913 return m_indirAssignMap;
3916 // Performs SSA conversion.
3919 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3920 void fgResetForSsa();
3922 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3924 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3925 inline bool fgExcludeFromSsa(unsigned lclNum);
3927 // The value numbers for this compilation.
3928 ValueNumStore* vnStore;
3931 ValueNumStore* GetValueNumStore()
3936 // Do value numbering (assign a value number to each
3938 void fgValueNumber();
3940 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3941 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3942 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3943 // match the element type of the array or fldSeq. When this type doesn't match
3944 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3946 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3949 FieldSeqNode* fldSeq,
3953 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3954 // has been parsed to yield the other input arguments. If evaluation of the address
3955 // can raise exceptions, those should be captured in the exception set "excVN."
3956 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3957 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3958 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3959 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3960 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3962 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3963 CORINFO_CLASS_HANDLE elemTypeEq,
3967 FieldSeqNode* fldSeq);
3969 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3970 // by evaluating the array index expression "tree". Returns the value number resulting from
3971 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3972 // "GT_IND" that does the dereference, and it is given the returned value number.
3973 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3975 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3976 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3978 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3980 // Utility functions for fgValueNumber.
3982 // Perform value-numbering for the trees in "blk".
3983 void fgValueNumberBlock(BasicBlock* blk);
3985 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3986 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3987 // assumed for the memoryKind at the start "entryBlk".
3988 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3990 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3991 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3992 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3994 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3996 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3998 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3999 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
4000 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
4002 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
4003 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
4005 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
4006 // value in that SSA #.
4007 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
4009 // The input 'tree' is a leaf node that is a constant
4010 // Assign the proper value number to the tree
4011 void fgValueNumberTreeConst(GenTreePtr tree);
4013 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
4014 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4016 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
4018 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
4020 // Does value-numbering for a block assignment.
4021 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
4023 // Does value-numbering for a cast tree.
4024 void fgValueNumberCastTree(GenTreePtr tree);
4026 // Does value-numbering for an intrinsic tree.
4027 void fgValueNumberIntrinsic(GenTreePtr tree);
4029 // Does value-numbering for a call. We interpret some helper calls.
4030 void fgValueNumberCall(GenTreeCall* call);
4032 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4033 void fgUpdateArgListVNs(GenTreeArgList* args);
4035 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4036 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4038 // Requires "helpCall" to be a helper call. Assigns it a value number;
4039 // we understand the semantics of some of the calls. Returns "true" if
4040 // the call may modify the heap (we assume arbitrary memory side effects if so).
4041 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4043 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
4044 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
4046 // These are the current value number for the memory implicit variables while
4047 // doing value numbering. These are the value numbers under the "liberal" interpretation
4048 // of memory values; the "conservative" interpretation needs no VN, since every access of
4049 // memory yields an unknown value.
4050 ValueNum fgCurMemoryVN[MemoryKindCount];
4052 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4053 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4054 // is 1, and the rest is an encoding of "elemTyp".
4055 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4057 if (elemStructType != nullptr)
4059 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4060 varTypeIsIntegral(elemTyp));
4061 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4062 return elemStructType;
4066 elemTyp = varTypeUnsignedToSigned(elemTyp);
4067 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4070 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4071 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4072 // the struct type of the element).
4073 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4075 size_t clsHndVal = size_t(clsHnd);
4076 if (clsHndVal & 0x1)
4078 return var_types(clsHndVal >> 1);
4086 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4087 var_types getJitGCType(BYTE gcType);
4089 enum structPassingKind
4091 SPK_Unknown, // Invalid value, never returned
4092 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4093 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4094 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4095 // parameters registers are used, then the stack will be used)
4096 // for X86 passed on the stack, for ARM32 passed in registers
4097 // or the stack or split between registers and the stack.
4098 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4100 }; // The struct is passed/returned by reference to a copy/buffer.
4102 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4103 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4104 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4105 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4107 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
4109 // Get the type that is used to pass values of the given struct type.
4110 // If you have already retrieved the struct size then pass it as the optional third argument
4112 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4113 structPassingKind* wbPassStruct,
4114 unsigned structSize = 0);
4116 // Get the type that is used to return values of the given struct type.
4117 // If you have already retrieved the struct size then pass it as the optional third argument
4119 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4120 structPassingKind* wbPassStruct = nullptr,
4121 unsigned structSize = 0);
4124 // Print a representation of "vnp" or "vn" on standard output.
4125 // If "level" is non-zero, we also print out a partial expansion of the value.
4126 void vnpPrint(ValueNumPair vnp, unsigned level);
4127 void vnPrint(ValueNum vn, unsigned level);
4130 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4132 // Dominator computation member functions
4133 // Not exposed outside Compiler
4135 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4137 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4138 // flow graph. We first assume the fields bbIDom on each
4139 // basic block are invalid. This computation is needed later
4140 // by fgBuildDomTree to build the dominance tree structure.
4141 // Based on: A Simple, Fast Dominance Algorithm
4142 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4144 void fgCompDominatedByExceptionalEntryBlocks();
4146 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4147 // Note: this is relatively slow compared to calling fgDominate(),
4148 // especially if dealing with a single block versus block check.
4150 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4152 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4154 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4156 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4158 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4160 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4161 // processed in topological sort, this function takes care of that.
4163 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4165 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4166 // Returns this as a set.
4168 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4169 // root nodes. Returns this as a set.
4172 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4175 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4176 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4179 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4180 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4181 // && postOrder(A) >= postOrder(B) making the computation O(1).
4182 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4184 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4186 void fgUpdateChangedFlowGraph();
4189 // Compute the predecessors of the blocks in the control flow graph.
4190 void fgComputePreds();
4192 // Remove all predecessor information.
4193 void fgRemovePreds();
4195 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4196 // before the full predecessors lists are computed.
4197 void fgComputeCheapPreds();
4200 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4202 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4212 // Initialize the per-block variable sets (used for liveness analysis).
4213 void fgInitBlockVarSets();
4215 // true if we've gone through and created GC Poll calls.
4216 bool fgGCPollsCreated;
4217 void fgMarkGCPollBlocks();
4218 void fgCreateGCPolls();
4219 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4221 // Requires that "block" is a block that returns from
4222 // a finally. Returns the number of successors (jump targets of
4223 // of blocks in the covered "try" that did a "LEAVE".)
4224 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4226 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4227 // a finally. Returns its "i"th successor (jump targets of
4228 // of blocks in the covered "try" that did a "LEAVE".)
4229 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4230 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4233 // Factor out common portions of the impls of the methods above.
4234 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4237 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4238 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4239 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4240 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4241 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4242 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4243 // we leave the entry associated with the block, but it will no longer be accessed.)
4244 struct SwitchUniqueSuccSet
4246 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4247 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4250 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4251 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4252 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4253 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4256 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4257 BlockToSwitchDescMap;
4260 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4261 // iteration over only the distinct successors.
4262 BlockToSwitchDescMap* m_switchDescMap;
4265 BlockToSwitchDescMap* GetSwitchDescMap()
4267 if (m_switchDescMap == nullptr)
4269 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4271 return m_switchDescMap;
4274 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4275 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4276 // we don't accidentally look up and return the wrong switch data.
4277 void InvalidateUniqueSwitchSuccMap()
4279 m_switchDescMap = nullptr;
4282 // Requires "switchBlock" to be a block that ends in a switch. Returns
4283 // the corresponding SwitchUniqueSuccSet.
4284 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4286 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4287 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4288 // remove it from "this", and ensure that "to" is a member.
4289 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4291 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4292 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4294 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4296 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4298 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4300 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4302 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4304 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4306 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4308 void fgRemoveBlockAsPred(BasicBlock* block);
4310 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4312 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4314 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4316 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4318 flowList* fgAddRefPred(BasicBlock* block,
4319 BasicBlock* blockPred,
4320 flowList* oldEdge = nullptr,
4321 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4324 void fgFindBasicBlocks();
4326 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4328 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4330 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4331 bool putInTryRegion,
4332 BasicBlock* startBlk,
4334 BasicBlock* nearBlk,
4335 BasicBlock* jumpBlk,
4338 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4340 void fgRemoveEmptyBlocks();
4342 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4344 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4346 void fgCreateLoopPreHeader(unsigned lnum);
4348 void fgUnreachableBlock(BasicBlock* block);
4350 void fgRemoveConditionalJump(BasicBlock* block);
4352 BasicBlock* fgLastBBInMainFunction();
4354 BasicBlock* fgEndBBAfterMainFunction();
4356 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4358 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4360 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4362 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4364 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4366 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4368 bool fgRenumberBlocks();
4370 bool fgExpandRarelyRunBlocks();
4372 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4374 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4376 enum FG_RELOCATE_TYPE
4378 FG_RELOCATE_TRY, // relocate the 'try' region
4379 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4381 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4383 #if FEATURE_EH_FUNCLETS
4384 #if defined(_TARGET_ARM_)
4385 void fgClearFinallyTargetBit(BasicBlock* block);
4386 #endif // defined(_TARGET_ARM_)
4387 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4388 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4389 void fgInsertFuncletPrologBlock(BasicBlock* block);
4390 void fgCreateFuncletPrologBlocks();
4391 void fgCreateFunclets();
4392 #else // !FEATURE_EH_FUNCLETS
4393 bool fgRelocateEHRegions();
4394 #endif // !FEATURE_EH_FUNCLETS
4396 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4398 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4400 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4402 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4404 bool fgOptimizeEmptyBlock(BasicBlock* block);
4406 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4408 bool fgOptimizeBranch(BasicBlock* bJump);
4410 bool fgOptimizeSwitchBranches(BasicBlock* block);
4412 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4414 bool fgOptimizeSwitchJumps();
4416 void fgPrintEdgeWeights();
4418 void fgComputeEdgeWeights();
4420 void fgReorderBlocks();
4422 void fgDetermineFirstColdBlock();
4424 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4426 bool fgUpdateFlowGraph(bool doTailDup = false);
4428 void fgFindOperOrder();
4430 // method that returns if you should split here
4431 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4433 void fgSetBlockOrder();
4435 void fgRemoveReturnBlock(BasicBlock* block);
4437 /* Helper code that has been factored out */
4438 inline void fgConvertBBToThrowBB(BasicBlock* block);
4440 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4441 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4442 GenTreePtr fgMakeTmpArgNode(
4443 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4445 // The following check for loops that don't execute calls
4446 bool fgLoopCallMarked;
4448 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4449 void fgLoopCallMark();
4451 void fgMarkLoopHead(BasicBlock* block);
4453 unsigned fgGetCodeEstimate(BasicBlock* block);
4456 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4457 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4458 bool fgDumpFlowGraph(Phases phase);
4460 #endif // DUMP_FLOWGRAPHS
4465 void fgDispBBLiveness(BasicBlock* block);
4466 void fgDispBBLiveness();
4467 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4468 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4469 void fgDispBasicBlocks(bool dumpTrees = false);
4470 void fgDumpStmtTree(GenTreePtr stmt, unsigned bbNum);
4471 void fgDumpBlock(BasicBlock* block);
4472 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4474 static fgWalkPreFn fgStress64RsltMulCB;
4475 void fgStress64RsltMul();
4476 void fgDebugCheckUpdate();
4477 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4478 void fgDebugCheckBlockLinks();
4479 void fgDebugCheckLinks(bool morphTrees = false);
4480 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4481 void fgDebugCheckFlags(GenTreePtr tree);
4482 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4483 void fgDebugCheckTryFinallyExits();
4486 #ifdef LEGACY_BACKEND
4487 static void fgOrderBlockOps(GenTreePtr tree,
4491 GenTreePtr* opsPtr, // OUT
4492 regMaskTP* regsPtr); // OUT
4493 #endif // LEGACY_BACKEND
4495 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4496 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4497 void fgTraverseRPO();
4499 //--------------------- Walking the trees in the IR -----------------------
4504 fgWalkPreFn* wtprVisitorFn;
4505 fgWalkPostFn* wtpoVisitorFn;
4506 void* pCallbackData; // user-provided data
4507 bool wtprLclsOnly; // whether to only visit lclvar nodes
4508 GenTreePtr parent; // parent of current node, provided to callback
4509 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4511 bool printModified; // callback can use this
4515 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4516 fgWalkPreFn* visitor,
4517 void* pCallBackData = nullptr,
4518 bool lclVarsOnly = false,
4519 bool computeStack = false);
4521 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4522 fgWalkPreFn* preVisitor,
4523 fgWalkPostFn* postVisitor,
4524 void* pCallBackData = nullptr);
4526 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4530 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4531 fgWalkPostFn* visitor,
4532 void* pCallBackData = nullptr,
4533 bool computeStack = false);
4535 // An fgWalkPreFn that looks for expressions that have inline throws in
4536 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4537 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4538 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4539 // properly propagated to parent trees). It returns WALK_CONTINUE
4541 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4542 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4543 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4545 /**************************************************************************
4547 *************************************************************************/
4550 friend class SsaBuilder;
4551 friend struct ValueNumberState;
4553 //--------------------- Detect the basic blocks ---------------------------
4555 BasicBlock** fgBBs; // Table of pointers to the BBs
4557 void fgInitBBLookup();
4558 BasicBlock* fgLookupBB(unsigned addr);
4560 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4562 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4564 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4566 void fgLinkBasicBlocks();
4568 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4570 void fgCheckBasicBlockControlFlow();
4572 void fgControlFlowPermitted(BasicBlock* blkSrc,
4573 BasicBlock* blkDest,
4574 BOOL IsLeave = false /* is the src a leave block */);
4576 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4578 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4580 void fgAdjustForAddressExposedOrWrittenThis();
4582 bool fgProfileData_ILSizeMismatch;
4583 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4584 ULONG fgProfileBufferCount;
4585 ULONG fgNumProfileRuns;
4587 unsigned fgStressBBProf()
4590 unsigned result = JitConfig.JitStressBBProf();
4593 if (compStressCompile(STRESS_BB_PROFILE, 15))
4604 bool fgHaveProfileData();
4605 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4606 void fgInstrumentMethod();
4609 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
4610 // or if we have some fake profile data for the stress mode
4611 bool fgIsUsingProfileWeights()
4613 return (fgHaveProfileData() || fgStressBBProf());
4616 // fgProfileRunsCount - returns total number of scenario runs for the profile data
4617 // or BB_UNITY_WEIGHT when we aren't using profile data.
4618 unsigned fgProfileRunsCount()
4620 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
4623 //-------- Insert a statement at the start or end of a basic block --------
4627 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4631 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4633 public: // Used by linear scan register allocation
4634 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4637 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4638 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4640 public: // Used by linear scan register allocation
4641 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4644 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4646 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4648 // Create a new temporary variable to hold the result of *ppTree,
4649 // and transform the graph accordingly.
4650 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4651 GenTree* fgMakeMultiUse(GenTree** ppTree);
4654 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4655 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4656 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4658 //-------- Determine the order in which the trees will be evaluated -------
4660 unsigned fgTreeSeqNum;
4661 GenTree* fgTreeSeqLst;
4662 GenTree* fgTreeSeqBeg;
4664 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4665 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4666 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4667 void fgSetStmtSeq(GenTree* tree);
4668 void fgSetBlockOrder(BasicBlock* block);
4670 //------------------------- Morphing --------------------------------------
4672 unsigned fgPtrArgCntCur;
4673 unsigned fgPtrArgCntMax;
4674 hashBv* fgOutgoingArgTemps;
4675 hashBv* fgCurrentlyInUseArgTemps;
4677 bool compCanEncodePtrArgCntMax();
4679 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4681 BasicBlock* fgSetRngChkTargetInner(SpecialCodeKind kind, bool delay, unsigned* stkDepth);
4684 void fgMoveOpsLeft(GenTreePtr tree);
4687 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4689 bool fgIsThrow(GenTreePtr tree);
4691 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4692 bool fgIsBlockCold(BasicBlock* block);
4694 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4696 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4698 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4700 bool fgMorphRelopToQmark(GenTreePtr tree);
4702 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4703 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4704 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4705 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4706 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4707 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4708 // small; hence the other fields of MorphAddrContext.
4709 enum MorphAddrContextKind
4714 struct MorphAddrContext
4716 MorphAddrContextKind m_kind;
4717 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4718 // top-level indirection and here have been constants.
