1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
104 // Or the more-general IAllocator interface.
105 void* __cdecl operator new(size_t n, IAllocator* alloc);
106 void* __cdecl operator new[](size_t n, IAllocator* alloc);
108 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
109 // caused these to be ambiguous with the global placement new operators.
110 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
111 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
114 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
115 #include "loopcloning.h"
117 /*****************************************************************************/
119 /* This is included here and not earlier as it needs the definition of "CSE"
120 * which is defined in the section above */
122 /*****************************************************************************/
124 unsigned genLog2(unsigned value);
125 unsigned genLog2(unsigned __int64 value);
127 var_types genActualType(var_types type);
128 var_types genUnsignedType(var_types type);
129 var_types genSignedType(var_types type);
131 unsigned ReinterpretHexAsDecimal(unsigned);
133 /*****************************************************************************/
136 #ifdef FEATURE_AVX_SUPPORT
137 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
138 #else // !FEATURE_AVX_SUPPORT
139 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
140 #endif // !FEATURE_AVX_SUPPORT
141 #else // !FEATURE_SIMD
142 const unsigned TEMP_MAX_SIZE = sizeof(double);
143 #endif // !FEATURE_SIMD
144 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
146 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
149 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
152 // The following holds the Local var info (scope information)
153 typedef const char* VarName; // Actual ASCII string
156 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
157 IL_OFFSET vsdLifeEnd; // instr offset of end of life
158 unsigned vsdVarNum; // (remapped) LclVarDsc number
161 VarName vsdName; // name of the var
164 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
165 // Also, it is the index of this entry in the info.compVarScopes array,
166 // which is useful since the array is also accessed via the
167 // compEnterScopeList and compExitScopeList sorted arrays.
170 /*****************************************************************************
172 * The following holds the local variable counts and the descriptor table.
175 // This is the location of a definition.
181 DefLoc() : m_blk(nullptr), m_tree(nullptr)
186 // This class encapsulates all info about a local variable that may vary for different SSA names
191 ValueNumPair m_vnPair;
199 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
204 // The constructor. Most things can just be zero'ed.
205 LclVarDsc(Compiler* comp);
207 // note this only packs because var_types is a typedef of unsigned char
208 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
210 unsigned char lvIsParam : 1; // is this a parameter?
211 unsigned char lvIsRegArg : 1; // is this a register argument?
212 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
214 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
215 // is to help determine whether to use block init in the prolog.
216 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
217 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
218 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
219 // variable is in the same register for the entire function.
220 unsigned char lvTracked : 1; // is this a tracked variable?
221 bool lvTrackedNonStruct()
223 return lvTracked && lvType != TYP_STRUCT;
225 unsigned char lvPinned : 1; // is this a pinned variable?
227 unsigned char lvMustInit : 1; // must be initialized
228 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
229 // global location, etc.
230 // We cannot reason reliably about the value of the variable.
231 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
232 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
236 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
238 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
239 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
240 // read/write its value.
241 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
242 // the variable to be
243 // in the stack (at least at those boundaries.)
244 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
245 // particular byte from an int).
246 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
247 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
249 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
250 unsigned char lvRefAssign : 1; // involved in pointer assignment
251 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
252 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
255 unsigned char lvArgWrite : 1; // variable is a parameter and STARG was used on it
256 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp
258 unsigned char lvIsBoolean : 1; // set if variable is boolean
260 unsigned char lvRngOptDone : 1; // considered for range check opt?
261 unsigned char lvLoopInc : 1; // incremented in the loop?
262 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
263 unsigned char lvArrIndx : 1; // used as an array index?
264 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
265 unsigned char lvArrIndxDom : 1; // index dominates loop exit
267 unsigned char lvSingleDef : 1; // variable has a single def
268 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
269 unsigned char lvVolatileHint : 1; // hint for AssertionProp
272 unsigned char lvSpilled : 1; // enregistered variable was spilled
273 #ifndef _TARGET_64BIT_
274 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
275 #endif // !_TARGET_64BIT_
276 #ifdef _TARGET_64BIT_
277 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
280 unsigned char lvKeepType : 1; // Don't change the type of this variable
281 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
283 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
285 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
286 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
288 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
289 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
290 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
291 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
292 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
294 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
295 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
298 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
299 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
300 // with (lvIsRegArg && lvIsHfa())
301 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
302 #endif // FEATURE_HFA
305 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
306 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
307 // Consider cleaning this up so this workaround is not required.
308 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
309 // I.e. there is no longer any reference to the struct directly.
310 // In this case we can simply remove this struct local.
312 #ifndef LEGACY_BACKEND
313 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
314 #endif // !LEGACY_BACKEND
317 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
318 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
319 unsigned char lvSIMDType : 1; // This is a SIMD struct
320 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
321 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
322 #endif // FEATURE_SIMD
323 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
326 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
328 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
329 // Valid on promoted struct local fields.
332 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
333 unsigned char lvFldOffset;
334 unsigned char lvFldOrdinal;
336 #if FEATURE_MULTIREG_ARGS
337 regNumber lvRegNumForSlot(unsigned slotNum)
343 else if (slotNum == 1)
345 return lvOtherArgReg;
349 assert(false && "Invalid slotNum!");
354 #endif // FEATURE_MULTIREG_ARGS
372 bool lvIsHfaRegArg() const
375 return _lvIsHfaRegArg;
381 void lvSetIsHfaRegArg()
384 _lvIsHfaRegArg = true;
388 bool lvHfaTypeIsFloat() const
391 return _lvHfaTypeIsFloat;
397 void lvSetHfaTypeIsFloat(bool value)
400 _lvHfaTypeIsFloat = value;
404 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
405 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
407 unsigned lvHfaSlots() const
410 assert(lvType == TYP_STRUCT);
412 return lvExactSize / sizeof(float);
413 #else // _TARGET_ARM64_
414 if (lvHfaTypeIsFloat())
416 return lvExactSize / sizeof(float);
420 return lvExactSize / sizeof(double);
422 #endif // _TARGET_ARM64_
425 // lvIsMultiRegArgOrRet()
426 // returns true if this is a multireg LclVar struct used in an argument context
427 // or if this is a multireg LclVar struct assigned from a multireg call
428 bool lvIsMultiRegArgOrRet()
430 return lvIsMultiRegArg || lvIsMultiRegRet;
434 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
435 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
436 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
437 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
438 // to non-zero if the variable gets the same register assignment for its entire
440 #if !defined(_TARGET_64BIT_)
441 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
442 #endif // !defined(_TARGET_64BIT_)
444 regNumberSmall _lvArgReg; // The register in which this argument is passed.
446 #if FEATURE_MULTIREG_ARGS
447 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
448 // Note this is defined but not used by ARM32
449 #endif // FEATURE_MULTIREG_ARGS
451 #ifndef LEGACY_BACKEND
453 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
454 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
456 #endif // !LEGACY_BACKEND
459 // The register number is stored in a small format (8 bits), but the getters return and the setters take
460 // a full-size (unsigned) format, to localize the casts here.
462 /////////////////////
464 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
466 regNumber GetRegNum() const
468 return (regNumber)_lvRegNum;
471 void SetRegNum(regNumber reg)
473 _lvRegNum = (regNumberSmall)reg;
474 assert(_lvRegNum == reg);
477 /////////////////////
479 #if defined(_TARGET_64BIT_)
480 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
482 regNumber GetOtherReg() const
484 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
485 // "unreachable code" warnings
489 void SetOtherReg(regNumber reg)
491 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
492 // "unreachable code" warnings
494 #else // !_TARGET_64BIT_
495 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
497 regNumber GetOtherReg() const
499 return (regNumber)_lvOtherReg;
502 void SetOtherReg(regNumber reg)
504 _lvOtherReg = (regNumberSmall)reg;
505 assert(_lvOtherReg == reg);
507 #endif // !_TARGET_64BIT_
509 /////////////////////
511 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
513 regNumber GetArgReg() const
515 return (regNumber)_lvArgReg;
518 void SetArgReg(regNumber reg)
520 _lvArgReg = (regNumberSmall)reg;
521 assert(_lvArgReg == reg);
524 #if FEATURE_MULTIREG_ARGS
525 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
527 regNumber GetOtherArgReg() const
529 return (regNumber)_lvOtherArgReg;
532 void SetOtherArgReg(regNumber reg)
534 _lvOtherArgReg = (regNumberSmall)reg;
535 assert(_lvOtherArgReg == reg);
537 #endif // FEATURE_MULTIREG_ARGS
540 // Is this is a SIMD struct?
541 bool lvIsSIMDType() const
546 // Is this is a SIMD struct which is used for SIMD intrinsic?