4719 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4720 // In that case, is the sum of those constant offsets.
4722 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4727 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4728 static MorphAddrContext s_CopyBlockMAC;
4731 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4732 var_types* baseTypeOut,
4734 unsigned* simdSizeOut,
4735 bool ignoreUsedInSIMDIntrinsic = false);
4736 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4737 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4738 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4739 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4741 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4742 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4743 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4745 #endif // FEATURE_SIMD
4746 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4747 GenTreePtr fgMorphCast(GenTreePtr tree);
4748 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4749 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4751 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4754 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4755 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4757 void fgFixupStructReturn(GenTreePtr call);
4758 GenTreePtr fgMorphLocalVar(GenTreePtr tree, bool forceRemorph);
4761 bool fgAddrCouldBeNull(GenTreePtr addr);
4764 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4765 bool fgCanFastTailCall(GenTreeCall* call);
4766 void fgMorphTailCall(GenTreeCall* call);
4767 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4768 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4769 fgArgTabEntryPtr argTabEntry,
4771 IL_OFFSETX callILOffset,
4772 GenTreePtr tmpAssignmentInsertionPoint,
4773 GenTreePtr paramAssignmentInsertionPoint);
4774 static int fgEstimateCallStackSize(GenTreeCall* call);
4775 GenTreePtr fgMorphCall(GenTreeCall* call);
4776 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4777 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4779 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4780 static fgWalkPreFn fgFindNonInlineCandidate;
4782 GenTreePtr fgOptimizeDelegateConstructor(GenTreeCall* call,
4783 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
4784 CORINFO_RESOLVED_TOKEN* ldftnToken);
4785 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4786 void fgAssignSetVarDef(GenTreePtr tree);
4787 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4788 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4789 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4790 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4791 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4792 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4793 void fgMorphUnsafeBlk(GenTreeObj* obj);
4794 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4795 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4796 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4797 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4798 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4799 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4800 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4802 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4803 GenTreePtr fgMorphConst(GenTreePtr tree);
4806 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4809 #if LOCAL_ASSERTION_PROP
4810 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4811 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4813 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4815 GenTreeStmt* fgMorphStmt;
4817 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4818 // used when morphing big offset.
4820 //----------------------- Liveness analysis -------------------------------
4822 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4823 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4825 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4826 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4827 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4829 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4831 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4833 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4834 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4836 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4837 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4839 void fgExtendDbgScopes();
4840 void fgExtendDbgLifetimes();
4843 void fgDispDebugScopes();
4846 //-------------------------------------------------------------------------
4848 // The following keeps track of any code we've added for things like array
4849 // range checking or explicit calls to enable GC, and so on.
4854 AddCodeDsc* acdNext;
4855 BasicBlock* acdDstBlk; // block to which we jump
4857 SpecialCodeKind acdKind; // what kind of a special block is this?
4858 unsigned short acdStkLvl;
4862 static unsigned acdHelper(SpecialCodeKind codeKind);
4864 AddCodeDsc* fgAddCodeList;
4866 bool fgRngChkThrowAdded;
4867 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4869 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4871 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4874 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4877 bool fgIsCodeAdded();
4879 bool fgIsThrowHlpBlk(BasicBlock* block);
4880 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4882 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4884 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4885 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4886 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4887 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4888 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4890 #if FEATURE_MULTIREG_RET
4891 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4892 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4893 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4894 #endif // FEATURE_MULTIREG_RET
4896 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4899 static fgWalkPreFn fgDebugCheckInlineCandidates;
4901 void CheckNoFatPointerCandidatesLeft();
4902 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4905 void fgPromoteStructs();
4906 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4907 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4909 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
4910 void fgMarkImplicitByRefArgs();
4912 // Change implicit byrefs' types from struct to pointer, and for any that were
4913 // promoted, create new promoted struct temps.
4914 void fgRetypeImplicitByRefArgs();
4916 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
4917 bool fgMorphImplicitByRefArgs(GenTreePtr tree);
4918 GenTreePtr fgMorphImplicitByRefArgs(GenTreePtr tree, bool isAddr);
4920 // Clear up annotations for any struct promotion temps created for implicit byrefs.
4921 void fgMarkDemotedImplicitByRefArgs();
4923 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4924 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4925 void fgMarkAddressExposedLocals();
4926 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4928 static fgWalkPreFn fgUpdateSideEffectsPre;
4929 static fgWalkPostFn fgUpdateSideEffectsPost;
4931 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4933 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4935 // The given local variable, required to be a struct variable, is being assigned via
4936 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4937 // the variable is not enregistered, and is therefore not promoted independently.
4938 void fgLclFldAssign(unsigned lclNum);
4940 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4941 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4942 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
4943 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4946 bool fgPrintInlinedMethods;
4949 bool fgIsBigOffset(size_t offset);
4951 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4952 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4953 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4954 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4955 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4958 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4959 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4963 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4964 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4971 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4974 void optRemoveRangeCheck(GenTreePtr tree, GenTreePtr stmt);
4975 bool optIsRangeCheckRemovable(GenTreePtr tree);
4978 static fgWalkPreFn optValidRangeCheckIndex;
4979 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4982 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4984 /**************************************************************************
4986 *************************************************************************/
4989 // Do hoisting for all loops.
4990 void optHoistLoopCode();
4992 // To represent sets of VN's that have already been hoisted in outer loops.
4993 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4994 typedef VNToBoolMap VNSet;
4996 struct LoopHoistContext
4999 // The set of variables hoisted in the current loop (or nullptr if there are none).
5000 VNSet* m_pHoistedInCurLoop;
5003 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
5004 VNSet m_hoistedInParentLoops;
5005 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
5006 // Previous decisions on loop-invariance of value numbers in the current loop.
5007 VNToBoolMap m_curLoopVnInvariantCache;
5009 VNSet* GetHoistedInCurLoop(Compiler* comp)
5011 if (m_pHoistedInCurLoop == nullptr)
5013 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
5015 return m_pHoistedInCurLoop;
5018 VNSet* ExtractHoistedInCurLoop()
5020 VNSet* res = m_pHoistedInCurLoop;
5021 m_pHoistedInCurLoop = nullptr;
5025 LoopHoistContext(Compiler* comp)
5026 : m_pHoistedInCurLoop(nullptr)
5027 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5028 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5033 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5034 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5035 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5036 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5038 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5039 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5040 // "m_hoistedInParentLoops".
5042 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5044 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5045 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5046 // expressions to "hoistInLoop".
5047 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5049 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5050 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
5052 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5053 // that are invariant in loop "lnum" (an index into the optLoopTable)
5054 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5055 // expressions to "hoistInLoop".
5056 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5057 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5058 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5059 bool optHoistLoopExprsForTree(GenTreePtr tree,
5061 LoopHoistContext* hoistCtxt,
5062 bool* firstBlockAndBeforeSideEffect,
5064 bool* pCctorDependent);
5066 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5067 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5069 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5070 // Constants and init values are always loop invariant.
5071 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5072 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5074 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5075 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5076 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5077 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5078 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
5080 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5081 // in the loop table.
5082 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5084 // Records the set of "side effects" of all loops: fields (object instance and static)
5085 // written to, and SZ-array element type equivalence classes updated.
5086 void optComputeLoopSideEffects();
5089 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5090 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5091 // static) written to, and SZ-array element type equivalence classes updated.
5092 void optComputeLoopNestSideEffects(unsigned lnum);
5094 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5095 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5097 // Hoist the expression "expr" out of loop "lnum".
5098 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
5101 void optOptimizeBools();
5104 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5106 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5109 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5111 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5112 // the loop into a "do-while" loop
5113 // Also finds all natural loops and records them in the loop table
5115 // Optionally clone loops in the loop table.
5116 void optCloneLoops();
5118 // Clone loop "loopInd" in the loop table.
5119 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5121 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5122 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5123 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5125 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5127 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5130 // This enumeration describes what is killed by a call.
5134 CALLINT_NONE, // no interference (most helpers)
5135 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5136 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5137 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5138 CALLINT_ALL, // kills everything (normal method call)
5142 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5143 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5144 // in bbNext order; we use comparisons on the bbNum to decide order.)
5145 // The blocks that define the body are
5146 // first <= top <= entry <= bottom .
5147 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5148 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5149 // Compiler::optFindNaturalLoops().
5152 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5153 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5154 // loop, but not the outer loop.)
5155 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5157 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5158 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5159 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5161 callInterf lpAsgCall; // "callInterf" for calls in the loop
5162 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5163 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5165 unsigned short lpFlags; // Mask of the LPFLG_* constants
5167 unsigned char lpExitCnt; // number of exits from the loop
5169 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5170 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5171 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5172 // (Actually, an "immediately" nested loop --
5173 // no other child of this loop is a parent of lpChild.)
5174 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5175 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5176 // by following "lpChild" then "lpSibling" links.
5178 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5179 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5181 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5182 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5183 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5185 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5186 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5188 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5189 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5190 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5191 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5193 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5194 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5195 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5197 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5198 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5199 // type are assigned to.
5201 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5202 // memory side effects. If this is set, the fields below
5203 // may not be accurate (since they become irrelevant.)
5204 bool lpContainsCall; // True if executing the loop body *may* execute a call
5206 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5207 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5209 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5211 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5212 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5214 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5216 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5217 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5219 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5220 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5222 JitSimplerHashBehavior>
5224 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5225 // instance fields modified
5228 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5229 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5231 JitSimplerHashBehavior>
5233 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5234 // arrays of that type are modified
5237 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5238 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5240 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5241 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5242 // (shifted left, with a low-order bit set to distinguish.)
5243 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5244 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5246 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5248 GenTreePtr lpIterTree; // The "i <op>= const" tree
5249 unsigned lpIterVar(); // iterator variable #
5250 int lpIterConst(); // the constant with which the iterator is incremented
5251 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5252 void VERIFY_lpIterTree();
5254 var_types lpIterOperType(); // For overflow instructions
5257 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5258 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5262 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5264 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5265 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5266 void VERIFY_lpTestTree();
5268 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5269 GenTreePtr lpIterator(); // the iterator node in the loop test
5270 GenTreePtr lpLimit(); // the limit node in the loop test
5272 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5273 // LPFLG_CONST_LIMIT
5274 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5276 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5277 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5278 // LPFLG_ARRLEN_LIMIT
5280 // Returns "true" iff "*this" contains the blk.
5281 bool lpContains(BasicBlock* blk)
5283 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5285 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5286 // to be equal, but requiring bottoms to be different.)
5287 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5289 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5292 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5293 // bottoms to be different.)
5294 bool lpContains(const LoopDsc& lp2)
5296 return lpContains(lp2.lpFirst, lp2.lpBottom);
5299 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5300 // (allowing firsts to be equal, but requiring bottoms to be different.)
5301 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5303 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5306 // Returns "true" iff "*this" is (properly) contained by "lp2"
5307 // (allowing firsts to be equal, but requiring bottoms to be different.)
5308 bool lpContainedBy(const LoopDsc& lp2)
5310 return lpContains(lp2.lpFirst, lp2.lpBottom);
5313 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5314 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5316 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5318 // Returns "true" iff "*this" is disjoint from "lp2".
5319 bool lpDisjoint(const LoopDsc& lp2)
5321 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5323 // Returns "true" iff the loop is well-formed (see code for defn).
5326 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5327 lpEntry->bbNum <= lpBottom->bbNum &&
5328 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5333 bool fgMightHaveLoop(); // returns true if there are any backedges
5334 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5337 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5338 unsigned char optLoopCount; // number of tracked loops
5340 bool optRecordLoop(BasicBlock* head,
5346 unsigned char exitCnt);
5349 unsigned optCallCount; // number of calls made in the method
5350 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5351 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5352 unsigned optLoopsCloned; // number of loops cloned in the current method.
5355 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5356 void optPrintLoopInfo(unsigned loopNum,
5358 BasicBlock* lpFirst,
5360 BasicBlock* lpEntry,
5361 BasicBlock* lpBottom,
5362 unsigned char lpExitCnt,
5364 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5365 void optPrintLoopInfo(unsigned lnum);
5366 void optPrintLoopRecording(unsigned lnum);
5368 void optCheckPreds();
5371 void optSetBlockWeights();
5373 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5375 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5377 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5379 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5380 unsigned optIsLoopIncrTree(GenTreePtr incr);
5381 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5382 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5383 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5384 bool optExtractInitTestIncr(BasicBlock* head,
5389 GenTreePtr* ppIncr);
5391 void optFindNaturalLoops();
5393 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5394 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5395 bool optCanonicalizeLoopNest(unsigned char loopInd);
5397 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5398 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5399 bool optCanonicalizeLoop(unsigned char loopInd);
5401 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5402 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5403 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5404 bool optLoopContains(unsigned l1, unsigned l2);
5406 // Requires "loopInd" to be a valid index into the loop table.
5407 // Updates the loop table by changing loop "loopInd", whose head is required
5408 // to be "from", to be "to". Also performs this transformation for any
5409 // loop nested in "loopInd" that shares the same head as "loopInd".
5410 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5412 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5413 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5414 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5416 // Marks the containsCall information to "lnum" and any parent loops.
5417 void AddContainsCallAllContainingLoops(unsigned lnum);
5418 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5419 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5420 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5421 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5422 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5423 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5425 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5426 // of "from".) Copies the jump destination from "from" to "to".
5427 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5429 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5430 unsigned optLoopDepth(unsigned lnum)
5432 unsigned par = optLoopTable[lnum].lpParent;
5433 if (par == BasicBlock::NOT_IN_LOOP)
5439 return 1 + optLoopDepth(par);
5443 void fgOptWhileLoop(BasicBlock* block);
5445 bool optComputeLoopRep(int constInit,
5448 genTreeOps iterOper,
5450 genTreeOps testOper,
5453 unsigned* iterCount);
5454 #if FEATURE_STACK_FP_X87
5457 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5458 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5459 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5460 #endif // FEATURE_STACK_FP_X87
5463 static fgWalkPreFn optIsVarAssgCB;
5466 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5468 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5470 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5472 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5474 /**************************************************************************
5475 * Optimization conditions
5476 *************************************************************************/
5478 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5479 bool optPentium4(void);
5480 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5481 bool optAvoidIntMult(void);
5486 // The following is the upper limit on how many expressions we'll keep track
5487 // of for the CSE analysis.
5489 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5491 static const int MIN_CSE_COST = 2;
5493 // Keeps tracked cse indices
5494 BitVecTraits* cseTraits;
5497 /* Generic list of nodes - used by the CSE logic */
5505 typedef struct treeLst* treeLstPtr;
5509 treeStmtLst* tslNext;
5510 GenTreePtr tslTree; // tree node
5511 GenTreePtr tslStmt; // statement containing the tree
5512 BasicBlock* tslBlock; // block containing the statement
5515 typedef struct treeStmtLst* treeStmtLstPtr;
5517 // The following logic keeps track of expressions via a simple hash table.
5521 CSEdsc* csdNextInBucket; // used by the hash table
5523 unsigned csdHashValue; // the orginal hashkey
5525 unsigned csdIndex; // 1..optCSECandidateCount
5526 char csdLiveAcrossCall; // 0 or 1
5528 unsigned short csdDefCount; // definition count
5529 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5531 unsigned csdDefWtCnt; // weighted def count
5532 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5534 GenTreePtr csdTree; // treenode containing the 1st occurance
5535 GenTreePtr csdStmt; // stmt containing the 1st occurance
5536 BasicBlock* csdBlock; // block containing the 1st occurance
5538 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5539 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5541 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5542 // number, this will reflect it; otherwise, NoVN.
5545 static const size_t s_optCSEhashSize;
5546 CSEdsc** optCSEhash;
5549 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, GenTreePtr, JitSimplerHashBehavior> NodeToNodeMap;
5551 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
5552 // re-numbered with the bound to improve range check elimination
5554 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
5555 void optCseUpdateCheckedBoundMap(GenTreePtr compare);
5559 CSEdsc* optCSEfindDsc(unsigned index);
5560 void optUnmarkCSE(GenTreePtr tree);
5562 // user defined callback data for the tree walk function optCSE_MaskHelper()
5563 struct optCSE_MaskData
5565 EXPSET_TP CSE_defMask;
5566 EXPSET_TP CSE_useMask;
5569 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5570 static fgWalkPreFn optCSE_MaskHelper;
5572 // This function walks all the node for an given tree
5573 // and return the mask of CSE definitions and uses for the tree
5575 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5577 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5578 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5579 bool optCSE_canSwap(GenTree* tree);
5581 static fgWalkPostFn optPropagateNonCSE;
5582 static fgWalkPreFn optHasNonCSEChild;
5584 static fgWalkPreFn optUnmarkCSEs;
5586 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5587 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5589 void optCleanupCSEs();
5592 void optEnsureClearCSEInfo();
5595 #endif // FEATURE_ANYCSE
5597 #if FEATURE_VALNUM_CSE
5598 /**************************************************************************
5599 * Value Number based CSEs
5600 *************************************************************************/
5603 void optOptimizeValnumCSEs();
5606 void optValnumCSE_Init();
5607 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5608 unsigned optValnumCSE_Locate();
5609 void optValnumCSE_InitDataFlow();
5610 void optValnumCSE_DataFlow();
5611 void optValnumCSE_Availablity();
5612 void optValnumCSE_Heuristic();
5613 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5615 #endif // FEATURE_VALNUM_CSE
5618 bool optDoCSE; // True when we have found a duplicate CSE tree
5619 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5620 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5621 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5622 unsigned optCSEstart; // The first local variable number that is a CSE
5623 unsigned optCSEcount; // The total count of CSE's introduced.
5624 unsigned optCSEweight; // The weight of the current block when we are
5625 // scanning for CSE expressions
5627 bool optIsCSEcandidate(GenTreePtr tree);
5629 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5631 bool lclNumIsTrueCSE(unsigned lclNum) const
5633 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5636 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5638 bool lclNumIsCSE(unsigned lclNum) const
5640 return lvaTable[lclNum].lvIsCSE;
5644 bool optConfigDisableCSE();
5645 bool optConfigDisableCSE2();
5647 void optOptimizeCSEs();
5649 #endif // FEATURE_ANYCSE
5657 unsigned ivaVar; // Variable we are interested in, or -1
5658 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5659 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5660 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5661 callInterf ivaMaskCall; // What kind of calls are there?
5664 static callInterf optCallInterf(GenTreeCall* call);
5667 // VN based copy propagation.
5668 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5669 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5670 LclNumToGenTreePtrStack;
5672 // Kill set to track variables with intervening definitions.
5673 VARSET_TP optCopyPropKillSet;
5675 // Copy propagation functions.
5676 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5677 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5678 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5679 bool optIsSsaLocal(GenTreePtr tree);
5680 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5681 void optVnCopyProp();
5683 /**************************************************************************
5684 * Early value propagation
5685 *************************************************************************/
5691 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5695 static unsigned GetHashCode(SSAName ssaNm)
5697 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5700 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5702 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5706 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5707 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5708 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5709 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5710 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5711 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5713 bool doesMethodHaveFatPointer()
5715 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5718 void setMethodHasFatPointer()
5720 optMethodFlags |= OMF_HAS_FATPOINTER;
5723 void clearMethodHasFatPointer()
5725 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5728 void addFatPointerCandidate(GenTreeCall* call)
5730 setMethodHasFatPointer();
5731 call->SetFatPointerCandidate();
5734 unsigned optMethodFlags;
5736 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5737 // No throughput diff was found with backward walk bound between 3-8.
5738 static const int optEarlyPropRecurBound = 5;
5740 enum class optPropKind
5748 bool gtIsVtableRef(GenTreePtr tree);
5749 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5750 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5751 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5752 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5753 bool optEarlyPropRewriteTree(GenTreePtr tree);
5754 bool optDoEarlyPropForBlock(BasicBlock* block);
5755 bool optDoEarlyPropForFunc();
5756 void optEarlyProp();
5757 void optFoldNullCheck(GenTreePtr tree);
5758 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5761 /**************************************************************************
5762 * Value/Assertion propagation
5763 *************************************************************************/
5765 // Data structures for assertion prop
5766 BitVecTraits* apTraits;
5769 enum optAssertionKind
5786 O1K_CONSTANT_LOOP_BND,
5807 optAssertionKind assertionKind;
5810 unsigned lclNum; // assigned to or property of this local var number
5818 struct AssertionDscOp1
5820 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5827 struct AssertionDscOp2
5829 optOp2Kind kind; // a const or copy assignment
5833 ssize_t iconVal; // integer
5834 unsigned iconFlags; // gtFlags
5836 struct Range // integer subrange
5850 bool IsCheckedBoundArithBound()
5852 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
5854 bool IsCheckedBoundBound()
5856 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
5858 bool IsConstantBound()
5860 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5861 op1.kind == O1K_CONSTANT_LOOP_BND);
5863 bool IsBoundsCheckNoThrow()
5865 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5868 bool IsCopyAssertion()
5870 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5873 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5875 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5876 a1->op2.kind == a2->op2.kind;
5879 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5881 if (kind == OAK_EQUAL)
5883 return kind2 == OAK_NOT_EQUAL;
5885 else if (kind == OAK_NOT_EQUAL)
5887 return kind2 == OAK_EQUAL;
5892 static ssize_t GetLowerBoundForIntegralType(var_types type)
5912 static ssize_t GetUpperBoundForIntegralType(var_types type)
5936 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5938 if (op1.kind != that->op1.kind)
5942 else if (op1.kind == O1K_ARR_BND)
5945 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
5949 return ((vnBased && (op1.vn == that->op1.vn)) ||
5950 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5954 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5956 if (op2.kind != that->op2.kind)
5962 case O2K_IND_CNS_INT:
5964 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5966 case O2K_CONST_LONG:
5967 return (op2.lconVal == that->op2.lconVal);
5969 case O2K_CONST_DOUBLE:
5970 // exact match because of positive and negative zero.
5971 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5973 case O2K_LCLVAR_COPY:
5975 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5976 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5979 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5982 // we will return false
5986 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5992 bool Complementary(AssertionDsc* that, bool vnBased)
5994 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5995 HasSameOp2(that, vnBased);
5998 bool Equals(AssertionDsc* that, bool vnBased)
6000 if (assertionKind != that->assertionKind)
6004 else if (assertionKind == OAK_NO_THROW)
6006 assert(op2.kind == O2K_INVALID);
6007 return HasSameOp1(that, vnBased);
6011 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
6017 static fgWalkPreFn optAddCopiesCallback;
6018 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
6019 unsigned optAddCopyLclNum;
6020 GenTreePtr optAddCopyAsgnNode;
6022 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6023 bool optAssertionPropagated; // set to true if we modified the trees
6024 bool optAssertionPropagatedCurrentStmt;
6026 GenTreePtr optAssertionPropCurrentTree;
6028 AssertionIndex* optComplementaryAssertionMap;
6029 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6030 // using the value of a local var) for each local var
6031 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6032 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6033 AssertionIndex optMaxAssertionCount;
6036 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6037 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6038 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
6039 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
6040 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
6041 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
6043 AssertionIndex GetAssertionCount()
6045 return optAssertionCount;
6047 ASSERT_TP* bbJtrueAssertionOut;
6048 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
6049 ValueNumToAssertsMap;
6050 ValueNumToAssertsMap* optValueNumToAsserts;
6052 // Assertion prop helpers.
6053 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6054 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6055 void optAssertionInit(bool isLocalProp);
6056 void optAssertionTraitsInit(AssertionIndex assertionCount);
6057 #if LOCAL_ASSERTION_PROP
6058 void optAssertionReset(AssertionIndex limit);
6059 void optAssertionRemove(AssertionIndex index);
6062 // Assertion prop data flow functions.
6063 void optAssertionPropMain();
6064 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
6065 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
6066 ASSERT_TP* optInitAssertionDataflowFlags();
6067 ASSERT_TP* optComputeAssertionGen();
6069 // Assertion Gen functions.
6070 void optAssertionGen(GenTreePtr tree);
6071 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
6072 AssertionInfo optCreateJTrueBoundsAssertion(GenTreePtr tree);
6073 AssertionInfo optAssertionGenJtrue(GenTreePtr tree);
6074 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
6075 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6076 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6078 // Assertion creation functions.
6079 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
6080 AssertionIndex optCreateAssertion(GenTreePtr op1,
6082 optAssertionKind assertionKind,
6083 AssertionDsc* assertion);
6084 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
6086 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6087 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6088 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6090 void optPrintVnAssertionMapping();
6092 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6094 // Used for respective assertion propagations.
6095 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
6096 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
6097 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
6098 bool optAssertionIsNonNull(GenTreePtr op,
6099 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6101 // Used for Relop propagation.
6102 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
6103 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6104 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6106 // Assertion prop for lcl var functions.
6107 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6108 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
6110 GenTreePtr stmt DEBUGARG(AssertionIndex index));
6111 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
6112 const GenTreePtr tree,
6113 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
6114 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
6116 // Assertion propagation functions.
6117 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6118 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6119 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6120 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6121 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6122 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6123 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6124 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6125 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6126 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
6127 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
6128 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, const GenTreePtr stmt);
6130 // Implied assertion functions.
6131 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6132 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6133 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6134 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6137 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6138 void optDebugCheckAssertion(AssertionDsc* assertion);
6139 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6141 void optAddCopies();
6142 #endif // ASSERTION_PROP
6144 /**************************************************************************
6146 *************************************************************************/
6149 struct LoopCloneVisitorInfo
6151 LoopCloneContext* context;
6154 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
6155 : context(context), loopNum(loopNum), stmt(nullptr)
6160 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6161 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6162 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
6163 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6164 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6165 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6166 void optObtainLoopCloningOpts(LoopCloneContext* context);
6167 bool optIsLoopClonable(unsigned loopInd);
6169 bool optCanCloneLoops();
6172 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6174 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6175 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6176 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6177 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6181 void optInsertLoopCloningStress(BasicBlock* head);
6183 #if COUNT_RANGECHECKS
6184 static unsigned optRangeChkRmv;
6185 static unsigned optRangeChkAll;
6194 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6199 RngChkDsc* rcdNextInBucket; // used by the hash table
6201 unsigned short rcdHashValue; // to make matching faster
6202 unsigned short rcdIndex; // 0..optRngChkCount-1
6204 GenTreePtr rcdTree; // the array index tree
6207 unsigned optRngChkCount;
6208 static const size_t optRngChkHashSize;
6210 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6211 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6213 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6216 bool optLoopsMarked;
6219 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6220 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6224 XX Does the register allocation and puts the remaining lclVars on the stack XX
6226 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6227 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6231 #ifndef LEGACY_BACKEND
6236 #else // LEGACY_BACKEND
6241 #endif // LEGACY_BACKEND
6243 #ifdef LEGACY_BACKEND
6245 void raAssignVars(); // register allocation
6246 #endif // LEGACY_BACKEND
6248 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6250 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6252 void raMarkStkVars();
6255 // Some things are used by both LSRA and regpredict allocators.
6257 FrameType rpFrameType;
6258 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6260 #ifdef LEGACY_BACKEND
6261 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6263 #endif // LEGACY_BACKEND
6265 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6267 #if FEATURE_FP_REGALLOC
6268 enum enumConfigRegisterFP
6270 CONFIG_REGISTER_FP_NONE = 0x0,
6271 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6272 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6273 CONFIG_REGISTER_FP_FULL = 0x3,
6275 enumConfigRegisterFP raConfigRegisterFP();
6276 #endif // FEATURE_FP_REGALLOC
6279 regMaskTP raConfigRestrictMaskFP();
6282 #ifndef LEGACY_BACKEND
6283 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6284 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6285 #else // LEGACY_BACKEND
6286 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6287 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6288 bool raNewBlocks; // True is we added killing blocks for FPU registers
6289 unsigned rpPasses; // Number of passes made by the register predicter
6290 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6291 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6292 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6293 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6294 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6295 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6296 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6297 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6298 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6299 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6300 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6301 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6303 bool rpRegAllocDone; // Set to true after we have completed register allocation
6305 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6307 void raSetupArgMasks(RegState* r);
6309 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6311 void raDumpVarIntf(); // Dump the variable to variable interference graph
6312 void raDumpRegIntf(); // Dump the variable to register interference graph
6314 void raAdjustVarIntf();
6316 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6318 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6320 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6321 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6323 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6325 static fgWalkPreFn rpMarkRegIntf;
6327 regMaskTP rpPredictAddressMode(
6328 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6330 void rpPredictRefAssign(unsigned lclNum);
6332 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6334 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6336 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6338 void rpPredictRegUse(); // Entry point
6340 unsigned raPredictTreeRegUse(GenTreePtr tree);
6341 unsigned raPredictListRegUse(GenTreePtr list);
6343 void raSetRegVarOrder(var_types regType,
6344 regNumber* customVarOrder,
6345 unsigned* customVarOrderSize,
6347 regMaskTP avoidReg);
6349 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6350 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6351 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6352 void raAddToStkPredict(unsigned val)
6354 unsigned newStkPredict = rpStkPredict + val;
6355 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6356 rpStkPredict = UINT_MAX - 1;
6358 rpStkPredict = newStkPredict;
6362 #if !FEATURE_FP_REGALLOC
6363 void raDispFPlifeInfo();
6367 regMaskTP genReturnRegForTree(GenTreePtr tree);
6368 #endif // LEGACY_BACKEND
6370 /* raIsVarargsStackArg is called by raMaskStkVars and by
6371 lvaSortByRefCount. It identifies the special case
6372 where a varargs function has a parameter passed on the
6373 stack, other than the special varargs handle. Such parameters
6374 require special treatment, because they cannot be tracked
6375 by the GC (their offsets in the stack are not known
6379 bool raIsVarargsStackArg(unsigned lclNum)
6383 LclVarDsc* varDsc = &lvaTable[lclNum];
6385 assert(varDsc->lvIsParam);
6387 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6389 #else // _TARGET_X86_
6393 #endif // _TARGET_X86_
6396 #ifdef LEGACY_BACKEND
6397 // Records the current prediction, if it's better than any previous recorded prediction.
6398 void rpRecordPrediction();
6399 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6400 void rpUseRecordedPredictionIfBetter();
6402 // Data members used in the methods above.
6403 unsigned rpBestRecordedStkPredict;
6404 struct VarRegPrediction
6406 bool m_isEnregistered;
6407 regNumberSmall m_regNum;
6408 regNumberSmall m_otherReg;
6410 VarRegPrediction* rpBestRecordedPrediction;
6411 #endif // LEGACY_BACKEND
6414 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6419 XX Get to the class and method info from the Execution Engine given XX
6420 XX tokens for the class and method XX
6422 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6423 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6427 /* These are the different addressing modes used to access a local var.
6428 * The JIT has to report the location of the locals back to the EE
6429 * for debugging purposes.
6435 VLT_REG_BYREF, // this type is currently only used for value types on X64
6438 VLT_STK_BYREF, // this type is currently only used for value types on X64
6452 siVarLocType vlType;
6455 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6457 // VLT_REG_BYREF -- the specified register contains the address of the variable
6465 // VLT_STK -- Any 32 bit value which is on the stack
6466 // eg. [ESP+0x20], or [EBP-0x28]
6467 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6468 // eg. mov EAX, [ESP+0x20]; [EAX]
6472 regNumber vlsBaseReg;
6473 NATIVE_OFFSET vlsOffset;
6476 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6485 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6486 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6494 regNumber vlrssBaseReg;
6495 NATIVE_OFFSET vlrssOffset;
6499 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6500 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6506 regNumber vlsrsBaseReg;
6507 NATIVE_OFFSET vlsrsOffset;
6513 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6514 // eg 2 DWords at [ESP+0x10]
6518 regNumber vls2BaseReg;
6519 NATIVE_OFFSET vls2Offset;
6522 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6523 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6530 // VLT_FIXED_VA -- fixed argument of a varargs function.
6531 // The argument location depends on the size of the variable
6532 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6533 // location of the first arg. This argument can then be accessed
6534 // relative to the position of the first arg
6538 unsigned vlfvOffset;
6545 void* rpValue; // pointer to the in-process
6546 // location of the value.