547 bool lvIsUsedInSIMDIntrinsic() const
549 return lvUsedInSIMDIntrinsic;
552 // If feature_simd not enabled, return false
553 bool lvIsSIMDType() const
557 bool lvIsUsedInSIMDIntrinsic() const
563 /////////////////////
565 #ifndef LEGACY_BACKEND
566 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
568 regNumber GetArgInitReg() const
570 return (regNumber)_lvArgInitReg;
573 void SetArgInitReg(regNumber reg)
575 _lvArgInitReg = (regNumberSmall)reg;
576 assert(_lvArgInitReg == reg);
579 /////////////////////
581 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
583 regPairNo GetArgInitRegPair() const
585 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
586 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
590 void SetArgInitRegPair(regPairNo regPair)
592 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
593 _lvArgInitRegPair = (regPairNoSmall)regPair;
594 assert(_lvArgInitRegPair == regPair);
597 /////////////////////
599 bool lvIsRegCandidate() const
601 return lvLRACandidate != 0;
604 bool lvIsInReg() const
606 return lvIsRegCandidate() && (lvRegNum != REG_STK);
609 #else // LEGACY_BACKEND
611 bool lvIsRegCandidate() const
613 return lvTracked != 0;
616 bool lvIsInReg() const
618 return lvRegister != 0;
621 #endif // LEGACY_BACKEND
623 regMaskTP lvRegMask() const
625 regMaskTP regMask = RBM_NONE;
626 if (varTypeIsFloating(TypeGet()))
628 if (lvRegNum != REG_STK)
630 regMask = genRegMaskFloat(lvRegNum, TypeGet());
635 if (lvRegNum != REG_STK)
637 regMask = genRegMask(lvRegNum);
640 // For longs we may have two regs
641 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
643 regMask |= genRegMask(lvOtherReg);
649 regMaskSmall lvPrefReg; // set of regs it prefers to live in
651 unsigned short lvVarIndex; // variable tracking index
652 unsigned short lvRefCnt; // unweighted (real) reference count
653 unsigned lvRefCntWtd; // weighted reference count
654 int lvStkOffs; // stack offset of home
655 unsigned lvExactSize; // (exact) size of the type in bytes
657 // Is this a promoted struct?
658 // This method returns true only for structs (including SIMD structs), not for
659 // locals that are split on a 32-bit target.
660 // It is only necessary to use this:
661 // 1) if only structs are wanted, and
662 // 2) if Lowering has already been done.
663 // Otherwise lvPromoted is valid.
664 bool lvPromotedStruct()
666 #if !defined(_TARGET_64BIT_)
667 return (lvPromoted && !varTypeIsLong(lvType));
668 #else // defined(_TARGET_64BIT_)
670 #endif // defined(_TARGET_64BIT_)
673 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
675 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
676 // where the struct itself is no longer used because all access is via its member fields.
677 // When that happens, the struct is marked as unused and its type has been changed to
678 // TYP_INT (to keep the GC tracking code from looking at it).
679 // See Compiler::raAssignVars() for details. For example:
680 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
681 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
682 // float V03.f1 (offs=0x00) -> V12 tmp7
683 // f8 (last use) (last use) $345
684 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
685 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
687 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
689 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
690 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
691 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
692 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
693 if ((lvType == TYP_SIMD12) && !lvIsParam)
695 assert(lvExactSize == 12);
698 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
700 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
703 unsigned lvSlotNum; // original slot # (if remapped)
705 typeInfo lvVerTypeInfo; // type info needed for verification
707 BYTE* lvGcLayout; // GC layout info for structs
710 BlockSet lvRefBlks; // Set of blocks that contain refs
711 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
712 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
714 var_types TypeGet() const
716 return (var_types)lvType;
718 bool lvStackAligned() const
720 assert(lvIsStructField);
721 return ((lvFldOffset % sizeof(void*)) == 0);
723 bool lvNormalizeOnLoad() const
725 return varTypeIsSmall(TypeGet()) &&
726 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
727 (lvIsParam || lvAddrExposed || lvIsStructField);
730 bool lvNormalizeOnStore()
732 return varTypeIsSmall(TypeGet()) &&
733 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
734 !(lvIsParam || lvAddrExposed || lvIsStructField);
737 void lvaResetSortAgainFlag(Compiler* pComp);
738 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
739 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
740 void setPrefReg(regNumber regNum, Compiler* pComp);
741 void addPrefReg(regMaskTP regMask, Compiler* pComp);
742 bool IsFloatRegType() const
744 return isFloatRegType(lvType) || lvIsHfaRegArg();
746 var_types GetHfaType() const
748 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
750 void SetHfaType(var_types type)
752 assert(varTypeIsFloating(type));
753 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
756 #ifndef LEGACY_BACKEND
757 var_types lvaArgType();
760 PerSsaArray lvPerSsaData;
763 // Keep track of the # of SsaNames, for a bounds check.
764 unsigned lvNumSsaNames;
767 // Returns the address of the per-Ssa data for the given ssaNum (which is required
768 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
769 // not an SSA variable).
770 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
772 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
773 assert(SsaConfig::RESERVED_SSA_NUM == 0);
774 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
775 assert(zeroBased < lvNumSsaNames);
776 return &lvPerSsaData.GetRef(zeroBased);
781 void PrintVarReg() const
783 if (isRegPairType(TypeGet()))
785 printf("%s:%s", getRegName(lvOtherReg), // hi32
786 getRegName(lvRegNum)); // lo32
790 printf("%s", getRegName(lvRegNum));
795 }; // class LclVarDsc
798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
799 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
803 XX The temporary lclVars allocated by the compiler for code generation XX
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
809 /*****************************************************************************
811 * The following keeps track of temporaries allocated in the stack frame
812 * during code-generation (after register allocation). These spill-temps are
813 * only used if we run out of registers while evaluating a tree.
815 * These are different from the more common temps allocated by lvaGrabTemp().
826 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
834 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
838 0); // temps must have a negative number (so they have a different number from all local variables)
839 tdOffs = BAD_TEMP_OFFSET;
843 IMPL_LIMITATION("too many spill temps");
848 bool tdLegalOffset() const
850 return tdOffs != BAD_TEMP_OFFSET;
854 int tdTempOffs() const
856 assert(tdLegalOffset());
859 void tdSetTempOffs(int offs)
862 assert(tdLegalOffset());
864 void tdAdjustTempOffs(int offs)
867 assert(tdLegalOffset());
870 int tdTempNum() const
875 unsigned tdTempSize() const
879 var_types tdTempType() const
885 // interface to hide linearscan implementation from rest of compiler
886 class LinearScanInterface
889 virtual void doLinearScan() = 0;
890 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
893 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
895 // Information about arrays: their element type and size, and the offset of the first element.
896 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
897 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
898 // for example, in value numbering of array index expressions.
901 var_types m_elemType;
902 CORINFO_CLASS_HANDLE m_elemStructType;
904 unsigned m_elemOffset;
906 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
910 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
911 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
916 // This enumeration names the phases into which we divide compilation. The phases should completely
917 // partition a compilation.
920 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
921 #include "compphases.h"
925 extern const char* PhaseNames[];
926 extern const char* PhaseEnums[];
927 extern const LPCWSTR PhaseShortNames[];
929 // The following enum provides a simple 1:1 mapping to CLR API's
930 enum API_ICorJitInfo_Names
932 #define DEF_CLR_API(name) API_##name,
933 #include "ICorJitInfo_API_names.h"
937 //---------------------------------------------------------------
941 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
942 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
943 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
944 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
945 // by "m_timerFailure" being true.
946 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
949 #ifdef FEATURE_JIT_METHOD_PERF
950 // The string names of the phases.
951 static const char* PhaseNames[];
953 static bool PhaseHasChildren[];
954 static int PhaseParent[];
956 unsigned m_byteCodeBytes;
957 unsigned __int64 m_totalCycles;
958 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
959 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
960 #if MEASURE_CLRAPI_CALLS
961 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
962 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
964 // For better documentation, we call EndPhase on
965 // non-leaf phases. We should also call EndPhase on the
966 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
967 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
968 // We add all such "redundant end phase" intervals to this variable below; we print
969 // it out in a report, so we can verify that it is, indeed, very small. If it ever
970 // isn't, this means that we're doing something significant between the end of the last
971 // declared subphase and the end of its parent.
972 unsigned __int64 m_parentPhaseEndSlop;
975 #if MEASURE_CLRAPI_CALLS
976 // The following measures the time spent inside each individual CLR API call.
977 unsigned m_allClrAPIcalls;
978 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
979 unsigned __int64 m_allClrAPIcycles;
980 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
981 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
982 #endif // MEASURE_CLRAPI_CALLS
984 CompTimeInfo(unsigned byteCodeBytes);
988 #ifdef FEATURE_JIT_METHOD_PERF
990 #if MEASURE_CLRAPI_CALLS
991 struct WrapICorJitInfo;
994 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
995 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
996 // The operation of adding a single method's timing to the summary may be performed concurrently by several
997 // threads, so it is protected by a lock.
998 // This class is intended to be used as a singleton type, with only a single instance.
999 class CompTimeSummaryInfo
1001 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1002 static CritSecObject s_compTimeSummaryLock;
1006 CompTimeInfo m_total;
1007 CompTimeInfo m_maximum;
1009 int m_numFilteredMethods;
1010 CompTimeInfo m_filtered;
1012 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1013 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1014 // If any OS interaction fails, returns 0.0.