6552 bool vlIsInReg(regNumber reg);
6553 bool vlIsOnStk(regNumber reg, signed offset);
6556 /*************************************************************************/
6561 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6562 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6563 CORINFO_CALLINFO_FLAGS flags,
6564 CORINFO_CALL_INFO* pResult);
6565 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6567 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6568 CORINFO_ACCESS_FLAGS flags,
6569 CORINFO_FIELD_INFO* pResult);
6573 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6575 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6577 bool IsSuperPMIException(unsigned code)
6579 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6581 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6582 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6583 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6584 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6585 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6586 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6587 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6588 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6592 case EXCEPTIONCODE_DebugBreakorAV:
6593 case EXCEPTIONCODE_MC:
6594 case EXCEPTIONCODE_LWM:
6595 case EXCEPTIONCODE_SASM:
6596 case EXCEPTIONCODE_SSYM:
6597 case EXCEPTIONCODE_CALLUTILS:
6598 case EXCEPTIONCODE_TYPEUTILS:
6599 case EXCEPTIONCODE_ASSERT:
6606 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6607 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6609 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6610 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6613 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6614 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6615 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6617 // VOM info, method sigs
6619 void eeGetSig(unsigned sigTok,
6620 CORINFO_MODULE_HANDLE scope,
6621 CORINFO_CONTEXT_HANDLE context,
6622 CORINFO_SIG_INFO* retSig);
6624 void eeGetCallSiteSig(unsigned sigTok,
6625 CORINFO_MODULE_HANDLE scope,
6626 CORINFO_CONTEXT_HANDLE context,
6627 CORINFO_SIG_INFO* retSig);
6629 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6631 // Method entry-points, instrs
6633 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6635 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6637 CORINFO_EE_INFO eeInfo;
6638 bool eeInfoInitialized;
6640 CORINFO_EE_INFO* eeGetEEInfo();
6642 // Gets the offset of a SDArray's first element
6643 unsigned eeGetArrayDataOffset(var_types type);
6644 // Gets the offset of a MDArray's first element
6645 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6647 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6649 // Returns the page size for the target machine as reported by the EE.
6650 inline size_t eeGetPageSize()
6652 return eeGetEEInfo()->osPageSize;
6655 // Returns the frame size at which we will generate a loop to probe the stack.
6656 inline size_t getVeryLargeFrameSize()
6659 // The looping probe code is 40 bytes, whereas the straight-line probing for
6660 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6661 // or greater, to generate smaller code.
6662 return 2 * eeGetPageSize();
6664 return 3 * eeGetPageSize();
6668 //------------------------------------------------------------------------
6669 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6671 // It represents Abi and target specific registers for the parameter.
6673 class VirtualStubParamInfo
6676 VirtualStubParamInfo(bool isCoreRTABI)
6678 #if defined(_TARGET_X86_)
6681 #elif defined(_TARGET_AMD64_)
6692 #elif defined(_TARGET_ARM_)
6703 #elif defined(_TARGET_ARM64_)
6707 #error Unsupported or unset target architecture
6710 #ifdef LEGACY_BACKEND
6711 #if defined(_TARGET_X86_)
6712 predict = PREDICT_REG_EAX;
6713 #elif defined(_TARGET_ARM_)
6714 predict = PREDICT_REG_R4;
6716 #error Unsupported or unset target architecture
6718 #endif // LEGACY_BACKEND
6721 regNumber GetReg() const
6726 _regMask_enum GetRegMask() const
6731 #ifdef LEGACY_BACKEND
6732 rpPredictReg GetPredict() const
6740 _regMask_enum regMask;
6742 #ifdef LEGACY_BACKEND
6743 rpPredictReg predict;
6747 VirtualStubParamInfo* virtualStubParamInfo;
6749 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6751 return eeGetEEInfo()->targetAbi == abi;
6754 inline bool generateCFIUnwindCodes()
6756 #ifdef UNIX_AMD64_ABI
6757 return IsTargetAbi(CORINFO_CORERT_ABI);
6765 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6767 // Debugging support - Line number info
6769 void eeGetStmtOffsets();
6771 unsigned eeBoundariesCount;
6773 struct boundariesDsc
6775 UNATIVE_OFFSET nativeIP;
6777 unsigned sourceReason;
6778 } * eeBoundaries; // Boundaries to report to EE
6779 void eeSetLIcount(unsigned count);
6780 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6784 static void eeDispILOffs(IL_OFFSET offs);
6785 static void eeDispLineInfo(const boundariesDsc* line);
6786 void eeDispLineInfos();
6789 // Debugging support - Local var info
6793 unsigned eeVarsCount;
6795 struct VarResultInfo
6797 UNATIVE_OFFSET startOffset;
6798 UNATIVE_OFFSET endOffset;
6802 void eeSetLVcount(unsigned count);
6803 void eeSetLVinfo(unsigned which,
6804 UNATIVE_OFFSET startOffs,
6805 UNATIVE_OFFSET length,
6810 const siVarLoc& loc);
6814 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6815 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6818 // ICorJitInfo wrappers
6820 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6822 void eeAllocUnwindInfo(BYTE* pHotCode,
6828 CorJitFuncKind funcKind);
6830 void eeSetEHcount(unsigned cEH);
6832 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6834 WORD eeGetRelocTypeHint(void* target);
6836 // ICorStaticInfo wrapper functions
6838 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6840 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6842 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6845 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6846 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6847 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6848 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6850 template <typename ParamType>
6851 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6853 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6856 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6858 // Utility functions
6860 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6863 const wchar_t* eeGetCPString(size_t stringHandle);
6866 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6868 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6869 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6871 static fgWalkPreFn CountSharedStaticHelper;
6872 static bool IsSharedStaticHelper(GenTreePtr tree);
6873 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6874 static bool IsGcSafePoint(GenTreePtr tree);
6876 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6877 // returns true/false if 'field' is a Jit Data offset
6878 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6879 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6880 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6882 /*****************************************************************************/
6887 enum TEMP_USAGE_TYPE
6893 static var_types tmpNormalizeType(var_types type);
6894 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6895 void tmpRlsTemp(TempDsc* temp);
6896 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6899 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6900 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6904 bool tmpAllFree() const;
6907 #ifndef LEGACY_BACKEND
6908 void tmpPreAllocateTemps(var_types type, unsigned count);
6909 #endif // !LEGACY_BACKEND
6912 #ifdef LEGACY_BACKEND
6913 unsigned tmpIntSpillMax; // number of int-sized spill temps
6914 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6915 #endif // LEGACY_BACKEND
6917 unsigned tmpCount; // Number of temps
6918 unsigned tmpSize; // Size of all the temps
6921 // Used by RegSet::rsSpillChk()
6922 unsigned tmpGetCount; // Temps which haven't been released yet
6925 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6927 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6928 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6931 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6932 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6936 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6937 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6941 CodeGenInterface* codeGen;
6943 // The following holds information about instr offsets in terms of generated code.
6947 IPmappingDsc* ipmdNext; // next line# record
6948 IL_OFFSETX ipmdILoffsx; // the instr offset
6949 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6950 bool ipmdIsLabel; // Can this code be a branch label?
6953 // Record the instr offset mapping to the generated code
6955 IPmappingDsc* genIPmappingList;
6956 IPmappingDsc* genIPmappingLast;
6958 // Managed RetVal - A side hash table meant to record the mapping from a
6959 // GT_CALL node to its IL offset. This info is used to emit sequence points
6960 // that can be used by debugger to determine the native offset at which the
6961 // managed RetVal will be available.
6963 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6964 // favor of a side table for two reasons: 1) We need IL offset for only those
6965 // GT_CALL nodes (created during importation) that correspond to an IL call and
6966 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6967 // structure and IL offset is needed only when generating debuggable code. Therefore
6968 // it is desirable to avoid memory size penalty in retail scenarios.
6969 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6970 CallSiteILOffsetTable;
6971 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6973 unsigned genReturnLocal; // Local number for the return value when applicable.
6974 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6976 // The following properties are part of CodeGenContext. Getters are provided here for
6977 // convenience and backward compatibility, but the properties can only be set by invoking
6978 // the setter on CodeGenContext directly.
6980 __declspec(property(get = getEmitter)) emitter* genEmitter;
6981 emitter* getEmitter()
6983 return codeGen->getEmitter();
6986 const bool isFramePointerUsed()
6988 return codeGen->isFramePointerUsed();
6991 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6992 bool getInterruptible()
6994 return codeGen->genInterruptible;
6996 void setInterruptible(bool value)
6998 codeGen->setInterruptible(value);
7002 const bool genDoubleAlign()
7004 return codeGen->doDoubleAlign();
7006 DWORD getCanDoubleAlign();
7007 bool shouldDoubleAlign(unsigned refCntStk,
7009 unsigned refCntWtdReg,
7010 unsigned refCntStkParam,
7011 unsigned refCntWtdStkDbl);
7012 #endif // DOUBLE_ALIGN
7014 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
7015 bool getFullPtrRegMap()
7017 return codeGen->genFullPtrRegMap;
7019 void setFullPtrRegMap(bool value)
7021 codeGen->setFullPtrRegMap(value);
7024 // Things that MAY belong either in CodeGen or CodeGenContext
7026 #if FEATURE_EH_FUNCLETS
7027 FuncInfoDsc* compFuncInfos;
7028 unsigned short compCurrFuncIdx;
7029 unsigned short compFuncInfoCount;
7031 unsigned short compFuncCount()
7033 assert(fgFuncletsCreated);
7034 return compFuncInfoCount;
7037 #else // !FEATURE_EH_FUNCLETS
7039 // This is a no-op when there are no funclets!
7040 void genUpdateCurrentFunclet(BasicBlock* block)
7045 FuncInfoDsc compFuncInfoRoot;
7047 static const unsigned compCurrFuncIdx = 0;
7049 unsigned short compFuncCount()
7054 #endif // !FEATURE_EH_FUNCLETS
7056 FuncInfoDsc* funCurrentFunc();
7057 void funSetCurrentFunc(unsigned funcIdx);
7058 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7059 unsigned int funGetFuncIdx(BasicBlock* block);
7063 VARSET_TP compCurLife; // current live variables
7064 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
7066 template <bool ForCodeGen>
7067 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
7069 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
7071 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
7074 template <bool ForCodeGen>
7075 void compUpdateLife(GenTreePtr tree);
7077 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
7078 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
7079 // use. (Can be more than one var in the case of dependently promoted struct vars.)
7080 template <bool ForCodeGen>
7081 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
7083 template <bool ForCodeGen>
7084 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7086 // Gets a register mask that represent the kill set for a helper call since
7087 // not all JIT Helper calls follow the standard ABI on the target architecture.
7088 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7090 // Gets a register mask that represent the kill set for a NoGC helper call.
7091 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7094 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7095 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7096 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7097 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7098 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7099 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7100 #endif // _TARGET_ARM_
7102 // 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
7104 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
7106 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7107 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7108 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
7109 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7110 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7111 // for the tracked var indices of the field vars, as in a live var set).
7112 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7114 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7116 if (m_promotedStructDeathVars == nullptr)
7118 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7120 return m_promotedStructDeathVars;
7124 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7125 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7129 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7130 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7133 #if !defined(__GNUC__)
7134 #pragma region Unwind information
7139 // Infrastructure functions: start/stop/reserve/emit.
7142 void unwindBegProlog();
7143 void unwindEndProlog();
7144 void unwindBegEpilog();
7145 void unwindEndEpilog();
7146 void unwindReserve();
7147 void unwindEmit(void* pHotCode, void* pColdCode);
7150 // Specific unwind information functions: called by code generation to indicate a particular
7151 // prolog or epilog unwindable instruction has been generated.
7154 void unwindPush(regNumber reg);
7155 void unwindAllocStack(unsigned size);
7156 void unwindSetFrameReg(regNumber reg, unsigned offset);
7157 void unwindSaveReg(regNumber reg, unsigned offset);
7159 #if defined(_TARGET_ARM_)
7160 void unwindPushMaskInt(regMaskTP mask);
7161 void unwindPushMaskFloat(regMaskTP mask);
7162 void unwindPopMaskInt(regMaskTP mask);
7163 void unwindPopMaskFloat(regMaskTP mask);
7164 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7165 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7166 // called via unwindPadding().
7167 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7168 // instruction and the current location.
7169 #endif // _TARGET_ARM_
7171 #if defined(_TARGET_ARM64_)
7173 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7174 // instruction and the current location.
7175 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7176 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7177 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7178 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7179 void unwindSaveNext(); // unwind code: save_next
7180 void unwindReturn(regNumber reg); // ret lr
7181 #endif // defined(_TARGET_ARM64_)
7184 // Private "helper" functions for the unwind implementation.
7188 #if FEATURE_EH_FUNCLETS
7189 void unwindGetFuncLocations(FuncInfoDsc* func,
7190 bool getHotSectionData,
7191 /* OUT */ emitLocation** ppStartLoc,
7192 /* OUT */ emitLocation** ppEndLoc);
7193 #endif // FEATURE_EH_FUNCLETS
7195 void unwindReserveFunc(FuncInfoDsc* func);
7196 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7198 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7200 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7201 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7203 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7205 #if defined(_TARGET_AMD64_)
7207 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7209 void unwindBegPrologWindows();
7210 void unwindPushWindows(regNumber reg);
7211 void unwindAllocStackWindows(unsigned size);
7212 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7213 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7215 #ifdef UNIX_AMD64_ABI
7216 void unwindBegPrologCFI();
7217 void unwindPushCFI(regNumber reg);
7218 void unwindAllocStackCFI(unsigned size);
7219 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7220 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7221 int mapRegNumToDwarfReg(regNumber reg);
7222 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7223 #endif // UNIX_AMD64_ABI
7224 #elif defined(_TARGET_ARM_)
7226 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7227 void unwindPushPopMaskFloat(regMaskTP mask);
7228 void unwindSplit(FuncInfoDsc* func);
7230 #endif // _TARGET_ARM_
7232 #if !defined(__GNUC__)
7233 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7237 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7238 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7242 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7243 XX that contains the distinguished, well-known SIMD type definitions). XX
7245 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7246 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7249 // Get highest available instruction set for floating point codegen
7250 InstructionSet getFloatingPointInstructionSet()
7252 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7255 return InstructionSet_AVX;
7260 return InstructionSet_SSE3_4;
7264 assert(canUseSSE2());
7265 return InstructionSet_SSE2;
7267 assert(!"getFPInstructionSet() is not implemented for target arch");
7269 return InstructionSet_NONE;
7273 // Get highest available instruction set for SIMD codegen
7274 InstructionSet getSIMDInstructionSet()
7276 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7277 return getFloatingPointInstructionSet();
7279 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7281 return InstructionSet_NONE;
7287 // Should we support SIMD intrinsics?
7290 // Have we identified any SIMD types?
7291 // This is currently used by struct promotion to avoid getting type information for a struct
7292 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7294 bool _usesSIMDTypes;
7295 bool usesSIMDTypes()
7297 return _usesSIMDTypes;
7299 void setUsesSIMDTypes(bool value)
7301 _usesSIMDTypes = value;
7304 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7305 // that require indexed access to the individual fields of the vector, which is not well supported
7306 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7307 unsigned lvaSIMDInitTempVarNum;
7310 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7311 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7312 CORINFO_CLASS_HANDLE SIMDIntHandle;
7313 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7314 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7315 CORINFO_CLASS_HANDLE SIMDShortHandle;
7316 CORINFO_CLASS_HANDLE SIMDByteHandle;
7317 CORINFO_CLASS_HANDLE SIMDLongHandle;
7318 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7319 CORINFO_CLASS_HANDLE SIMDULongHandle;
7320 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7321 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7322 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7323 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7325 // Get the handle for a SIMD type.
7326 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7328 if (simdBaseType == TYP_FLOAT)
7333 return SIMDVector2Handle;
7335 return SIMDVector3Handle;
7337 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7339 return SIMDVector4Handle;
7348 assert(simdType == getSIMDVectorType());
7349 switch (simdBaseType)
7352 return SIMDFloatHandle;
7354 return SIMDDoubleHandle;
7356 return SIMDIntHandle;
7358 return SIMDUShortHandle;
7360 return SIMDUShortHandle;
7362 return SIMDUByteHandle;
7364 return SIMDShortHandle;
7366 return SIMDByteHandle;
7368 return SIMDLongHandle;
7370 return SIMDUIntHandle;
7372 return SIMDULongHandle;
7374 assert(!"Didn't find a class handle for simdType");
7376 return NO_CLASS_HANDLE;
7380 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7381 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7382 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7384 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7385 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7386 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7387 bool isSIMDTypeLocal(GenTree* tree)
7389 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7392 // Returns true if the type of the tree is a byref of TYP_SIMD
7393 bool isAddrOfSIMDType(GenTree* tree)
7395 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7397 switch (tree->OperGet())
7400 return varTypeIsSIMD(tree->gtGetOp1());
7402 case GT_LCL_VAR_ADDR:
7403 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7406 return isSIMDTypeLocal(tree);
7413 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7415 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7416 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7417 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7420 // Returns base type of a TYP_SIMD local.