1015 double CyclesPerSecond();
1017 // This can use what ever data you want to determine if the value to be added
1018 // belongs in the filtered section (it's always included in the unfiltered section)
1019 bool IncludedInFilteredData(CompTimeInfo& info);
1022 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1023 static CompTimeSummaryInfo s_compTimeSummary;
1025 CompTimeSummaryInfo()
1026 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1030 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1031 // This is thread safe.
1032 void AddInfo(CompTimeInfo& info, bool includePhases);
1034 // Print the summary information to "f".
1035 // This is not thread-safe; assumed to be called by only one thread.
1036 void Print(FILE* f);
1039 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1040 // and when the current phase started. This is intended to be part of a Compilation object. This is
1041 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1045 unsigned __int64 m_start; // Start of the compilation.
1046 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1047 #if MEASURE_CLRAPI_CALLS
1048 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1049 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1050 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1051 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1052 static double s_cyclesPerSec; // Cached for speedier measurements
1055 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1057 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1059 static CritSecObject s_csvLock; // Lock to protect the time log file.
1060 void PrintCsvMethodStats(Compiler* comp);
1063 void* operator new(size_t);
1064 void* operator new[](size_t);
1065 void operator delete(void*);
1066 void operator delete[](void*);
1069 // Initialized the timer instance
1070 JitTimer(unsigned byteCodeSize);
1072 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1074 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1077 static void PrintCsvHeader();
1079 // Ends the current phase (argument is for a redundant check).
1080 void EndPhase(Phases phase);
1082 #if MEASURE_CLRAPI_CALLS
1083 // Start and end a timed CLR API call.
1084 void CLRApiCallEnter(unsigned apix);
1085 void CLRApiCallLeave(unsigned apix);
1086 #endif // MEASURE_CLRAPI_CALLS
1088 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1089 // and adds it to "sum".
1090 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1092 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1093 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1094 // "m_info" to true.
1095 bool GetThreadCycles(unsigned __int64* cycles)
1097 bool res = CycleTimer::GetThreadCyclesS(cycles);
1100 m_info.m_timerFailure = true;
1105 #endif // FEATURE_JIT_METHOD_PERF
1107 //------------------- Function/Funclet info -------------------------------
1108 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1110 FUNC_ROOT, // The main/root function (always id==0)
1111 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1112 FUNC_FILTER, // a funclet associated with an EH filter
1115 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1122 BYTE funFlags; // Currently unused, just here for padding
1123 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1124 // funclet. It is only valid if funKind field indicates this is a
1125 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1127 #if defined(_TARGET_AMD64_)
1129 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1130 emitLocation* startLoc;
1131 emitLocation* endLoc;
1132 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1133 emitLocation* coldEndLoc;
1134 UNWIND_INFO unwindHeader;
1135 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1136 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1137 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1138 unsigned unwindCodeSlot;
1140 #ifdef UNIX_AMD64_ABI
1141 jitstd::vector<CFI_CODE>* cfiCodes;
1142 #endif // UNIX_AMD64_ABI
1144 #elif defined(_TARGET_ARMARCH_)
1146 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1147 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1148 // Note: we only have a pointer here instead of the actual object,
1149 // to save memory in the JIT case (compared to the NGEN case),
1150 // where we don't have any cold section.
1151 // Note 2: we currently don't support hot/cold splitting in functions
1152 // with EH, so uwiCold will be NULL for all funclets.
1154 #endif // _TARGET_ARMARCH_
1156 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1157 // that isn't shared between the main function body and funclets.
1160 struct fgArgTabEntry
1163 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1166 otherRegNum = REG_NA;
1167 isStruct = false; // is this a struct arg
1169 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1171 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1173 // it will point at the actual argument in the gtCallLateArgs list.
1174 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1176 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1178 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1180 unsigned numRegs; // Count of number of registers that this argument uses
1182 // A slot is a pointer sized region in the OutArg area.
1183 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1184 unsigned numSlots; // Count of number of slots that this argument uses
1186 unsigned alignment; // 1 or 2 (slots/registers)
1187 unsigned lateArgInx; // index into gtCallLateArgs list
1188 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1190 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1191 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1192 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1193 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1194 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1195 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1196 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1197 // previous arguments.
1198 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1199 // to be on the stack despite its arg list position.
1201 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1202 bool isStruct : 1; // True if this is a struct arg
1204 regNumber otherRegNum; // The (second) register to use when passing this argument.
1206 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1207 #elif defined(_TARGET_X86_)
1208 __declspec(property(get = getIsStruct)) bool isStruct;
1211 return varTypeIsStruct(node);
1213 #endif // _TARGET_X86_
1216 void SetIsHfaRegArg(bool hfaRegArg)
1218 isHfaRegArg = hfaRegArg;
1221 void SetIsBackFilled(bool backFilled)
1223 isBackFilled = backFilled;
1226 bool IsBackFilled() const
1228 return isBackFilled;
1230 #else // !_TARGET_ARM_
1231 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1233 void SetIsHfaRegArg(bool hfaRegArg)
1237 void SetIsBackFilled(bool backFilled)
1241 bool IsBackFilled() const
1245 #endif // !_TARGET_ARM_
1251 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1253 //-------------------------------------------------------------------------
1255 // The class fgArgInfo is used to handle the arguments
1256 // when morphing a GT_CALL node.
1261 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1262 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1263 unsigned argCount; // Updatable arg count value
1264 unsigned nextSlotNum; // Updatable slot count value
1265 unsigned stkLevel; // Stack depth when we make this call (for x86)
1267 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1268 bool hasRegArgs; // true if we have one or more register arguments
1269 bool hasStackArgs; // true if we have one or more stack arguments
1270 bool argsComplete; // marker for state
1271 bool argsSorted; // marker for state
1272 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1275 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1278 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1279 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1281 fgArgTabEntryPtr AddRegArg(
1282 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1284 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1285 fgArgTabEntryPtr AddRegArg(
1292 const bool isStruct,
1293 const regNumber otherRegNum = REG_NA,
1294 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1295 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1297 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1301 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1303 void RemorphReset();
1304 fgArgTabEntryPtr RemorphRegArg(
1305 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1307 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1309 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1311 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1313 void ArgsComplete();
1317 void EvalArgsToTemps();
1319 void RecordStkLevel(unsigned stkLvl);
1320 unsigned RetrieveStkLevel();
1326 fgArgTabEntryPtr* ArgTable()
1330 unsigned GetNextSlotNum()
1340 return hasStackArgs;
1342 bool AreArgsComplete() const
1344 return argsComplete;
1349 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1350 // We have the ability to mark source expressions with "Test Labels."
1351 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1352 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1354 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1357 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1358 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1359 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1360 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1361 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1364 struct TestLabelAndNum
1369 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1374 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1376 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1379 // This class implements the "IAllocator" interface, so that we can use
1380 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1382 class CompAllocator : public IAllocator
1385 #if MEASURE_MEM_ALLOC
1389 CompAllocator(Compiler* comp, CompMemKind cmk)
1391 #if MEASURE_MEM_ALLOC
1397 inline void* Alloc(size_t sz);
1399 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1401 // For the compiler's no-release allocator, free operations are no-ops.
1408 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1409 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1411 XX The big guy. The sections are currently organized as : XX
1413 XX o GenTree and BasicBlock XX
1425 XX o PrologScopeInfo XX
1426 XX o CodeGenerator XX
1431 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1432 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1437 friend class emitter;
1438 friend class UnwindInfo;
1439 friend class UnwindFragmentInfo;
1440 friend class UnwindEpilogInfo;
1441 friend class JitTimer;
1442 friend class LinearScan;
1443 friend class fgArgInfo;
1444 friend class Rationalizer;
1446 friend class Lowering;
1447 friend class CSE_DataFlow;
1448 friend class CSE_Heuristic;
1449 friend class CodeGenInterface;
1450 friend class CodeGen;
1451 friend class LclVarDsc;
1452 friend class TempDsc;
1454 friend class ObjectAllocator;
1456 #ifndef _TARGET_64BIT_
1457 friend class DecomposeLongs;
1458 #endif // !_TARGET_64BIT_
1461 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1462 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1464 XX Misc structs definitions XX
1466 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1467 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1471 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1490 bool dumpIRDataflow;
1491 bool dumpIRBlockHeaders;
1493 LPCWSTR dumpIRPhase;
1494 LPCWSTR dumpIRFormat;
1496 bool shouldUseVerboseTrees();
1497 bool asciiTrees; // If true, dump trees using only ASCII characters
1498 bool shouldDumpASCIITrees();
1499 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1500 bool shouldUseVerboseSsa();
1501 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1502 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1504 const char* VarNameToStr(VarName name)
1509 DWORD expensiveDebugCheckLevel;
1512 #if FEATURE_MULTIREG_RET
1513 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1514 #endif // FEATURE_MULTIREG_RET
1517 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1518 #endif // ARM_SOFTFP
1520 //-------------------------------------------------------------------------
1521 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1522 // HFAs are one to four element structs where each element is the same
1523 // type, either all float or all double. They are treated specially
1524 // in the ARM Procedure Call Standard, specifically, they are passed in
1525 // floating-point registers instead of the general purpose registers.