7421 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7422 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7424 if (isSIMDTypeLocal(tree))
7426 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7432 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7434 return info.compCompHnd->isInSIMDModule(clsHnd);
7437 bool isSIMDClass(typeInfo* pTypeInfo)
7439 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7442 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7443 // if it is not a SIMD type or is an unsupported base type.
7444 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7446 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7448 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7451 // Get SIMD Intrinsic info given the method handle.
7452 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7453 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7454 CORINFO_METHOD_HANDLE methodHnd,
7455 CORINFO_SIG_INFO* sig,
7458 var_types* baseType,
7459 unsigned* sizeBytes);
7461 // Pops and returns GenTree node from importers type stack.
7462 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7463 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7465 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7466 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7468 // Creates a GT_SIMD tree for Select operation
7469 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7471 unsigned simdVectorSize,
7476 // Creates a GT_SIMD tree for Min/Max operation
7477 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7478 CORINFO_CLASS_HANDLE typeHnd,
7480 unsigned simdVectorSize,
7484 // Transforms operands and returns the SIMD intrinsic to be applied on
7485 // transformed operands to obtain given relop result.
7486 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7487 CORINFO_CLASS_HANDLE typeHnd,
7488 unsigned simdVectorSize,
7489 var_types* baseType,
7493 // Creates a GT_SIMD tree for Abs intrinsic.
7494 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7496 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7497 // Transforms operands and returns the SIMD intrinsic to be applied on
7498 // transformed operands to obtain == comparison result.
7499 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7500 unsigned simdVectorSize,
7504 // Transforms operands and returns the SIMD intrinsic to be applied on
7505 // transformed operands to obtain > comparison result.
7506 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7507 unsigned simdVectorSize,
7511 // Transforms operands and returns the SIMD intrinsic to be applied on
7512 // transformed operands to obtain >= comparison result.
7513 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7514 unsigned simdVectorSize,
7518 // Transforms operands and returns the SIMD intrinsic to be applied on
7519 // transformed operands to obtain >= comparison result in case of int32
7520 // and small int base type vectors.
7521 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7522 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7523 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7525 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7526 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7527 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7528 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7529 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7531 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7532 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7533 GenTreePtr newobjThis,
7534 CORINFO_CLASS_HANDLE clsHnd,
7535 CORINFO_METHOD_HANDLE method,
7536 CORINFO_SIG_INFO* sig,
7539 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7541 // Whether SIMD vector occupies part of SIMD register.
7542 // SSE2: vector2f/3f are considered sub register SIMD types.
7543 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7544 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7546 unsigned sizeBytes = 0;
7547 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7548 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7551 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7553 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7556 // Get the type for the hardware SIMD vector.
7557 // This is the maximum SIMD type supported for this target.
7558 var_types getSIMDVectorType()
7560 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7567 assert(canUseSSE2());
7571 assert(!"getSIMDVectorType() unimplemented on target arch");
7576 // Get the size of the SIMD type in bytes
7577 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7579 unsigned sizeBytes = 0;
7580 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7584 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7585 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7587 // Get the the number of elements of basetype of SIMD vector given by its type handle
7588 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7590 // Get preferred alignment of SIMD type.
7591 int getSIMDTypeAlignment(var_types simdType);
7593 // Get the number of bytes in a SIMD Vector for the current compilation.
7594 unsigned getSIMDVectorRegisterByteLength()
7596 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7599 return YMM_REGSIZE_BYTES;
7603 assert(canUseSSE2());
7604 return XMM_REGSIZE_BYTES;
7607 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7612 // The minimum and maximum possible number of bytes in a SIMD vector.
7613 unsigned int maxSIMDStructBytes()
7615 return getSIMDVectorRegisterByteLength();
7617 unsigned int minSIMDStructBytes()
7619 return emitTypeSize(TYP_SIMD8);
7622 #ifdef FEATURE_AVX_SUPPORT
7623 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7624 static const unsigned maxPossibleSIMDStructBytes = 32;
7625 #else // !FEATURE_AVX_SUPPORT
7626 static const unsigned maxPossibleSIMDStructBytes = 16;
7627 #endif // !FEATURE_AVX_SUPPORT
7629 // Returns the codegen type for a given SIMD size.
7630 var_types getSIMDTypeForSize(unsigned size)
7632 var_types simdType = TYP_UNDEF;
7635 simdType = TYP_SIMD8;
7637 else if (size == 12)
7639 simdType = TYP_SIMD12;
7641 else if (size == 16)
7643 simdType = TYP_SIMD16;
7645 #ifdef FEATURE_AVX_SUPPORT
7646 else if (size == 32)
7648 simdType = TYP_SIMD32;
7650 #endif // FEATURE_AVX_SUPPORT
7653 noway_assert(!"Unexpected size for SIMD type");
7658 unsigned getSIMDInitTempVarNum()
7660 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7662 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7663 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7665 return lvaSIMDInitTempVarNum;
7668 #endif // FEATURE_SIMD
7671 //------------------------------------------------------------------------
7672 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7674 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7675 // candidate for enregistration.
7677 unsigned largestEnregisterableStructSize()
7680 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7681 if (vectorRegSize > TARGET_POINTER_SIZE)
7683 return vectorRegSize;
7686 #endif // FEATURE_SIMD
7688 return TARGET_POINTER_SIZE;
7693 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7694 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7695 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7697 // Is this var is of type simd struct?
7698 bool lclVarIsSIMDType(unsigned varNum)
7700 LclVarDsc* varDsc = lvaTable + varNum;
7701 return varDsc->lvIsSIMDType();
7704 // Is this Local node a SIMD local?
7705 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7707 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7710 // Returns true if the TYP_SIMD locals on stack are aligned at their
7711 // preferred byte boundary specified by getSIMDTypeAlignment().
7713 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7714 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7715 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7716 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7717 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7718 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7719 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7722 bool isSIMDTypeLocalAligned(unsigned varNum)
7724 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7725 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7728 int off = lvaFrameAddress(varNum, &ebpBased);
7729 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7730 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7731 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7734 #endif // FEATURE_SIMD
7739 // Whether SSE2 is available
7740 bool canUseSSE2() const
7742 #ifdef _TARGET_XARCH_
7743 return opts.compCanUseSSE2;
7749 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7750 bool CanUseSSE3_4() const
7752 #ifdef _TARGET_XARCH_
7753 return opts.compCanUseSSE3_4;
7759 bool canUseAVX() const
7761 #ifdef FEATURE_AVX_SUPPORT
7762 return opts.compCanUseAVX;
7769 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7770 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7774 XX Generic info about the compilation and the method being compiled. XX
7775 XX It is responsible for driving the other phases. XX
7776 XX It is also responsible for all the memory management. XX
7778 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7779 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7783 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7785 InlineResult* compInlineResult; // The result of importing the inlinee method.
7787 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7788 bool compJmpOpUsed; // Does the method do a JMP
7789 bool compLongUsed; // Does the method use TYP_LONG
7790 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7791 bool compTailCallUsed; // Does the method do a tailcall
7792 bool compLocallocUsed; // Does the method use localloc.
7793 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7794 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7795 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7797 // NOTE: These values are only reliable after
7798 // the importing is completely finished.
7800 #ifdef LEGACY_BACKEND
7801 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7802 // we can iterate over these efficiently.
7805 #if CPU_USES_BLOCK_MOVE
7806 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7810 // State information - which phases have completed?
7811 // These are kept together for easy discoverability
7813 bool bRangeAllowStress;
7814 bool compCodeGenDone;
7815 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7816 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7817 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7818 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7821 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7822 bool fgLocalVarLivenessChanged;
7824 bool compStackProbePrologDone;
7826 #ifndef LEGACY_BACKEND
7828 #endif // !LEGACY_BACKEND
7829 bool compRationalIRForm;
7831 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7833 bool compGeneratingProlog;
7834 bool compGeneratingEpilog;
7835 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7836 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7837 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7838 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7839 bool getNeedsGSSecurityCookie() const
7841 return compNeedsGSSecurityCookie;
7843 void setNeedsGSSecurityCookie()
7845 compNeedsGSSecurityCookie = true;
7848 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7849 // frame layout calculations, this is the level we are currently
7852 //---------------------------- JITing options -----------------------------
7865 JitFlags* jitFlags; // all flags passed from the EE
7866 unsigned compFlags; // method attributes
7868 codeOptimize compCodeOpt; // what type of code optimizations
7872 #ifdef _TARGET_XARCH_
7873 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7874 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7876 #ifdef FEATURE_AVX_SUPPORT
7877 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7878 #endif // FEATURE_AVX_SUPPORT
7879 #endif // _TARGET_XARCH_
7881 // optimize maximally and/or favor speed over size?
7883 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7884 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7885 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7886 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7887 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7889 // Maximun number of locals before turning off the inlining
7890 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7893 unsigned instrCount;
7894 unsigned lvRefCount;
7895 bool compMinOptsIsSet;
7897 bool compMinOptsIsUsed;
7899 inline bool MinOpts()
7901 assert(compMinOptsIsSet);
7902 compMinOptsIsUsed = true;
7905 inline bool IsMinOptsSet()
7907 return compMinOptsIsSet;
7910 inline bool MinOpts()
7914 inline bool IsMinOptsSet()
7916 return compMinOptsIsSet;
7919 inline void SetMinOpts(bool val)
7921 assert(!compMinOptsIsUsed);
7922 assert(!compMinOptsIsSet || (compMinOpts == val));
7924 compMinOptsIsSet = true;
7927 // true if the CLFLG_* for an optimization is set.
7928 inline bool OptEnabled(unsigned optFlag)
7930 return !!(compFlags & optFlag);
7933 #ifdef FEATURE_READYTORUN_COMPILER
7934 inline bool IsReadyToRun()
7936 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7939 inline bool IsReadyToRun()
7945 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7946 // PInvoke transitions inline (e.g. when targeting CoreRT).
7947 inline bool ShouldUsePInvokeHelpers()
7949 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7952 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7954 inline bool IsReversePInvoke()
7956 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7959 // true if we must generate code compatible with JIT32 quirks
7960 inline bool IsJit32Compat()
7962 #if defined(_TARGET_X86_)
7963 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7969 // true if we must generate code compatible with Jit64 quirks
7970 inline bool IsJit64Compat()
7972 #if defined(_TARGET_AMD64_)
7973 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7974 #elif !defined(FEATURE_CORECLR)
7981 bool compScopeInfo; // Generate the LocalVar info ?
7982 bool compDbgCode; // Generate debugger-friendly code?
7983 bool compDbgInfo; // Gather debugging info?
7986 #ifdef PROFILING_SUPPORTED
7987 bool compNoPInvokeInlineCB;
7989 static const bool compNoPInvokeInlineCB;
7993 bool compGcChecks; // Check arguments and return values to ensure they are sane
7994 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7995 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7999 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
8000 // to be allocated on the stack.
8001 // It will be set to true in the following cases:
8002 // 1. When the method being compiled has a declarative security
8003 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
8004 // This is also the case when we inject a prolog and epilog in the method.
8006 // 2. When the method being compiled has imperative security (i.e. the method
8007 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
8009 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
8011 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
8012 // which gets reported as a GC root to stackwalker.
8013 // (See also ICodeManager::GetAddrOfSecurityObject.)
8018 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
8019 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
8023 #ifdef UNIX_AMD64_ABI
8024 // This flag is indicating if there is a need to align the frame.
8025 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8026 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8027 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8028 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8029 // there are calls and making sure the frame alignment logic is executed.
8030 bool compNeedToAlignFrame;
8031 #endif // UNIX_AMD64_ABI
8033 bool compProcedureSplitting; // Separate cold code from hot code
8035 bool genFPorder; // Preserve FP order (operations are non-commutative)
8036 bool genFPopt; // Can we do frame-pointer-omission optimization?
8037 bool altJit; // True if we are an altjit and are compiling this method
8040 bool optRepeat; // Repeat optimizer phases k times
8044 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8045 bool dspCode; // Display native code generated
8046 bool dspEHTable; // Display the EH table reported to the VM
8047 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8048 bool dspEmit; // Display emitter output
8049 bool dspLines; // Display source-code lines intermixed with native code output
8050 bool dmpHex; // Display raw bytes in hex of native code output
8051 bool varNames; // Display variables names in native code output
8052 bool disAsm; // Display native code as it is generated
8053 bool disAsmSpilled; // Display native code when any register spilling occurs
8054 bool disDiffable; // Makes the Disassembly code 'diff-able'
8055 bool disAsm2; // Display native code after it is generated using external disassembler
8056 bool dspOrder; // Display names of each of the methods that we ngen/jit
8057 bool dspUnwind; // Display the unwind info output
8058 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8059 bool compLongAddress; // Force using large pseudo instructions for long address
8060 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8061 bool dspGCtbls; // Display the GC tables
8065 bool doLateDisasm; // Run the late disassembler
8066 #endif // LATE_DISASM
8068 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8069 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8070 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8071 static const bool dspGCtbls = true;
8074 // We need stack probes to guarantee that we won't trigger a stack overflow
8075 // when calling unmanaged code until they get a chance to set up a frame, because
8076 // the EE will have no idea where it is.
8078 // We will only be doing this currently for hosted environments. Unfortunately
8079 // we need to take care of stubs, so potentially, we will have to do the probes
8080 // for any call. We have a plan for not needing for stubs though
8081 bool compNeedStackProbes;
8083 #ifdef PROFILING_SUPPORTED
8084 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8085 // This option helps make the JIT behave as if it is running under a profiler.
8086 bool compJitELTHookEnabled;
8087 #endif // PROFILING_SUPPORTED
8089 #if FEATURE_TAILCALL_OPT
8090 // Whether opportunistic or implicit tail call optimization is enabled.
8091 bool compTailCallOpt;
8092 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8093 bool compTailCallLoopOpt;
8097 static const bool compUseSoftFP = true;
8098 #else // !ARM_SOFTFP
8099 static const bool compUseSoftFP = false;
8102 GCPollType compGCPollType;
8106 static bool s_pAltJitExcludeAssembliesListInitialized;
8107 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8112 template <typename T>
8115 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8118 template <typename T>
8121 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8124 static int dspTreeID(GenTree* tree)
8126 return tree->gtTreeID;
8128 static void printTreeID(GenTree* tree)
8130 if (tree == nullptr)
8136 printf("[%06d]", dspTreeID(tree));
8143 #define STRESS_MODES \
8147 /* "Variations" stress areas which we try to mix up with each other. */ \
8148 /* These should not be exhaustively used as they might */ \
8149 /* hide/trivialize other areas */ \
8152 STRESS_MODE(DBL_ALN) \
8153 STRESS_MODE(LCL_FLDS) \
8154 STRESS_MODE(UNROLL_LOOPS) \
8155 STRESS_MODE(MAKE_CSE) \
8156 STRESS_MODE(LEGACY_INLINE) \
8157 STRESS_MODE(CLONE_EXPR) \
8158 STRESS_MODE(USE_FCOMI) \
8159 STRESS_MODE(USE_CMOV) \
8161 STRESS_MODE(BB_PROFILE) \
8162 STRESS_MODE(OPT_BOOLS_GC) \
8163 STRESS_MODE(REMORPH_TREES) \
8164 STRESS_MODE(64RSLT_MUL) \
8165 STRESS_MODE(DO_WHILE_LOOPS) \
8166 STRESS_MODE(MIN_OPTS) \
8167 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8168 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8169 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8170 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8171 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8172 STRESS_MODE(NULL_OBJECT_CHECK) \
8173 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8174 STRESS_MODE(RANDOM_INLINE) \
8175 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8176 STRESS_MODE(GENERIC_VARN) \
8178 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8180 STRESS_MODE(COUNT_VARN) \
8182 /* "Check" stress areas that can be exhaustively used if we */ \
8183 /* dont care about performance at all */ \
8185 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8186 STRESS_MODE(CHK_FLOW_UPDATE) \
8187 STRESS_MODE(EMITTER) \
8188 STRESS_MODE(CHK_REIMPORT) \
8189 STRESS_MODE(FLATFP) \
8190 STRESS_MODE(GENERIC_CHECK) \
8195 #define STRESS_MODE(mode) STRESS_##mode,
8202 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8203 BYTE compActiveStressModes[STRESS_COUNT];
8206 #define MAX_STRESS_WEIGHT 100
8208 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8212 bool compInlineStress()
8214 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8217 bool compRandomInlineStress()
8219 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8224 bool compTailCallStress()
8227 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8233 codeOptimize compCodeOpt()
8236 // Switching between size & speed has measurable throughput impact
8237 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8238 // DEBUG, but should generate identical code between CHK & RET builds,
8239 // so that's not acceptable.
8240 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8241 // Investigate the cause of the throughput regression.
8243 return opts.compCodeOpt;
8245 return BLENDED_CODE;
8249 //--------------------- Info about the procedure --------------------------
8253 COMP_HANDLE compCompHnd;
8254 CORINFO_MODULE_HANDLE compScopeHnd;
8255 CORINFO_CLASS_HANDLE compClassHnd;
8256 CORINFO_METHOD_HANDLE compMethodHnd;
8257 CORINFO_METHOD_INFO* compMethodInfo;
8259 BOOL hasCircularClassConstraints;
8260 BOOL hasCircularMethodConstraints;
8262 #if defined(DEBUG) || defined(LATE_DISASM)
8263 const char* compMethodName;
8264 const char* compClassName;
8265 const char* compFullName;
8266 #endif // defined(DEBUG) || defined(LATE_DISASM)
8268 #if defined(DEBUG) || defined(INLINE_DATA)
8269 // Method hash is logcally const, but computed
8271 mutable unsigned compMethodHashPrivate;
8272 unsigned compMethodHash() const;
8273 #endif // defined(DEBUG) || defined(INLINE_DATA)
8275 #ifdef PSEUDORANDOM_NOP_INSERTION
8276 // things for pseudorandom nop insertion
8277 unsigned compChecksum;
8281 // The following holds the FLG_xxxx flags for the method we're compiling.
8284 // The following holds the class attributes for the method we're compiling.
8285 unsigned compClassAttr;
8287 const BYTE* compCode;
8288 IL_OFFSET compILCodeSize; // The IL code size
8289 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8290 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8291 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8292 // (2) the code is hot/cold split, and we issued less code than we expected
8293 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8295 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8296 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8297 bool compIsContextful : 1; // contextful method
8298 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8299 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8300 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8301 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8302 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8304 var_types compRetType; // Return type of the method as declared in IL
8305 var_types compRetNativeType; // Normalized return type as per target arch ABI
8306 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8307 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8309 #if FEATURE_FASTTAILCALL
8310 unsigned compArgRegCount; // Number of incoming integer argument registers used for incoming arguments
8311 unsigned compFloatArgRegCount; // Number of incoming floating argument registers used for incoming arguments
8312 size_t compArgStackSize; // Incoming argument stack size in bytes
8313 #endif // FEATURE_FASTTAILCALL
8315 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8316 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8317 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8318 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8319 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8320 unsigned compMaxStack;
8321 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8322 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8324 unsigned compCallUnmanaged; // count of unmanaged calls
8325 unsigned compLvFrameListRoot; // lclNum for the Frame root
8326 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8327 // You should generally use compHndBBtabCount instead: it is the
8328 // current number of EH clauses (after additions like synchronized
8329 // methods and funclets, and removals like unreachable code deletion).
8331 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8332 // and the VM expects that, or the JIT is a "self-host" compiler
8333 // (e.g., x86 hosted targeting x86) and the VM expects that.
8335 /* The following holds IL scope information about local variables.
8338 unsigned compVarScopesCount;
8339 VarScopeDsc* compVarScopes;
8341 /* The following holds information about instr offsets for
8342 * which we need to report IP-mappings
8345 IL_OFFSET* compStmtOffsets; // sorted
8346 unsigned compStmtOffsetsCount;
8347 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8349 #define CPU_X86 0x0100 // The generic X86 CPU
8350 #define CPU_X86_PENTIUM_4 0x0110
8352 #define CPU_X64 0x0200 // The generic x64 CPU
8353 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8354 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8356 #define CPU_ARM 0x0300 // The generic ARM CPU
8358 unsigned genCPU; // What CPU are we running on
8361 // Returns true if the method being compiled returns a non-void and non-struct value.
8362 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8363 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8364 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8365 // Methods returning such structs are considered to return non-struct return value and
8366 // this method returns true in that case.
8367 bool compMethodReturnsNativeScalarType()
8369 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8372 // Returns true if the method being compiled returns RetBuf addr as its return value
8373 bool compMethodReturnsRetBufAddr()
8375 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8376 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8378 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8379 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8380 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8381 // methods with hidden RetBufArg.
8383 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8384 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8385 // returning the address of RetBuf.
8387 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8388 // to be returned in RAX.
8389 CLANG_FORMAT_COMMENT_ANCHOR;
8391 #ifdef _TARGET_AMD64_
8392 return (info.compRetBuffArg != BAD_VAR_NUM);
8393 #else // !_TARGET_AMD64_
8394 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8395 #endif // !_TARGET_AMD64_
8398 // Returns true if the method returns a value in more than one return register
8399 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8400 // TODO-ARM64: Does this apply for ARM64 too?
8401 bool compMethodReturnsMultiRegRetType()
8403 #if FEATURE_MULTIREG_RET
8404 #if defined(_TARGET_X86_)
8405 // On x86 only 64-bit longs are returned in multiple registers
8406 return varTypeIsLong(info.compRetNativeType);
8407 #else // targets: X64-UNIX, ARM64 or ARM32
8408 // On all other targets that support multireg return values:
8409 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8410 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8411 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8412 #endif // TARGET_XXX
8414 #else // not FEATURE_MULTIREG_RET
8416 // For this architecture there are no multireg returns
8419 #endif // FEATURE_MULTIREG_RET
8422 #if FEATURE_MULTIREG_ARGS
8423 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8424 // return the gcPtr layout for the pointers sized fields
8425 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8426 #endif // FEATURE_MULTIREG_ARGS
8428 // Returns true if the method being compiled returns a value
8429 bool compMethodHasRetVal()
8431 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8432 compMethodReturnsMultiRegRetType();
8437 void compDispLocalVars();
8441 //-------------------------- Global Compiler Data ------------------------------------
8444 static unsigned s_compMethodsCount; // to produce unique label names
8445 unsigned compGenTreeID;
8446 unsigned compBasicBlockID;
8449 BasicBlock* compCurBB; // the current basic block in process
8450 GenTreePtr compCurStmt; // the current statement in process
8452 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8455 // The following is used to create the 'method JIT info' block.
8456 size_t compInfoBlkSize;
8457 BYTE* compInfoBlkAddr;
8459 EHblkDsc* compHndBBtab; // array of EH data
8460 unsigned compHndBBtabCount; // element count of used elements in EH data array
8461 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8463 #if defined(_TARGET_X86_)
8465 //-------------------------------------------------------------------------
8466 // Tracking of region covered by the monitor in synchronized methods
8467 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8468 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8470 #endif // !_TARGET_X86_
8472 Phases previousCompletedPhase; // the most recently completed phase
8474 //-------------------------------------------------------------------------
8475 // The following keeps track of how many bytes of local frame space we've
8476 // grabbed so far in the current function, and how many argument bytes we
8477 // need to pop when we return.
8480 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8482 // Count of callee-saved regs we pushed in the prolog.
8483 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8484 // In case of Amd64 this doesn't include float regs saved on stack.
8485 unsigned compCalleeRegsPushed;
8487 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8488 // Mask of callee saved float regs on stack.
8489 regMaskTP compCalleeFPRegsSavedMask;
8491 #ifdef _TARGET_AMD64_
8492 // Quirk for VS debug-launch scenario to work:
8493 // Bytes of padding between save-reg area and locals.
8494 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8495 unsigned compVSQuirkStackPaddingNeeded;
8496 bool compQuirkForPPPflag;
8499 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8501 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8502 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8503 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8505 //-------------------------------------------------------------------------
8507 static void compStartup(); // One-time initialization
8508 static void compShutdown(); // One-time finalization
8510 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8513 static void compDisplayStaticSizes(FILE* fout);
8515 //------------ Some utility functions --------------
8517 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8518 void** ppIndirection); /* OUT */
8520 // Several JIT/EE interface functions return a CorInfoType, and also return a
8521 // class handle as an out parameter if the type is a value class. Returns the
8522 // size of the type these describe.
8523 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8526 // Components used by the compiler may write unit test suites, and
8527 // have them run within this method. They will be run only once per process, and only
8528 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8529 // These should fail by asserting.
8530 void compDoComponentUnitTestsOnce();
8533 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8534 CORINFO_MODULE_HANDLE classPtr,
8535 COMP_HANDLE compHnd,
8536 CORINFO_METHOD_INFO* methodInfo,
8537 void** methodCodePtr,
8538 ULONG* methodCodeSize,
8539 JitFlags* compileFlags);
8540 void compCompileFinish();
8541 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8542 COMP_HANDLE compHnd,
8543 CORINFO_METHOD_INFO* methodInfo,
8544 void** methodCodePtr,
8545 ULONG* methodCodeSize,
8546 JitFlags* compileFlags,
8547 CorInfoInstantiationVerification instVerInfo);
8549 ArenaAllocator* compGetAllocator();
8551 #if MEASURE_MEM_ALLOC
8553 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8557 unsigned allocCnt; // # of allocs
8558 UINT64 allocSz; // total size of those alloc.
8559 UINT64 allocSzMax; // Maximum single allocation.
8560 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8561 UINT64 nraTotalSizeAlloc;
8562 UINT64 nraTotalSizeUsed;
8564 static const char* s_CompMemKindNames[]; // Names of the kinds.
8566 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8568 for (int i = 0; i < CMK_Count; i++)
8570 allocSzByKind[i] = 0;
8573 MemStats(const MemStats& ms)
8574 : allocCnt(ms.allocCnt)
8575 , allocSz(ms.allocSz)
8576 , allocSzMax(ms.allocSzMax)
8577 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8578 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8580 for (int i = 0; i < CMK_Count; i++)
8582 allocSzByKind[i] = ms.allocSzByKind[i];
8586 // Until we have ubiquitous constructors.
8589 this->MemStats::MemStats();
8592 void AddAlloc(size_t sz, CompMemKind cmk)
8596 if (sz > allocSzMax)
8600 allocSzByKind[cmk] += sz;
8603 void Print(FILE* f); // Print these stats to f.
8604 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8606 MemStats genMemStats;
8608 struct AggregateMemStats : public MemStats
8612 AggregateMemStats() : MemStats(), nMethods(0)
8616 void Add(const MemStats& ms)
8619 allocCnt += ms.allocCnt;
8620 allocSz += ms.allocSz;
8621 allocSzMax = max(allocSzMax, ms.allocSzMax);
8622 for (int i = 0; i < CMK_Count; i++)
8624 allocSzByKind[i] += ms.allocSzByKind[i];
8626 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8627 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8630 void Print(FILE* f); // Print these stats to jitstdout.
8633 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8634 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8635 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8637 #endif // MEASURE_MEM_ALLOC
8639 #if LOOP_HOIST_STATS
8640 unsigned m_loopsConsidered;
8641 bool m_curLoopHasHoistedExpression;
8642 unsigned m_loopsWithHoistedExpressions;
8643 unsigned m_totalHoistedExpressions;
8645 void AddLoopHoistStats();
8646 void PrintPerMethodLoopHoistStats();
8648 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8649 static unsigned s_loopsConsidered;
8650 static unsigned s_loopsWithHoistedExpressions;
8651 static unsigned s_totalHoistedExpressions;
8653 static void PrintAggregateLoopHoistStats(FILE* f);
8654 #endif // LOOP_HOIST_STATS
8656 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8657 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8658 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8659 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8660 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8661 void compFreeMem(void*);
8663 bool compIsForImportOnly();
8664 bool compIsForInlining();
8665 bool compDonotInline();
8668 const char* compLocalVarName(unsigned varNum, unsigned offs);
8669 VarName compVarName(regNumber reg, bool isFloatReg = false);
8670 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8671 const char* compRegPairName(regPairNo regPair);
8672 const char* compRegNameForSize(regNumber reg, size_t size);
8673 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8674 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8675 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8678 //-------------------------------------------------------------------------
8680 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8682 struct VarScopeMapInfo
8684 VarScopeListNode* head;
8685 VarScopeListNode* tail;
8686 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8688 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8695 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8696 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8698 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8699 VarNumToScopeDscMap;
8701 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8702 VarNumToScopeDscMap* compVarScopeMap;
8704 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8706 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8708 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8710 void compInitVarScopeMap();
8712 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8713 // enter scope, sorted by instr offset
8714 unsigned compNextEnterScope;
8716 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8717 // go out of scope, sorted by instr offset
8718 unsigned compNextExitScope;
8720 void compInitScopeLists();
8722 void compResetScopeLists();
8724 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8726 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8728 void compProcessScopesUntil(unsigned offset,
8730 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8731 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8734 void compDispScopeLists();
8737 bool compIsProfilerHookNeeded();
8739 //-------------------------------------------------------------------------
8740 /* Statistical Data Gathering */
8742 void compJitStats(); // call this function and enable
8743 // various ifdef's below for statistical data
8746 void compCallArgStats();
8747 static void compDispCallArgStats(FILE* fout);
8750 //-------------------------------------------------------------------------
8757 ArenaAllocator* compAllocator;
8760 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8761 // suitable for use by utilcode collection types.
8762 IAllocator* compAsIAllocator;
8764 #if MEASURE_MEM_ALLOC
8765 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8766 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8767 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8769 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8771 #endif // MEASURE_MEM_ALLOC
8773 void compFunctionTraceStart();
8774 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8777 size_t compMaxUncheckedOffsetForNullObject;
8779 void compInitOptions(JitFlags* compileFlags);
8781 void compSetProcessor();
8782 void compInitDebuggingInfo();
8783 void compSetOptimizationLevel();
8784 #ifdef _TARGET_ARMARCH_
8785 bool compRsvdRegCheck(FrameLayoutState curState);
8787 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8789 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8790 void ResetOptAnnotations();
8792 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8793 void RecomputeLoopInfo();
8795 #ifdef PROFILING_SUPPORTED
8796 // Data required for generating profiler Enter/Leave/TailCall hooks
8798 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8799 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8800 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8803 #ifdef _TARGET_AMD64_
8804 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8807 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8808 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8810 IAllocator* getAllocator()
8812 return compAsIAllocator;
8815 #if MEASURE_MEM_ALLOC
8816 IAllocator* getAllocatorBitset()
8818 return compAsIAllocatorBitset;
8820 IAllocator* getAllocatorGC()
8822 return compAsIAllocatorGC;
8824 IAllocator* getAllocatorLoopHoist()
8826 return compAsIAllocatorLoopHoist;
8828 #else // !MEASURE_MEM_ALLOC
8829 IAllocator* getAllocatorBitset()
8831 return compAsIAllocator;
8833 IAllocator* getAllocatorGC()
8835 return compAsIAllocator;
8837 IAllocator* getAllocatorLoopHoist()
8839 return compAsIAllocator;
8841 #endif // !MEASURE_MEM_ALLOC
8844 IAllocator* getAllocatorDebugOnly()
8846 #if MEASURE_MEM_ALLOC
8847 return compAsIAllocatorDebugOnly;
8848 #else // !MEASURE_MEM_ALLOC
8849 return compAsIAllocator;
8850 #endif // !MEASURE_MEM_ALLOC
8855 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8856 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8860 XX Checks for type compatibility and merges types XX
8862 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8863 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8867 // Set to TRUE if verification cannot be skipped for this method
8868 // If we detect unverifiable code, we will lazily check
8869 // canSkipMethodVerification() to see if verification is REALLY needed.