1528 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1529 bool IsHfa(GenTreePtr tree);
1531 var_types GetHfaType(GenTreePtr tree);
1532 unsigned GetHfaCount(GenTreePtr tree);
1534 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1535 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1537 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1538 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1540 //-------------------------------------------------------------------------
1541 // The following is used for validating format of EH table
1545 typedef struct EHNodeDsc* pEHNodeDsc;
1547 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1548 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1561 EHBlockType ehnBlockType; // kind of EH block
1562 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1563 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1564 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1566 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1567 pEHNodeDsc ehnChild; // leftmost nested block
1569 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1570 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1572 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1573 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1575 inline void ehnSetTryNodeType()
1577 ehnBlockType = TryNode;
1579 inline void ehnSetFilterNodeType()
1581 ehnBlockType = FilterNode;
1583 inline void ehnSetHandlerNodeType()
1585 ehnBlockType = HandlerNode;
1587 inline void ehnSetFinallyNodeType()
1589 ehnBlockType = FinallyNode;
1591 inline void ehnSetFaultNodeType()
1593 ehnBlockType = FaultNode;
1596 inline BOOL ehnIsTryBlock()
1598 return ehnBlockType == TryNode;
1600 inline BOOL ehnIsFilterBlock()
1602 return ehnBlockType == FilterNode;
1604 inline BOOL ehnIsHandlerBlock()
1606 return ehnBlockType == HandlerNode;
1608 inline BOOL ehnIsFinallyBlock()
1610 return ehnBlockType == FinallyNode;
1612 inline BOOL ehnIsFaultBlock()
1614 return ehnBlockType == FaultNode;
1617 // returns true if there is any overlap between the two nodes
1618 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1620 if (node1->ehnStartOffset < node2->ehnStartOffset)
1622 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1626 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1630 // fails with BADCODE if inner is not completely nested inside outer
1631 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1633 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1637 //-------------------------------------------------------------------------
1638 // Exception handling functions
1641 #if !FEATURE_EH_FUNCLETS
1643 bool ehNeedsShadowSPslots()
1645 return (info.compXcptnsCount || opts.compDbgEnC);
1648 // 0 for methods with no EH
1649 // 1 for methods with non-nested EH, or where only the try blocks are nested
1650 // 2 for a method with a catch within a catch
1652 unsigned ehMaxHndNestingCount;
1654 #endif // !FEATURE_EH_FUNCLETS
1656 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1657 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1659 bool bbInCatchHandlerILRange(BasicBlock* blk);
1660 bool bbInFilterILRange(BasicBlock* blk);
1661 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1662 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1663 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1664 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1665 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1667 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1668 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1670 // Returns true if "block" is the start of a try region.
1671 bool bbIsTryBeg(BasicBlock* block);
1673 // Returns true if "block" is the start of a handler or filter region.
1674 bool bbIsHandlerBeg(BasicBlock* block);
1676 // Returns true iff "block" is where control flows if an exception is raised in the
1677 // try region, and sets "*regionIndex" to the index of the try for the handler.
1678 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1679 // block of the filter, but not for the filter's handler.
1680 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1682 bool ehHasCallableHandlers();
1684 // Return the EH descriptor for the given region index.
1685 EHblkDsc* ehGetDsc(unsigned regionIndex);
1687 // Return the EH index given a region descriptor.
1688 unsigned ehGetIndex(EHblkDsc* ehDsc);
1690 // Return the EH descriptor index of the enclosing try, for the given region index.
1691 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1693 // Return the EH descriptor index of the enclosing handler, for the given region index.
1694 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1696 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1697 // block is not in a 'try' region).
1698 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1700 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1701 // if this block is not in a filter or handler region).
1702 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1704 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1705 // nullptr if this block's exceptions propagate to caller).
1706 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1708 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1709 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1710 bool ehIsBlockEHLast(BasicBlock* block);
1712 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1714 // Return the region index of the most nested EH region this block is in.
1715 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1717 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1718 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1720 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1721 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1722 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1723 // (It can never be a filter.)
1724 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1726 // A block has been deleted. Update the EH table appropriately.
1727 void ehUpdateForDeletedBlock(BasicBlock* block);
1729 // Determine whether a block can be deleted while preserving the EH normalization rules.
1730 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1732 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1733 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1735 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1736 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1737 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1738 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1739 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1740 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1741 // lives in a filter.)
1742 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1744 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1745 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1746 // (nullptr if the last block is the last block in the program).
1747 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1748 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1751 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1752 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1753 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1756 #if FEATURE_EH_FUNCLETS
1757 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1758 // if there is a filter that protects a region with a nested EH clause (such as a
1759 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1760 // genFuncletProlog() for more details. However, the VM seems to use it for more
1761 // purposes, maybe including debugging. Until we are sure otherwise, always create
1762 // a PSPSym for functions with any EH.
1763 bool ehNeedsPSPSym() const
1767 #else // _TARGET_X86_
1768 return compHndBBtabCount > 0;
1769 #endif // _TARGET_X86_
1772 bool ehAnyFunclets(); // Are there any funclets in this function?
1773 unsigned ehFuncletCount(); // Return the count of funclets in the function
1775 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1776 #else // !FEATURE_EH_FUNCLETS
1777 bool ehAnyFunclets()
1781 unsigned ehFuncletCount()
1786 unsigned bbThrowIndex(BasicBlock* blk)
1788 return blk->bbTryIndex;
1789 } // Get the index to use as the cache key for sharing throw blocks
1790 #endif // !FEATURE_EH_FUNCLETS
1792 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1793 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1794 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1795 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1796 // convenient to also consider it a predecessor.)
1797 flowList* BlockPredsWithEH(BasicBlock* blk);
1799 // This table is useful for memoization of the method above.
1800 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1802 BlockToFlowListMap* m_blockToEHPreds;
1803 BlockToFlowListMap* GetBlockToEHPreds()
1805 if (m_blockToEHPreds == nullptr)
1807 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1809 return m_blockToEHPreds;
1812 void* ehEmitCookie(BasicBlock* block);
1813 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1815 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1817 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1819 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1821 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1823 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1825 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1827 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1829 void fgAllocEHTable();
1831 void fgRemoveEHTableEntry(unsigned XTnum);
1833 #if FEATURE_EH_FUNCLETS
1835 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1837 #endif // FEATURE_EH_FUNCLETS
1841 #endif // !FEATURE_EH
1843 void fgSortEHTable();
1845 // Causes the EH table to obey some well-formedness conditions, by inserting
1846 // empty BB's when necessary:
1847 // * No block is both the first block of a handler and the first block of a try.
1848 // * No block is the first block of multiple 'try' regions.
1849 // * No block is the last block of multiple EH regions.
1850 void fgNormalizeEH();
1851 bool fgNormalizeEHCase1();
1852 bool fgNormalizeEHCase2();
1853 bool fgNormalizeEHCase3();
1856 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1857 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1858 void fgVerifyHandlerTab();
1859 void fgDispHandlerTab();
1862 bool fgNeedToSortEHTable;
1864 void verInitEHTree(unsigned numEHClauses);
1865 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1866 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1867 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1868 void verCheckNestingLevel(EHNodeDsc* initRoot);
1871 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1872 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1874 XX GenTree and BasicBlock XX
1876 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1878 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1879 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1882 // Functions to create nodes
1883 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1886 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1888 // For binary opers.
1889 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1891 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1893 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1894 var_types type = TYP_I_IMPL,
1895 GenTreePtr op1 = nullptr,
1896 GenTreePtr op2 = nullptr);
1898 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1900 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1902 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1904 GenTreePtr gtNewJmpTableNode();
1905 GenTreePtr gtNewIconHandleNode(
1906 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1908 unsigned gtTokenToIconFlags(unsigned token);
1910 GenTreePtr gtNewIconEmbHndNode(void* value,
1913 unsigned handle1 = 0,
1914 void* handle2 = nullptr,
1915 void* compileTimeHandle = nullptr);
1917 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1918 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1919 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1920 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1922 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1924 GenTreePtr gtNewLconNode(__int64 value);
1926 GenTreePtr gtNewDconNode(double value);
1928 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1930 GenTreePtr gtNewZeroConNode(var_types type);
1932 GenTreePtr gtNewOneConNode(var_types type);
1935 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1936 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1939 GenTreeBlk* gtNewBlkOpNode(
1940 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1942 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1945 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1948 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1949 void gtSetObjGcInfo(GenTreeObj* objNode);
1950 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1951 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1953 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1955 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1957 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1958 CORINFO_METHOD_HANDLE handle,
1960 GenTreeArgList* args,
1961 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1963 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1965 GenTreeArgList* args,
1966 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1968 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1971 GenTreeArgList* args = nullptr);
1973 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1976 GenTreeSIMD* gtNewSIMDNode(
1977 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1978 GenTreeSIMD* gtNewSIMDNode(var_types type,
1981 SIMDIntrinsicID simdIntrinsicID,
1984 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
1987 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1988 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
1989 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
1991 GenTreePtr gtNewCodeRef(BasicBlock* block);
1993 GenTreePtr gtNewFieldRef(
1994 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
1996 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
1998 GenTreeArgList* gtNewArgList(GenTreePtr op);
1999 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2000 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2002 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2003 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2004 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2005 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2007 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2009 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2011 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2012 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2013 CORINFO_ACCESS_FLAGS access,
2014 CORINFO_FIELD_INFO* pFieldInfo,
2016 CORINFO_CLASS_HANDLE structType,
2019 GenTreePtr gtNewNothingNode();
2021 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2023 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2025 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2027 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2029 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2031 //------------------------------------------------------------------------
2032 // Other GenTree functions
2034 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2036 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2037 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2038 // IntCnses with value `deepVarVal`.