8870 BOOL tiVerificationNeeded;
8872 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8873 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8874 BOOL tiIsVerifiableCode;
8876 // Set to TRUE if runtime callout is needed for this method
8877 BOOL tiRuntimeCalloutNeeded;
8879 // Set to TRUE if security prolog/epilog callout is needed for this method
8880 // Note: This flag is different than compNeedSecurityCheck.
8881 // compNeedSecurityCheck means whether or not a security object needs
8882 // to be allocated on the stack, which is currently true for EnC as well.
8883 // tiSecurityCalloutNeeded means whether or not security callouts need
8884 // to be inserted in the jitted code.
8885 BOOL tiSecurityCalloutNeeded;
8887 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8888 // This support is necessary to suport attributes that are not described in
8889 // for example, signatures. For example, the permanent home byref (byref that
8890 // points to the gc heap), isn't a property of method signatures, therefore,
8891 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8892 // but when deciding if we need to reimport a block, we need to take these
8894 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8896 // Returns TRUE if child is equal to or a subtype of parent.
8897 // normalisedForStack indicates that both types are normalised for the stack
8898 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8900 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8901 // *pDest is modified to represent the merged type. Sets "*changed" to true
8902 // if this changes "*pDest".
8903 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8905 // Set pDest from the primitive value type.
8906 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8908 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8911 // <BUGNUM> VSW 471305
8912 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8913 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8914 // We use a "short" as we need to push/pop this scope.
8916 short compRegSetCheckLevel;
8920 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8921 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8923 XX IL verification stuff XX
8926 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8927 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8931 // The following is used to track liveness of local variables, initialization
8932 // of valueclass constructors, and type safe use of IL instructions.
8934 // dynamic state info needed for verification
8935 EntryState verCurrentState;
8937 // this ptr of object type .ctors are considered intited only after
8938 // the base class ctor is called, or an alternate ctor is called.
8939 // An uninited this ptr can be used to access fields, but cannot
8940 // be used to call a member function.
8941 BOOL verTrackObjCtorInitState;
8943 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8945 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8946 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8947 void verInitCurrentState();
8948 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8950 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8951 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8952 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8954 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8955 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8956 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8957 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8958 typeInfo verMakeTypeInfo(CorInfoType ciType,
8959 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8960 BOOL verIsSDArray(typeInfo ti);
8961 typeInfo verGetArrayElemType(typeInfo ti);
8963 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8964 BOOL verNeedsVerification();
8965 BOOL verIsByRefLike(const typeInfo& ti);
8966 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8968 // generic type variables range over types that satisfy IsBoxable
8969 BOOL verIsBoxable(const typeInfo& ti);
8971 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8972 DEBUGARG(unsigned line));
8973 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8974 DEBUGARG(unsigned line));
8975 bool verCheckTailCallConstraint(OPCODE opcode,
8976 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8977 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8978 // on a type parameter?
8979 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8980 // return false to the caller.
8981 // If false, it will throw.
8983 bool verIsBoxedValueType(typeInfo ti);
8985 void verVerifyCall(OPCODE opcode,
8986 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8987 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8989 bool readonlyCall, // is this a "readonly." call?
8990 const BYTE* delegateCreateStart,
8991 const BYTE* codeAddr,
8992 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8994 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8996 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8997 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8998 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8999 const CORINFO_FIELD_INFO& fieldInfo,
9000 const typeInfo* tiThis,
9002 BOOL allowPlainStructAsThis = FALSE);
9003 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
9004 void verVerifyThisPtrInitialised();
9005 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
9007 // Register allocator
9008 void raInitStackFP();
9009 void raEnregisterVarsPrePassStackFP();
9010 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
9011 void raEnregisterVarsPostPassStackFP();
9012 void raGenerateFPRefCounts();
9013 void raEnregisterVarsStackFP();
9014 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
9016 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
9017 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
9019 // returns true if enregistering v1 would save more mem accesses than v2
9020 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
9023 void raDumpHeightsStackFP();
9024 void raDumpVariableRegIntfFloat();
9027 #if FEATURE_STACK_FP_X87
9029 // Currently, we use FP transition blocks in only 2 situations:
9031 // -conditional jump on longs where FP stack differs with target: it's not strictly
9032 // necessary, but its low frequency and the code would get complicated if we try to
9033 // inline the FP stack adjustment, as we have a lot of special casing going on to try
9034 // minimize the way we generate the jump code.
9035 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
9036 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
9038 // However, transition blocks have 2 problems
9040 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
9041 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
9042 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
9043 // in the right place without preordering them), this causes us to have to generate the transition
9044 // blocks in the cold area if we want procedure splitting.
9047 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
9048 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
9049 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
9050 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
9051 // a big change in the exception.
9053 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
9054 // optimizations. For these 2 cases:
9056 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
9057 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
9058 // a switch statement.
9060 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
9061 // current procedure splitting and exception code have.
9062 bool compMayHaveTransitionBlocks;
9064 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
9066 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
9068 unsigned raCntStkStackFP;
9069 unsigned raCntWtdStkDblStackFP;
9070 unsigned raCntStkParamDblStackFP;
9072 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
9073 // TODO: Do we want to put this in LclVarDsc?
9074 unsigned raPayloadStackFP[lclMAX_TRACKED];
9075 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9077 // Useful for debugging
9078 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
9080 #endif // FEATURE_STACK_FP_X87
9083 // One line log function. Default level is 0. Increasing it gives you
9084 // more log information
9086 // levels are currently unused: #define JITDUMP(level,...) ();
9087 void JitLogEE(unsigned level, const char* fmt, ...);
9089 bool compDebugBreak;
9091 bool compJitHaltMethod();
9096 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9097 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9099 XX GS Security checks for unsafe buffers XX
9101 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9102 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9105 struct ShadowParamVarInfo
9107 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9108 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9110 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9112 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
9113 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9114 // slots and update all trees to refer to shadow slots is done immediately after
9115 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9116 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9117 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9118 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9119 // creating a shadow slot even though this routine returns true.
9121 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9122 // required. There are two cases under which a reg arg could potentially be used from its
9124 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9125 // b) LSRA spills it
9127 // Possible solution to address case (a)
9128 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9129 // in this routine. Note that live out of exception handler is something we may not be
9130 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9131 // Therefore, for methods with exception handling and need GS cookie check we might have
9132 // to take conservative approach.
9134 // Possible solution to address case (b)
9135 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9136 // create a new spill temp if the method needs GS cookie check.
9137 return varDsc->lvIsParam;
9138 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
9139 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9146 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9151 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9152 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9153 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9155 void gsGSChecksInitCookie(); // Grabs cookie variable
9156 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9157 bool gsFindVulnerableParams(); // Shadow param analysis code
9158 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9160 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9161 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9163 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9164 // This can be overwritten by setting complus_JITInlineSize env variable.
9166 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9169 #ifdef FEATURE_JIT_METHOD_PERF
9170 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9171 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9173 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9174 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9176 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9178 #if MEASURE_CLRAPI_CALLS
9179 // Thin wrappers that call into JitTimer (if present).
9180 inline void CLRApiCallEnter(unsigned apix);
9181 inline void CLRApiCallLeave(unsigned apix);
9184 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9185 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9190 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9191 // These variables are associated with maintaining SQM data about compile time.
9192 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9193 // in the current compilation.
9194 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9195 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9196 // the inlining phase in the current compilation.
9197 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9199 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9200 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9201 // type-loading and class initialization).
9202 void RecordStateAtEndOfInlining();
9203 // Assumes being called at the end of compilation. Update the SQM state.
9204 void RecordStateAtEndOfCompilation();
9206 #ifdef FEATURE_CLRSQM
9207 // Does anything SQM related necessary at process shutdown time.
9208 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9209 #endif // FEATURE_CLRSQM
9212 #if FUNC_INFO_LOGGING
9213 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9214 // filename to write it to.
9215 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9216 #endif // FUNC_INFO_LOGGING
9218 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9220 // Is the compilation in a full trust context?
9221 bool compIsFullTrust();
9224 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9225 #endif // MEASURE_NOWAY
9227 #ifndef FEATURE_TRACELOGGING
9228 // Should we actually fire the noway assert body and the exception handler?
9229 bool compShouldThrowOnNoway();
9230 #else // FEATURE_TRACELOGGING
9231 // Should we actually fire the noway assert body and the exception handler?
9232 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9234 // Telemetry instance to use per method compilation.
9235 JitTelemetry compJitTelemetry;
9237 // Get common parameters that have to be logged with most telemetry data.
9238 void compGetTelemetryDefaults(const char** assemblyName,
9239 const char** scopeName,
9240 const char** methodName,
9241 unsigned* methodHash);
9242 #endif // !FEATURE_TRACELOGGING
9246 NodeToTestDataMap* m_nodeTestData;
9248 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9249 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9250 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9251 // Current kept in this.
9253 NodeToTestDataMap* GetNodeTestData()
9255 Compiler* compRoot = impInlineRoot();
9256 if (compRoot->m_nodeTestData == nullptr)
9258 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9260 return compRoot->m_nodeTestData;
9263 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9265 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9266 // currently occur in the AST graph.
9267 NodeToIntMap* FindReachableNodesInNodeTestData();
9269 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9270 // test data, associate that data with "to".
9271 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9273 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9274 // have annotations, attach similar annotations to the corresponding nodes in "to".
9275 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9277 // These are the methods that test that the various conditions implied by the
9278 // test attributes are satisfied.
9279 void JitTestCheckSSA(); // SSA builder tests.
9280 void JitTestCheckVN(); // Value numbering tests.
9283 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9285 FieldSeqStore* m_fieldSeqStore;
9287 FieldSeqStore* GetFieldSeqStore()
9289 Compiler* compRoot = impInlineRoot();
9290 if (compRoot->m_fieldSeqStore == nullptr)
9292 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9293 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9294 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9296 return compRoot->m_fieldSeqStore;
9299 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9301 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9302 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9303 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9304 // attach the field sequence directly to the address node.
9305 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9307 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9309 // Don't need to worry about inlining here
9310 if (m_zeroOffsetFieldMap == nullptr)
9312 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9314 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9315 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9317 return m_zeroOffsetFieldMap;
9320 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9321 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9322 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9323 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9324 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9325 // record the the field sequence using the ZeroOffsetFieldMap described above.
9327 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9328 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9329 // CoreRT. Such case is handled same as the default case.
9330 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9332 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9334 NodeToArrayInfoMap* m_arrayInfoMap;
9336 NodeToArrayInfoMap* GetArrayInfoMap()
9338 Compiler* compRoot = impInlineRoot();
9339 if (compRoot->m_arrayInfoMap == nullptr)
9341 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9342 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9343 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9345 return compRoot->m_arrayInfoMap;
9348 //-----------------------------------------------------------------------------------------------------------------
9349 // Compiler::TryGetArrayInfo:
9350 // Given an indirection node, checks to see whether or not that indirection represents an array access, and
9351 // if so returns information about the array.
9354 // indir - The `GT_IND` node.
9355 // arrayInfo (out) - Information about the accessed array if this function returns true. Undefined otherwise.
9358 // True if the `GT_IND` node represents an array access; false otherwise.
9359 inline bool TryGetArrayInfo(GenTreeIndir* indir, ArrayInfo* arrayInfo)
9361 if ((indir->gtFlags & GTF_IND_ARR_INDEX) == 0)
9366 if (indir->gtOp1->OperIs(GT_INDEX_ADDR))
9368 GenTreeIndexAddr* const indexAddr = indir->gtOp1->AsIndexAddr();
9369 *arrayInfo = ArrayInfo(indexAddr->gtElemType, indexAddr->gtElemSize, indexAddr->gtElemOffset,
9370 indexAddr->gtStructElemClass);
9374 bool found = GetArrayInfoMap()->Lookup(indir, arrayInfo);
9379 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9381 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9382 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9383 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9384 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9386 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9388 // Use the same map for GCHeap and ByrefExposed when their states match.
9389 memoryKind = ByrefExposed;
9392 assert(memoryKind < MemoryKindCount);
9393 Compiler* compRoot = impInlineRoot();
9394 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9396 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9397 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9398 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9400 return compRoot->m_memorySsaMap[memoryKind];
9403 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9404 CORINFO_CLASS_HANDLE m_refAnyClass;
9405 CORINFO_FIELD_HANDLE GetRefanyDataField()
9407 if (m_refAnyClass == nullptr)
9409 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9411 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9413 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9415 if (m_refAnyClass == nullptr)
9417 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9419 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9423 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9425 #if ALLVARSET_COUNTOPS
9426 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9429 static HelperCallProperties s_helperCallProperties;
9431 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9432 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9433 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9436 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9439 unsigned __int8* offset0,
9440 unsigned __int8* offset1);
9442 void GetStructTypeOffset(CORINFO_CLASS_HANDLE typeHnd,
9445 unsigned __int8* offset0,
9446 unsigned __int8* offset1);
9448 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9449 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9451 void fgMorphMultiregStructArgs(GenTreeCall* call);
9452 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9454 }; // end of class Compiler
9456 // Inline methods of CompAllocator.
9457 void* CompAllocator::Alloc(size_t sz)
9459 #if MEASURE_MEM_ALLOC
9460 return m_comp->compGetMem(sz, m_cmk);
9462 return m_comp->compGetMem(sz);
9466 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9468 #if MEASURE_MEM_ALLOC
9469 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9471 return m_comp->compGetMemArray(elems, elemSize);
9475 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9476 inline LclVarDsc::LclVarDsc(Compiler* comp)
9477 : // Initialize the ArgRegs to REG_STK.
9478 // The morph will do the right thing to change
9479 // to the right register if passed in register.
9482 #if FEATURE_MULTIREG_ARGS
9483 _lvOtherArgReg(REG_STK)
9485 #endif // FEATURE_MULTIREG_ARGS
9487 lvRefBlks(BlockSetOps::UninitVal())
9489 #endif // ASSERTION_PROP
9490 lvPerSsaData(comp->getAllocator())
9494 //---------------------------------------------------------------------------------------------------------------------
9495 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9497 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9498 // shown in parentheses):
9500 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9501 // of a misnomer, as the first entry will always be the current node.
9503 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9504 // argument before visiting the node's operands.
9506 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9507 // argument after visiting the node's operands.
9509 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9510 // `DoPreOrder` must be true if this option is true.
9512 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9513 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9514 // visited before the first).