2039 GenTreePtr gtCloneExpr(
2040 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2042 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2043 // `varNum` to int constants with value `varVal`.
2044 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2046 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2049 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2051 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2053 // Returns "true" iff the complexity (not formally defined, but first interpretation
2054 // is #of nodes in subtree) of "tree" is greater than "limit".
2055 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2056 // before they have been set.)
2057 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2059 bool gtCompareTree(GenTree* op1, GenTree* op2);
2061 GenTreePtr gtReverseCond(GenTree* tree);
2063 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2065 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2067 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2069 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2072 unsigned gtHashValue(GenTree* tree);
2074 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2077 void gtPrepareCost(GenTree* tree);
2078 bool gtIsLikelyRegVar(GenTree* tree);
2080 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2082 // Returns true iff the secondNode can be swapped with firstNode.
2083 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2085 unsigned gtSetEvalOrder(GenTree* tree);
2087 #if FEATURE_STACK_FP_X87
2089 void gtComputeFPlvls(GenTreePtr tree);
2090 #endif // FEATURE_STACK_FP_X87
2092 void gtSetStmtInfo(GenTree* stmt);
2094 // Returns "true" iff "node" has any of the side effects in "flags".
2095 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2097 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2098 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2100 // Appends 'expr' in front of 'list'
2101 // 'list' will typically start off as 'nullptr'
2102 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2103 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2105 void gtExtractSideEffList(GenTreePtr expr,
2107 unsigned flags = GTF_SIDE_EFFECT,
2108 bool ignoreRoot = false);
2110 GenTreePtr gtGetThisArg(GenTreePtr call);
2112 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2113 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2114 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2115 // the given "fldHnd", is such an object pointer.
2116 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2118 // Return true if call is a recursive call; return false otherwise.
2119 bool gtIsRecursiveCall(GenTreeCall* call)
2121 return (call->gtCallMethHnd == info.compMethodHnd);
2124 //-------------------------------------------------------------------------
2126 GenTreePtr gtFoldExpr(GenTreePtr tree);
2129 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2130 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2131 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2132 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2133 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2134 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2135 // optimizations for now.
2136 __attribute__((optnone))
2138 gtFoldExprConst(GenTreePtr tree);
2139 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2140 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2142 //-------------------------------------------------------------------------
2143 // Get the handle, if any.
2144 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2145 // Get the handle, and assert if not found.
2146 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2148 //-------------------------------------------------------------------------
2149 // Functions to display the trees
2152 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2154 void gtDispVN(GenTreePtr tree);
2155 void gtDispConst(GenTreePtr tree);
2156 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2157 void gtDispNodeName(GenTreePtr tree);
2158 void gtDispRegVal(GenTreePtr tree);
2170 void gtDispChild(GenTreePtr child,
2171 IndentStack* indentStack,
2173 __in_opt const char* msg = nullptr,
2174 bool topOnly = false);
2175 void gtDispTree(GenTreePtr tree,
2176 IndentStack* indentStack = nullptr,
2177 __in_opt const char* msg = nullptr,
2178 bool topOnly = false,
2179 bool isLIR = false);
2180 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2181 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2182 char* gtGetLclVarName(unsigned lclNum);
2183 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2184 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2185 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2186 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2187 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2188 void gtDispFieldSeq(FieldSeqNode* pfsn);
2190 void gtDispRange(LIR::ReadOnlyRange const& range);
2192 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2194 void gtDispLIRNode(GenTree* node);
2206 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2207 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2210 static fgWalkPreFn gtAssertColonCond;
2212 static fgWalkPreFn gtMarkColonCond;
2213 static fgWalkPreFn gtClearColonCond;
2215 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2216 bool gtHasCatchArg(GenTreePtr tree);
2217 bool gtHasUnmanagedCall(GenTreePtr tree);
2219 typedef ArrayStack<GenTree*> GenTreeStack;
2221 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2222 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2224 //=========================================================================
2225 // BasicBlock functions
2227 // This is a debug flag we will use to assert when creating block during codegen
2228 // as this interferes with procedure splitting. If you know what you're doing, set
2229 // it to true before creating the block. (DEBUG only)
2230 bool fgSafeBasicBlockCreation;
2233 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2236 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2237 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2241 XX The variables to be used by the code generator. XX
2243 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2244 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2248 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2249 // be placed in the stack frame and it's fields must be laid out sequentially.
2251 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2252 // a local variable that can be enregistered or placed in the stack frame.
2253 // The fields do not need to be laid out sequentially
2255 enum lvaPromotionType
2257 PROMOTION_TYPE_NONE, // The struct local is not promoted
2258 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2259 // and its field locals are independent of its parent struct local.
2260 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2261 // but its field locals depend on its parent struct local.
2264 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2265 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2267 /*****************************************************************************/
2269 enum FrameLayoutState
2272 INITIAL_FRAME_LAYOUT,
2273 PRE_REGALLOC_FRAME_LAYOUT,
2274 REGALLOC_FRAME_LAYOUT,
2275 TENTATIVE_FRAME_LAYOUT,
2280 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2281 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2282 bool lvaSortAgain; // true: We need to sort the lvaTable
2283 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2284 unsigned lvaCount; // total number of locals
2286 unsigned lvaRefCount; // total number of references to locals
2287 LclVarDsc* lvaTable; // variable descriptor table
2288 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2290 LclVarDsc** lvaRefSorted; // table sorted by refcount
2292 unsigned short lvaTrackedCount; // actual # of locals being tracked
2293 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2295 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2296 // Only for AMD64 System V cache the first caller stack homed argument.
2297 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2298 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2301 VARSET_TP lvaTrackedVars; // set of tracked variables
2303 #ifndef _TARGET_64BIT_
2304 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2306 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2308 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2309 // It that changes, this changes. VarSets from different epochs
2310 // cannot be meaningfully combined.
2312 unsigned GetCurLVEpoch()
2317 // reverse map of tracked number to var number
2318 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2320 #ifdef LEGACY_BACKEND
2321 // variable interference graph
2322 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2325 // variable preference graph
2326 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2330 // # of procs compiled a with double-aligned stack
2331 static unsigned s_lvaDoubleAlignedProcsCount;
2335 // Getters and setters for address-exposed and do-not-enregister local var properties.
2336 bool lvaVarAddrExposed(unsigned varNum);
2337 void lvaSetVarAddrExposed(unsigned varNum);
2338 bool lvaVarDoNotEnregister(unsigned varNum);
2340 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2341 enum DoNotEnregisterReason
2346 DNER_VMNeedsStackAddr,
2347 DNER_LiveInOutOfHandler,
2348 DNER_LiveAcrossUnmanagedCall,
2349 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2350 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2351 #ifdef JIT32_GCENCODER
2356 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2358 unsigned lvaVarargsHandleArg;
2360 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2362 #endif // _TARGET_X86_
2364 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2365 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2366 #if FEATURE_FIXED_OUT_ARGS
2367 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2369 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2370 // that tracks whether the lock has been taken
2372 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2373 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2374 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2376 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2377 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2379 #if FEATURE_FIXED_OUT_ARGS
2380 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2381 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2382 #endif // FEATURE_FIXED_OUT_ARGS
2385 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2386 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2387 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2388 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2389 // this variable to be this scratch word whenever struct promotion occurs.
2390 unsigned lvaPromotedStructAssemblyScratchVar;
2391 #endif // _TARGET_ARM_
2394 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2395 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2398 bool lvaGenericsContextUsed;
2400 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2401 // CORINFO_GENERICS_CTXT_FROM_THIS?
2402 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2404 //-------------------------------------------------------------------------
2405 // All these frame offsets are inter-related and must be kept in sync
2407 #if !FEATURE_EH_FUNCLETS
2408 // This is used for the callable handlers
2409 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2410 #endif // FEATURE_EH_FUNCLETS
2412 unsigned lvaCachedGenericContextArgOffs;
2413 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2416 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2418 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2420 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2421 // after the reg predict we will use a computed maxTmpSize
2422 // which is based upon the number of spill temps predicted by reg predict
2423 // All this is necessary because if we under-estimate the size of the spill
2424 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2426 // Pre codegen max spill temp size.