9516 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9518 // A simple pre-order visitor might look something like the following:
9520 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9525 // DoPreOrder = true
9528 // unsigned m_count;
9530 // CountingVisitor(Compiler* compiler)
9531 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9535 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9541 // This visitor would then be used like so:
9543 // CountingVisitor countingVisitor(compiler);
9544 // countingVisitor.WalkTree(root);
9546 template <typename TVisitor>
9547 class GenTreeVisitor
9550 typedef Compiler::fgWalkResult fgWalkResult;
9554 ComputeStack = false,
9556 DoPostOrder = false,
9557 DoLclVarsOnly = false,
9558 UseExecutionOrder = false,
9561 Compiler* m_compiler;
9562 ArrayStack<GenTree*> m_ancestors;
9564 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler)
9566 assert(compiler != nullptr);
9568 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9569 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9572 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9574 return fgWalkResult::WALK_CONTINUE;
9577 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9579 return fgWalkResult::WALK_CONTINUE;
9583 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9585 assert(use != nullptr);
9587 GenTree* node = *use;
9589 if (TVisitor::ComputeStack)
9591 m_ancestors.Push(node);
9594 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9595 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9597 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9598 if (result == fgWalkResult::WALK_ABORT)
9604 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9610 switch (node->OperGet())
9615 case GT_LCL_VAR_ADDR:
9616 case GT_LCL_FLD_ADDR:
9617 if (TVisitor::DoLclVarsOnly)
9619 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9620 if (result == fgWalkResult::WALK_ABORT)
9636 case GT_MEMORYBARRIER:
9641 case GT_START_NONGC:
9643 #if !FEATURE_EH_FUNCLETS
9645 #endif // !FEATURE_EH_FUNCLETS
9647 #ifndef LEGACY_BACKEND
9649 #endif // LEGACY_BACKEND
9652 case GT_CLS_VAR_ADDR:
9656 case GT_PINVOKE_PROLOG:
9657 case GT_PINVOKE_EPILOG:
9661 // Lclvar unary operators
9662 case GT_STORE_LCL_VAR:
9663 case GT_STORE_LCL_FLD:
9664 if (TVisitor::DoLclVarsOnly)
9666 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9667 if (result == fgWalkResult::WALK_ABORT)
9674 // Standard unary operators
9702 GenTreeUnOp* const unOp = node->AsUnOp();
9703 if (unOp->gtOp1 != nullptr)
9705 result = WalkTree(&unOp->gtOp1, unOp);
9706 if (result == fgWalkResult::WALK_ABORT)
9717 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9719 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9720 if (result == fgWalkResult::WALK_ABORT)
9724 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9725 if (result == fgWalkResult::WALK_ABORT)
9729 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9730 if (result == fgWalkResult::WALK_ABORT)
9737 case GT_ARR_BOUNDS_CHECK:
9740 #endif // FEATURE_SIMD
9742 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9744 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9745 if (result == fgWalkResult::WALK_ABORT)
9749 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9750 if (result == fgWalkResult::WALK_ABORT)
9759 GenTreeField* const field = node->AsField();
9761 if (field->gtFldObj != nullptr)
9763 result = WalkTree(&field->gtFldObj, field);
9764 if (result == fgWalkResult::WALK_ABORT)
9774 GenTreeArrElem* const arrElem = node->AsArrElem();
9776 result = WalkTree(&arrElem->gtArrObj, arrElem);
9777 if (result == fgWalkResult::WALK_ABORT)
9782 const unsigned rank = arrElem->gtArrRank;
9783 for (unsigned dim = 0; dim < rank; dim++)
9785 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9786 if (result == fgWalkResult::WALK_ABORT)
9796 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9798 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9799 if (result == fgWalkResult::WALK_ABORT)
9803 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9804 if (result == fgWalkResult::WALK_ABORT)
9808 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9809 if (result == fgWalkResult::WALK_ABORT)
9818 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9820 GenTree** op1Use = &dynBlock->gtOp1;
9821 GenTree** op2Use = &dynBlock->gtDynamicSize;
9823 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9825 std::swap(op1Use, op2Use);
9828 result = WalkTree(op1Use, dynBlock);
9829 if (result == fgWalkResult::WALK_ABORT)
9833 result = WalkTree(op2Use, dynBlock);
9834 if (result == fgWalkResult::WALK_ABORT)
9841 case GT_STORE_DYN_BLK:
9843 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9845 GenTree** op1Use = &dynBlock->gtOp1;
9846 GenTree** op2Use = &dynBlock->gtOp2;
9847 GenTree** op3Use = &dynBlock->gtDynamicSize;
9849 if (TVisitor::UseExecutionOrder)
9851 if (dynBlock->IsReverseOp())
9853 std::swap(op1Use, op2Use);
9855 if (dynBlock->gtEvalSizeFirst)
9857 std::swap(op3Use, op2Use);
9858 std::swap(op2Use, op1Use);
9862 result = WalkTree(op1Use, dynBlock);
9863 if (result == fgWalkResult::WALK_ABORT)
9867 result = WalkTree(op2Use, dynBlock);
9868 if (result == fgWalkResult::WALK_ABORT)
9872 result = WalkTree(op3Use, dynBlock);
9873 if (result == fgWalkResult::WALK_ABORT)
9882 GenTreeCall* const call = node->AsCall();
9884 if (call->gtCallObjp != nullptr)
9886 result = WalkTree(&call->gtCallObjp, call);
9887 if (result == fgWalkResult::WALK_ABORT)
9893 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
9895 result = WalkTree(args->pCurrent(), call);
9896 if (result == fgWalkResult::WALK_ABORT)
9902 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
9904 result = WalkTree(args->pCurrent(), call);
9905 if (result == fgWalkResult::WALK_ABORT)
9911 if (call->gtCallType == CT_INDIRECT)
9913 if (call->gtCallCookie != nullptr)
9915 result = WalkTree(&call->gtCallCookie, call);
9916 if (result == fgWalkResult::WALK_ABORT)
9922 result = WalkTree(&call->gtCallAddr, call);
9923 if (result == fgWalkResult::WALK_ABORT)
9929 if (call->gtControlExpr != nullptr)
9931 result = WalkTree(&call->gtControlExpr, call);
9932 if (result == fgWalkResult::WALK_ABORT)
9944 assert(node->OperIsBinary());
9946 GenTreeOp* const op = node->AsOp();
9948 GenTree** op1Use = &op->gtOp1;
9949 GenTree** op2Use = &op->gtOp2;
9951 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
9953 std::swap(op1Use, op2Use);
9956 if (*op1Use != nullptr)
9958 result = WalkTree(op1Use, op);
9959 if (result == fgWalkResult::WALK_ABORT)
9965 if (*op2Use != nullptr)
9967 result = WalkTree(op2Use, op);
9968 if (result == fgWalkResult::WALK_ABORT)
9978 // Finally, visit the current node
9979 if (TVisitor::DoPostOrder)
9981 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
9984 if (TVisitor::ComputeStack)
9993 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
9994 class GenericTreeWalker final
9995 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
10000 ComputeStack = computeStack,
10001 DoPreOrder = doPreOrder,
10002 DoPostOrder = doPostOrder,
10003 DoLclVarsOnly = doLclVarsOnly,
10004 UseExecutionOrder = useExecutionOrder,
10008 Compiler::fgWalkData* m_walkData;
10011 GenericTreeWalker(Compiler::fgWalkData* walkData)
10012 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
10013 walkData->compiler)
10014 , m_walkData(walkData)
10016 assert(walkData != nullptr);
10020 walkData->parentStack = &this->m_ancestors;
10024 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
10026 m_walkData->parent = user;
10027 return m_walkData->wtprVisitorFn(use, m_walkData);
10030 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
10032 m_walkData->parent = user;
10033 return m_walkData->wtpoVisitorFn(use, m_walkData);
10037 class IncLclVarRefCountsVisitor final : public GenTreeVisitor<IncLclVarRefCountsVisitor>
10043 DoLclVarsOnly = true
10046 IncLclVarRefCountsVisitor(Compiler* compiler);
10047 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
10049 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10052 class DecLclVarRefCountsVisitor final : public GenTreeVisitor<DecLclVarRefCountsVisitor>
10058 DoLclVarsOnly = true
10061 DecLclVarRefCountsVisitor(Compiler* compiler);
10062 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user);
10064 static Compiler::fgWalkResult WalkTree(Compiler* compiler, GenTree* tree);
10068 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10069 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10071 XX Miscellaneous Compiler stuff XX
10073 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10074 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10077 // Values used to mark the types a stack slot is used for
10079 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10080 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10081 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10082 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10083 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10084 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10085 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10086 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10088 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10090 /*****************************************************************************
10092 * Variables to keep track of total code amounts.
10097 extern size_t grossVMsize;
10098 extern size_t grossNCsize;
10099 extern size_t totalNCsize;
10101 extern unsigned genMethodICnt;
10102 extern unsigned genMethodNCnt;
10103 extern size_t gcHeaderISize;
10104 extern size_t gcPtrMapISize;
10105 extern size_t gcHeaderNSize;
10106 extern size_t gcPtrMapNSize;
10108 #endif // DISPLAY_SIZES
10110 /*****************************************************************************
10112 * Variables to keep track of basic block counts (more data on 1 BB methods)
10115 #if COUNT_BASIC_BLOCKS
10116 extern Histogram bbCntTable;
10117 extern Histogram bbOneBBSizeTable;
10120 /*****************************************************************************
10122 * Used by optFindNaturalLoops to gather statistical information such as
10123 * - total number of natural loops
10124 * - number of loops with 1, 2, ... exit conditions
10125 * - number of loops that have an iterator (for like)
10126 * - number of loops that have a constant iterator
10131 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10132 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10133 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10134 extern unsigned totalLoopCount; // counts the total number of natural loops
10135 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10136 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10137 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10138 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10140 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10141 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10142 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10143 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10144 extern Histogram loopCountTable; // Histogram of loop counts
10145 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10147 #endif // COUNT_LOOPS
10149 /*****************************************************************************
10150 * variables to keep track of how many iterations we go in a dataflow pass
10155 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10156 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10158 #endif // DATAFLOW_ITER
10160 #if MEASURE_BLOCK_SIZE
10161 extern size_t genFlowNodeSize;
10162 extern size_t genFlowNodeCnt;
10163 #endif // MEASURE_BLOCK_SIZE
10165 #if MEASURE_NODE_SIZE
10166 struct NodeSizeStats
10170 genTreeNodeCnt = 0;
10171 genTreeNodeSize = 0;
10172 genTreeNodeActualSize = 0;
10175 // Count of tree nodes allocated.
10176 unsigned __int64 genTreeNodeCnt;
10178 // The size we allocate.
10179 unsigned __int64 genTreeNodeSize;
10181 // The actual size of the node. Note that the actual size will likely be smaller
10182 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10183 // a smaller node to a larger one. TODO-Cleanup: add stats on
10184 // SetOper()/ChangeOper() usage to quantify this.
10185 unsigned __int64 genTreeNodeActualSize;
10187 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10188 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10189 extern Histogram genTreeNcntHist;
10190 extern Histogram genTreeNsizHist;
10191 #endif // MEASURE_NODE_SIZE
10193 /*****************************************************************************
10194 * Count fatal errors (including noway_asserts).
10198 extern unsigned fatal_badCode;
10199 extern unsigned fatal_noWay;
10200 extern unsigned fatal_NOMEM;
10201 extern unsigned fatal_noWayAssertBody;
10203 extern unsigned fatal_noWayAssertBodyArgs;
10205 extern unsigned fatal_NYI;
10206 #endif // MEASURE_FATAL
10208 /*****************************************************************************
10212 #ifdef _TARGET_XARCH_
10214 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10215 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10216 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10218 const instruction INS_AND = INS_and;
10219 const instruction INS_OR = INS_or;
10220 const instruction INS_XOR = INS_xor;
10221 const instruction INS_NEG = INS_neg;
10222 const instruction INS_TEST = INS_test;
10223 const instruction INS_MUL = INS_imul;
10224 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10225 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10226 const instruction INS_BREAKPOINT = INS_int3;
10227 const instruction INS_ADDC = INS_adc;
10228 const instruction INS_SUBC = INS_sbb;
10229 const instruction INS_NOT = INS_not;
10233 #ifdef _TARGET_ARM_
10235 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10236 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10237 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10239 const instruction INS_AND = INS_and;
10240 const instruction INS_OR = INS_orr;
10241 const instruction INS_XOR = INS_eor;
10242 const instruction INS_NEG = INS_rsb;
10243 const instruction INS_TEST = INS_tst;
10244 const instruction INS_MUL = INS_mul;
10245 const instruction INS_MULADD = INS_mla;
10246 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10247 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10248 const instruction INS_BREAKPOINT = INS_bkpt;
10249 const instruction INS_ADDC = INS_adc;
10250 const instruction INS_SUBC = INS_sbc;
10251 const instruction INS_NOT = INS_mvn;
10253 const instruction INS_ABS = INS_vabs;
10254 const instruction INS_ROUND = INS_invalid;
10255 const instruction INS_SQRT = INS_vsqrt;
10259 #ifdef _TARGET_ARM64_
10261 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10262 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10263 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10265 const instruction INS_AND = INS_and;
10266 const instruction INS_OR = INS_orr;
10267 const instruction INS_XOR = INS_eor;
10268 const instruction INS_NEG = INS_neg;
10269 const instruction INS_TEST = INS_tst;
10270 const instruction INS_MUL = INS_mul;
10271 const instruction INS_MULADD = INS_madd;
10272 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10273 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10274 const instruction INS_BREAKPOINT = INS_bkpt;
10275 const instruction INS_ADDC = INS_adc;
10276 const instruction INS_SUBC = INS_sbc;
10277 const instruction INS_NOT = INS_mvn;
10279 const instruction INS_ABS = INS_fabs;
10280 const instruction INS_ROUND = INS_frintn;
10281 const instruction INS_SQRT = INS_fsqrt;
10285 /*****************************************************************************/
10287 extern const BYTE genTypeSizes[];
10288 extern const BYTE genTypeAlignments[];
10289 extern const BYTE genTypeStSzs[];
10290 extern const BYTE genActualTypes[];
10292 /*****************************************************************************/
10294 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10295 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10297 #ifdef _TARGET_ARM_
10298 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10299 #elif defined(_TARGET_ARM64_)
10300 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10303 /*****************************************************************************/
10305 #define REG_CORRUPT regNumber(REG_NA + 1)
10306 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
10307 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
10309 /*****************************************************************************/
10311 extern BasicBlock dummyBB;
10313 /*****************************************************************************/
10314 /*****************************************************************************/
10316 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10317 // nodes of a statement in execution order.
10318 // __stmt: a GT_STMT type GenTree*
10319 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10321 #define foreach_treenode_execution_order(__node, __stmt) \
10322 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10324 // foreach_block: An iterator over all blocks in the function.
10325 // __compiler: the Compiler* object
10326 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10328 #define foreach_block(__compiler, __block) \
10329 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10331 /*****************************************************************************/
10332 /*****************************************************************************/
10336 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10338 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10339 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10341 XX Debugging helpers XX
10343 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10344 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10347 /*****************************************************************************/
10348 /* The following functions are intended to be called from the debugger, to dump
10349 * various data structures. The can be used in the debugger Watch or Quick Watch
10350 * windows. They are designed to be short to type and take as few arguments as
10351 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10352 * See the function definition comment for more details.
10355 void cBlock(Compiler* comp, BasicBlock* block);
10356 void cBlocks(Compiler* comp);
10357 void cBlocksV(Compiler* comp);
10358 void cTree(Compiler* comp, GenTree* tree);
10359 void cTrees(Compiler* comp);
10360 void cEH(Compiler* comp);
10361 void cVar(Compiler* comp, unsigned lclNum);
10362 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10363 void cVars(Compiler* comp);
10364 void cVarsFinal(Compiler* comp);
10365 void cBlockPreds(Compiler* comp, BasicBlock* block);
10366 void cReach(Compiler* comp);
10367 void cDoms(Compiler* comp);
10368 void cLiveness(Compiler* comp);
10369 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10371 void cFuncIR(Compiler* comp);
10372 void cBlockIR(Compiler* comp, BasicBlock* block);
10373 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10374 void cTreeIR(Compiler* comp, GenTree* tree);
10375 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10376 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10377 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10378 int cOperandIR(Compiler* comp, GenTree* operand);
10379 int cLeafIR(Compiler* comp, GenTree* tree);
10380 int cIndirIR(Compiler* comp, GenTree* tree);
10381 int cListIR(Compiler* comp, GenTree* list);
10382 int cSsaNumIR(Compiler* comp, GenTree* tree);
10383 int cValNumIR(Compiler* comp, GenTree* tree);
10384 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10386 void dBlock(BasicBlock* block);
10389 void dTree(GenTree* tree);
10392 void dVar(unsigned lclNum);
10393 void dVarDsc(LclVarDsc* varDsc);
10396 void dBlockPreds(BasicBlock* block);
10400 void dCVarSet(VARSET_VALARG_TP vars);
10402 void dVarSet(VARSET_VALARG_TP vars);
10403 void dRegMask(regMaskTP mask);
10406 void dBlockIR(BasicBlock* block);
10407 void dTreeIR(GenTree* tree);
10408 void dLoopIR(Compiler::LoopDsc* loop);
10409 void dLoopNumIR(unsigned loopNum);
10410 int dTabStopIR(int curr, int tabstop);
10411 int dTreeTypeIR(GenTree* tree);
10412 int dTreeKindsIR(GenTree* tree);
10413 int dTreeFlagsIR(GenTree* tree);
10414 int dOperandIR(GenTree* operand);
10415 int dLeafIR(GenTree* tree);
10416 int dIndirIR(GenTree* tree);
10417 int dListIR(GenTree* list);
10418 int dSsaNumIR(GenTree* tree);
10419 int dValNumIR(GenTree* tree);
10420 int dDependsIR(GenTree* comma);
10423 GenTree* dFindTree(GenTree* tree, unsigned id);
10424 GenTree* dFindTree(unsigned id);
10425 GenTreeStmt* dFindStmt(unsigned id);
10426 BasicBlock* dFindBlock(unsigned bbNum);
10430 #include "compiler.hpp" // All the shared inline functions
10432 /*****************************************************************************/
10433 #endif //_COMPILER_H_
10434 /*****************************************************************************/