2427 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2429 //-------------------------------------------------------------------------
2431 unsigned lvaGetMaxSpillTempSize();
2433 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2434 #endif // _TARGET_ARM_
2435 void lvaAssignFrameOffsets(FrameLayoutState curState);
2436 void lvaFixVirtualFrameOffsets();
2438 #ifndef LEGACY_BACKEND
2439 void lvaUpdateArgsWithInitialReg();
2440 #endif // !LEGACY_BACKEND
2442 void lvaAssignVirtualFrameOffsetsToArgs();
2443 #ifdef UNIX_AMD64_ABI
2444 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2445 #else // !UNIX_AMD64_ABI
2446 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2447 #endif // !UNIX_AMD64_ABI
2448 void lvaAssignVirtualFrameOffsetsToLocals();
2449 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2450 #ifdef _TARGET_AMD64_
2451 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2452 bool lvaIsCalleeSavedIntRegCountEven();
2454 void lvaAlignFrame();
2455 void lvaAssignFrameOffsetsToPromotedStructs();
2456 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2459 void lvaDumpRegLocation(unsigned lclNum);
2460 void lvaDumpFrameLocation(unsigned lclNum);
2461 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2462 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2463 // layout state defined by lvaDoneFrameLayout
2466 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2467 // to avoid bugs from borderline cases.
2468 #define MAX_FrameSize 0x3FFFFFFF
2469 void lvaIncrementFrameSize(unsigned size);
2471 unsigned lvaFrameSize(FrameLayoutState curState);
2473 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2474 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2476 // Returns the caller-SP-relative offset for the local variable "varNum."
2477 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2479 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2480 int lvaGetSPRelativeOffset(unsigned varNum);
2482 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2483 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2485 //------------------------ For splitting types ----------------------------
2487 void lvaInitTypeRef();
2489 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2490 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2491 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2492 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2493 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2494 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2496 void lvaInitVarDsc(LclVarDsc* varDsc,
2498 CorInfoType corInfoType,
2499 CORINFO_CLASS_HANDLE typeHnd,
2500 CORINFO_ARG_LIST_HANDLE varList,
2501 CORINFO_SIG_INFO* varSig);
2503 static unsigned lvaTypeRefMask(var_types type);
2505 var_types lvaGetActualType(unsigned lclNum);
2506 var_types lvaGetRealType(unsigned lclNum);
2508 //-------------------------------------------------------------------------
2512 unsigned lvaLclSize(unsigned varNum);
2513 unsigned lvaLclExactSize(unsigned varNum);
2515 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2517 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2518 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2519 // the return result.
2520 bool lvaLclVarRefsAccum(
2521 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2523 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2524 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2525 // and (destructively) unions "trkedVars" into "*result".
2526 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2528 ALLVARSET_VALARG_TP allVars,
2529 VARSET_VALARG_TP trkdVars);
2531 bool lvaHaveManyLocals() const;
2533 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2534 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2535 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2538 void lvaSortByRefCount();
2539 void lvaDumpRefCounts();
2541 void lvaMarkLocalVars(BasicBlock* block);
2543 void lvaMarkLocalVars(); // Local variable ref-counting
2545 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2547 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2549 static fgWalkPreFn lvaIncRefCntsCB;
2550 void lvaIncRefCnts(GenTreePtr tree);
2552 static fgWalkPreFn lvaDecRefCntsCB;
2553 void lvaDecRefCnts(GenTreePtr tree);
2554 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2555 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2556 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2559 struct lvaStressLclFldArgs
2561 Compiler* m_pCompiler;
2565 static fgWalkPreFn lvaStressLclFldCB;
2566 void lvaStressLclFld();
2568 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2569 void lvaDispVarSet(VARSET_VALARG_TP set);
2574 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2576 int lvaFrameAddress(int varNum, bool* pFPbased);
2579 bool lvaIsParameter(unsigned varNum);
2580 bool lvaIsRegArgument(unsigned varNum);
2581 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2582 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2583 // that writes to arg0
2585 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2586 // (this is an overload of lvIsTemp because there are no temp parameters).
2587 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2588 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2589 bool lvaIsImplicitByRefLocal(unsigned varNum)
2591 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2592 LclVarDsc* varDsc = &(lvaTable[varNum]);
2593 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2595 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2598 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2602 // Returns true if this local var is a multireg struct
2603 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2605 // If the class is a TYP_STRUCT, get/set a class handle describing it
2607 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2608 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2610 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2612 // Info about struct fields
2613 struct lvaStructFieldInfo
2615 CORINFO_FIELD_HANDLE fldHnd;
2616 unsigned char fldOffset;
2617 unsigned char fldOrdinal;
2620 CORINFO_CLASS_HANDLE fldTypeHnd;
2623 // Info about struct to be promoted.
2624 struct lvaStructPromotionInfo
2626 CORINFO_CLASS_HANDLE typeHnd;
2628 bool requiresScratchVar;
2631 unsigned char fieldCnt;
2632 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2634 lvaStructPromotionInfo()
2635 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2640 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2641 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2642 lvaStructPromotionInfo* StructPromotionInfo,
2644 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2645 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2646 #if !defined(_TARGET_64BIT_)
2647 void lvaPromoteLongVars();
2648 #endif // !defined(_TARGET_64BIT_)
2649 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2650 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2651 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2652 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2653 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2654 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2655 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2657 #if defined(FEATURE_SIMD)
2658 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2660 assert(varDsc->lvType == TYP_SIMD12);
2661 assert(varDsc->lvExactSize == 12);
2663 #if defined(_TARGET_64BIT_)
2664 assert(varDsc->lvSize() == 16);
2666 #else // !defined(_TARGET_64BIT_)
2668 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2669 // already does this calculation. However, we also need to prevent mapping types if the var is a
2670 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2671 // However, we don't know this until late, so we may have already pretended the field is bigger
2673 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2682 #endif // !defined(_TARGET_64BIT_)
2684 #endif // defined(FEATURE_SIMD)
2686 BYTE* lvaGetGcLayout(unsigned varNum);
2687 bool lvaTypeIsGC(unsigned varNum);
2688 unsigned lvaGSSecurityCookie; // LclVar number
2689 bool lvaTempsHaveLargerOffsetThanVars();
2691 unsigned lvaSecurityObject; // variable representing the security object on the stack
2692 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2694 #if FEATURE_EH_FUNCLETS
2695 unsigned lvaPSPSym; // variable representing the PSPSym
2698 InlineInfo* impInlineInfo;
2699 InlineStrategy* m_inlineStrategy;
2701 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2702 Compiler* impInlineRoot();
2704 #if defined(DEBUG) || defined(INLINE_DATA)
2705 unsigned __int64 getInlineCycleCount()
2707 return m_compCycles;
2709 #endif // defined(DEBUG) || defined(INLINE_DATA)
2711 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2712 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2714 //=========================================================================
2716 //=========================================================================
2719 //---------------- Local variable ref-counting ----------------------------
2722 BasicBlock* lvaMarkRefsCurBlock;
2723 GenTreePtr lvaMarkRefsCurStmt;
2725 BasicBlock::weight_t lvaMarkRefsWeight;
2727 static fgWalkPreFn lvaMarkLclRefsCallback;
2728 void lvaMarkLclRefs(GenTreePtr tree);
2730 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2731 PerSsaArray lvHeapPerSsaData;
2732 unsigned lvHeapNumSsaNames;
2735 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2736 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2737 // not an SSA variable).
2738 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2740 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2741 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2743 assert(ssaNum < lvHeapNumSsaNames);
2744 return &lvHeapPerSsaData.GetRef(ssaNum);
2748 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2749 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2753 XX Imports the given method and converts it to semantic trees XX
2755 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2756 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2762 void impImport(BasicBlock* method);
2764 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2765 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2766 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2767 CORINFO_CLASS_HANDLE impGetStringClass();
2768 CORINFO_CLASS_HANDLE impGetObjectClass();
2770 //=========================================================================
2772 //=========================================================================
2775 //-------------------- Stack manipulation ---------------------------------
2777 unsigned impStkSize; // Size of the full stack
2779 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2781 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2783 struct SavedStack // used to save/restore stack contents.
2785 unsigned ssDepth; // number of values on stack
2786 StackEntry* ssTrees; // saved tree values
2789 bool impIsPrimitive(CorInfoType type);
2790 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2792 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2793 void impPushOnStackNoType(GenTreePtr tree);
2795 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2796 void impPushNullObjRefOnStack();
2797 StackEntry impPopStack();
2798 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2799 GenTreePtr impPopStack(typeInfo& ti);
2800 StackEntry& impStackTop(unsigned n = 0);
2802 void impSaveStackState(SavedStack* savePtr, bool copy);
2803 void impRestoreStackState(SavedStack* savePtr);
2805 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2806 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2807 CORINFO_CALL_INFO* pCallInfo);
2809 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2811 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2813 bool impCanPInvokeInline();
2814 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2815 void impCheckForPInvokeCall(
2816 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2817 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2818 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2820 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2821 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2822 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2824 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2825 CORINFO_METHOD_HANDLE calleeMethodHnd,
2826 CORINFO_CLASS_HANDLE delegateTypeHnd);
2828 var_types impImportCall(OPCODE opcode,
2829 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2830 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2832 GenTreePtr newobjThis,
2834 CORINFO_CALL_INFO* callInfo,
2835 IL_OFFSET rawILOffset);
2837 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2839 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2841 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2844 var_types impImportJitTestLabelMark(int numArgs);
2847 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2849 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2851 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2852 CORINFO_ACCESS_FLAGS access,
2853 CORINFO_FIELD_INFO* pFieldInfo,
2856 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2858 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2860 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2862 void impImportLeave(BasicBlock* block);
2863 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2864 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2865 CORINFO_CLASS_HANDLE clsHnd,
2866 CORINFO_METHOD_HANDLE method,
2867 CORINFO_SIG_INFO* sig,
2871 CorInfoIntrinsics* pIntrinsicID);
2872 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2873 CORINFO_SIG_INFO* sig,
2876 CorInfoIntrinsics intrinsicID);
2877 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2879 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2881 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2882 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2883 CORINFO_THIS_TRANSFORM transform);
2885 //----------------- Manipulating the trees and stmts ----------------------
2887 GenTreePtr impTreeList; // Trees for the BB being imported
2888 GenTreePtr impTreeLast; // The last tree for the current BB
2892 CHECK_SPILL_ALL = -1,
2893 CHECK_SPILL_NONE = -2
2897 void impBeginTreeList();
2898 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2899 void impEndTreeList(BasicBlock* block);
2900 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2901 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2902 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2903 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2904 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2905 void impAssignTempGen(unsigned tmp,
2908 GenTreePtr* pAfterStmt = nullptr,
2909 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2910 BasicBlock* block = nullptr);
2911 void impAssignTempGen(unsigned tmpNum,
2913 CORINFO_CLASS_HANDLE structHnd,
2915 GenTreePtr* pAfterStmt = nullptr,
2916 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2917 BasicBlock* block = nullptr);
2918 GenTreePtr impCloneExpr(GenTreePtr tree,
2920 CORINFO_CLASS_HANDLE structHnd,
2922 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2923 GenTreePtr impAssignStruct(GenTreePtr dest,
2925 CORINFO_CLASS_HANDLE structHnd,
2927 GenTreePtr* pAfterStmt = nullptr,
2928 BasicBlock* block = nullptr);
2929 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2931 CORINFO_CLASS_HANDLE structHnd,
2933 GenTreePtr* pAfterStmt = nullptr,
2934 BasicBlock* block = nullptr);
2936 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2937 CORINFO_CLASS_HANDLE structHnd,
2941 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2942 BYTE* gcLayout = nullptr,
2943 unsigned* numGCVars = nullptr,
2944 var_types* simdBaseType = nullptr);
2946 GenTreePtr impNormStructVal(GenTreePtr structVal,
2947 CORINFO_CLASS_HANDLE structHnd,
2949 bool forceNormalization = false);
2951 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2952 BOOL* pRuntimeLookup = nullptr,
2953 BOOL mustRestoreHandle = FALSE,
2954 BOOL importParent = FALSE);
2956 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2957 BOOL* pRuntimeLookup = nullptr,
2958 BOOL mustRestoreHandle = FALSE)
2960 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2963 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2964 CORINFO_LOOKUP* pLookup,
2966 void* compileTimeHandle);
2968 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2970 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2971 CORINFO_LOOKUP* pLookup,
2972 void* compileTimeHandle);
2974 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2976 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2977 CorInfoHelpFunc helper,
2979 GenTreeArgList* arg = nullptr,
2980 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
2982 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
2984 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2987 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
2988 CORINFO_CLASS_HANDLE typeClass,
2992 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
2993 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
2994 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
2995 static bool IsMathIntrinsic(GenTreePtr tree);
2998 //----------------- Importing the method ----------------------------------
3000 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3003 unsigned impCurOpcOffs;
3004 const char* impCurOpcName;
3005 bool impNestedStackSpill;
3007 // For displaying instrs with generated native code (-n:B)
3008 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3009 void impNoteLastILoffs();
3012 /* IL offset of the stmt currently being imported. It gets set to
3013 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3014 updated at IL offsets for which we have to report mapping info.
3015 It also includes flag bits, so use jitGetILoffs()
3016 to get the actual IL offset value.
3019 IL_OFFSETX impCurStmtOffs;
3020 void impCurStmtOffsSet(IL_OFFSET offs);
3022 void impNoteBranchOffs();
3024 unsigned impInitBlockLineInfo();
3026 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3027 bool impIsThis(GenTreePtr obj);
3028 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3029 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3030 bool impIsAnySTLOC(OPCODE opcode)
3032 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3033 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3036 GenTreeArgList* impPopList(unsigned count,
3038 CORINFO_SIG_INFO* sig,
3039 GenTreeArgList* prefixTree = nullptr);
3041 GenTreeArgList* impPopRevList(unsigned count,
3043 CORINFO_SIG_INFO* sig,
3044 unsigned skipReverseCount = 0);
3047 * Get current IL offset with stack-empty info incoporated
3049 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3051 //---------------- Spilling the importer stack ----------------------------
3057 SavedStack pdSavedStack;
3058 ThisInitState pdThisPtrInit;
3061 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3062 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3064 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3065 ExpandArray<BYTE> impPendingBlockMembers;
3067 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3068 // Operates on the map in the top-level ancestor.
3069 BYTE impGetPendingBlockMember(BasicBlock* blk)
3071 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3074 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3075 // Operates on the map in the top-level ancestor.
3076 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3078 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3081 bool impCanReimport;
3083 bool impSpillStackEntry(unsigned level,
3087 bool bAssertOnRecursion,
3092 void impSpillStackEnsure(bool spillLeaves = false);
3093 void impEvalSideEffects();
3094 void impSpillSpecialSideEff();
3095 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3096 void impSpillValueClasses();
3097 void impSpillEvalStack();
3098 static fgWalkPreFn impFindValueClasses;
3099 void impSpillLclRefs(ssize_t lclNum);
3101 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3103 void impImportBlockCode(BasicBlock* block);
3105 void impReimportMarkBlock(BasicBlock* block);
3106 void impReimportMarkSuccessors(BasicBlock* block);
3108 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3110 void impImportBlockPending(BasicBlock* block);
3112 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3113 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3114 // for the block, but instead, just re-uses the block's existing EntryState.
3115 void impReimportBlockPending(BasicBlock* block);
3117 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3119 void impImportBlock(BasicBlock* block);
3121 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3122 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3123 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3124 // the variables that will be used -- and for all the predecessors of those successors, and the
3125 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3126 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3127 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3128 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3129 // of local variable numbers, so we represent them with the base local variable number), returns that.
3130 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3131 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3132 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3133 // on which kind of member of the clique the block is).
3134 unsigned impGetSpillTmpBase(BasicBlock* block);
3136 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3137 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3138 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3139 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3140 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3141 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3142 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3143 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3144 // successors receive a native int. Similarly float and double are unified to double.
3145 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3146 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3147 // predecessors, so they insert an upcast if needed).
3148 void impReimportSpillClique(BasicBlock* block);
3150 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3151 // block, and represent the predecessor and successor members of the clique currently being computed.
3152 // *** Access to these will need to be locked in a parallel compiler.
3153 ExpandArray<BYTE> impSpillCliquePredMembers;
3154 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3162 // Abstract class for receiving a callback while walking a spill clique
3163 class SpillCliqueWalker
3166 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3169 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3170 class SetSpillTempsBase : public SpillCliqueWalker
3175 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3178 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3181 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3182 class ReimportSpillClique : public SpillCliqueWalker
3187 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3190 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3193 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3194 // predecessor or successor within the spill clique
3195 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3197 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3198 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3199 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3200 void impRetypeEntryStateTemps(BasicBlock* blk);
3202 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3203 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3205 void impPushVar(GenTree* op, typeInfo tiRetVal);
3206 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3207 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3209 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3211 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3212 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3213 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3216 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3219 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3220 struct BlockListNode
3223 BlockListNode* m_next;
3224 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3227 void* operator new(size_t sz, Compiler* comp);
3229 BlockListNode* impBlockListNodeFreeList;
3231 BlockListNode* AllocBlockListNode();
3232 void FreeBlockListNode(BlockListNode* node);
3234 bool impIsValueType(typeInfo* pTypeInfo);
3235 var_types mangleVarArgsType(var_types type);
3238 regNumber getCallArgIntRegister(regNumber floatReg);
3239 regNumber getCallArgFloatRegister(regNumber intReg);
3240 #endif // FEATURE_VARARG
3243 static unsigned jitTotalMethodCompiled;
3247 static LONG jitNestingLevel;
3250 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3252 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3254 // STATIC inlining decision based on the IL code.
3255 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3256 CORINFO_METHOD_INFO* methInfo,
3258 InlineResult* inlineResult);
3260 void impCheckCanInline(GenTreePtr call,
3261 CORINFO_METHOD_HANDLE fncHandle,
3263 CORINFO_CONTEXT_HANDLE exactContextHnd,
3264 InlineCandidateInfo** ppInlineCandidateInfo,
3265 InlineResult* inlineResult);
3267 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3268 GenTreePtr curArgVal,
3270 InlineResult* inlineResult);
3272 void impInlineInitVars(InlineInfo* pInlineInfo);
3274 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3276 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3278 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3280 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3281 GenTreePtr variableBeingDereferenced,
3282 InlArgInfo* inlArgInfo);
3284 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3286 bool impTailCallRetTypeCompatible(var_types callerRetType,
3287 CORINFO_CLASS_HANDLE callerRetTypeClass,
3288 var_types calleeRetType,
3289 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3291 bool impIsTailCallILPattern(bool tailPrefixed,
3293 const BYTE* codeAddrOfNextOpcode,
3294 const BYTE* codeEnd,
3296 bool* IsCallPopRet = nullptr);
3298 bool impIsImplicitTailCallCandidate(
3299 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3302 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3303 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3307 XX Info about the basic-blocks, their contents and the flow analysis XX
3309 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3310 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3314 BasicBlock* fgFirstBB; // Beginning of the basic block list
3315 BasicBlock* fgLastBB; // End of the basic block list
3316 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3317 #if FEATURE_EH_FUNCLETS
3318 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3320 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3322 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3323 unsigned fgEdgeCount; // # of control flow edges between the BBs
3324 unsigned fgBBcount; // # of BBs in the method
3326 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3328 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3329 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3330 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3331 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3333 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3334 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3335 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3336 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3337 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3338 // index). The arrays are of size fgBBNumMax + 1.
3339 unsigned* fgDomTreePreOrder;
3340 unsigned* fgDomTreePostOrder;
3342 bool fgBBVarSetsInited;
3344 // Allocate array like T* a = new T[fgBBNumMax + 1];
3345 // Using helper so we don't keep forgetting +1.
3346 template <typename T>
3347 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3349 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3352 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3353 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3354 // cannot be meaningfully combined. Note that new blocks can be created with higher
3355 // block numbers without changing the basic block epoch. These blocks *cannot*
3356 // participate in a block set until the blocks are all renumbered, causing the epoch
3357 // to change. This is useful if continuing to use previous block sets is valuable.
3358 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3359 unsigned fgCurBBEpoch;
3361 unsigned GetCurBasicBlockEpoch()
3363 return fgCurBBEpoch;
3366 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3367 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3368 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3369 unsigned fgCurBBEpochSize;
3371 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3372 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3373 unsigned fgBBSetCountInSizeTUnits;
3375 void NewBasicBlockEpoch()
3377 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3379 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3381 fgCurBBEpochSize = fgBBNumMax + 1;
3382 fgBBSetCountInSizeTUnits =
3383 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3386 // All BlockSet objects are now invalid!
3387 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3388 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3392 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3393 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3394 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3395 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3397 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3398 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3399 // array of size_t bitsets), then print that out.
3400 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3407 void EnsureBasicBlockEpoch()
3409 if (fgCurBBEpochSize != fgBBNumMax + 1)
3411 NewBasicBlockEpoch();
3415 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3416 void fgEnsureFirstBBisScratch();
3417 bool fgFirstBBisScratch();
3418 bool fgBBisScratch(BasicBlock* block);
3420 void fgExtendEHRegionBefore(BasicBlock* block);
3421 void fgExtendEHRegionAfter(BasicBlock* block);
3423 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3425 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3427 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3430 BasicBlock* nearBlk,
3431 bool putInFilter = false,
3432 bool runRarely = false,
3433 bool insertAtEnd = false);
3435 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3437 bool runRarely = false,
3438 bool insertAtEnd = false);
3440 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3442 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3443 BasicBlock* afterBlk,
3444 unsigned xcptnIndex,
3445 bool putInTryRegion);
3447 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3448 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3449 void fgUnlinkBlock(BasicBlock* block);
3451 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3452 bool fgMultipleNots;
3455 bool fgModified; // True if the flow graph has been modified recently
3456 bool fgComputePredsDone; // Have we computed the bbPreds list
3457 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3458 bool fgDomsComputed; // Have we computed the dominator sets?
3459 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3461 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3462 bool fgHasPostfix; // any postfix ++/-- found?
3463 unsigned fgIncrCount; // number of increment nodes found
3465 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3469 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3470 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3473 bool fgRemoveRestOfBlock; // true if we know that we will throw
3474 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3476 // There are two modes for ordering of the trees.
3477 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3478 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3479 // by traversing the tree according to the order of the operands.
3480 // - In FGOrderLinear, the dominant ordering is the linear order.
3487 FlowGraphOrder fgOrder;
3489 // The following are boolean flags that keep track of the state of internal data structures
3491 bool fgStmtListThreaded;
3492 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3493 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3494 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3495 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3496 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3497 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3498 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3499 // This is derived from the profile data
3500 // or is BB_UNITY_WEIGHT when we don't have profile data
3502 #if FEATURE_EH_FUNCLETS
3503 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3504 #endif // FEATURE_EH_FUNCLETS
3506 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3507 // since fgMorphTree can be called from several places
3508 bool fgExpandInline; // indicates that we are creating tree for the inliner
3510 bool impBoxTempInUse; // the temp below is valid and available
3511 unsigned impBoxTemp; // a temporary that is used for boxing
3514 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3515 // and we are trying to compile again in a "safer", minopts mode?
3519 unsigned impInlinedCodeSize;
3522 //-------------------------------------------------------------------------
3530 void fgRemoveEmptyTry();
3532 void fgRemoveEmptyFinally();
3534 void fgCloneFinally();
3536 void fgCleanupContinuation(BasicBlock* continuation);
3538 void fgUpdateFinallyTargetFlags();
3540 GenTreePtr fgGetCritSectOfStaticMethod();
3542 #if !defined(_TARGET_X86_)
3544 void fgAddSyncMethodEnterExit();
3546 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3548 void fgConvertSyncReturnToLeave(BasicBlock* block);
3550 #endif // !_TARGET_X86_
3552 void fgAddReversePInvokeEnterExit();
3554 bool fgMoreThanOneReturnBlock();
3556 // The number of separate return points in the method.
3557 unsigned fgReturnCount;
3559 void fgAddInternal();
3561 bool fgFoldConditional(BasicBlock* block);
3563 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3564 void fgMorphBlocks();
3566 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3568 void fgCheckArgCnt();
3569 void fgSetOptions();
3572 static fgWalkPreFn fgAssertNoQmark;
3573 void fgPreExpandQmarkChecks(GenTreePtr expr);
3574 void fgPostExpandQmarkChecks();
3575 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3578 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3580 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3581 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3582 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3583 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3584 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3586 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3587 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3588 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3589 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3591 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3592 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3593 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3594 void fgExpandQmarkNodes();
3598 // Do "simple lowering." This functionality is (conceptually) part of "general"
3599 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3600 void fgSimpleLowering();
3602 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3604 GenTreePtr fgInitThisClass();
3606 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3608 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3610 void fgLocalVarLiveness();
3612 void fgLocalVarLivenessInit();
3614 #ifdef LEGACY_BACKEND
3615 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3617 void fgPerNodeLocalVarLiveness(GenTree* node);
3619 void fgPerBlockLocalVarLiveness();
3621 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3623 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3625 // This is used in the liveness computation, as a temporary. When we use the
3626 // arbitrary-length VarSet representation, it is better not to allocate a new one
3628 VARSET_TP fgMarkIntfUnionVS;
3630 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3632 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3634 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3636 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3638 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3640 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3642 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3643 GenTreePtr startNode,
3645 VARSET_VALARG_TP volatileVars,
3646 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3648 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3650 bool fgRemoveDeadStore(GenTree** pTree,
3654 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3656 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3658 // For updating liveset during traversal AFTER fgComputeLife has completed
3659 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3660 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3662 // Returns the set of live variables after endTree,
3663 // assuming that liveSet is the set of live variables BEFORE tree.
3664 // Requires that fgComputeLife has completed, and that tree is in the same
3665 // statement as endTree, and that it comes before endTree in execution order
3667 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3669 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3670 while (tree != nullptr && tree != endTree->gtNext)
3672 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3673 tree = tree->gtNext;
3675 assert(tree == endTree->gtNext);
3679 void fgInterBlockLocalVarLiveness();
3681 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3682 // "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
3683 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3684 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3685 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3686 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3687 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3689 if (m_opAsgnVarDefSsaNums == nullptr)
3691 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3693 return m_opAsgnVarDefSsaNums;
3696 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3697 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3698 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3700 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3702 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3703 // Except: assumes that lcl is a def, and if it is
3704 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3705 // rather than the "use" SSA number recorded in the tree "lcl".
3706 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3708 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3709 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3710 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3711 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3712 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3714 // (byref addrS1 = &s1,
3715 // *(addrS1 * offsetof(f0)) = s2f0,
3717 // *(addrS1 * offsetof(fn)) = s2fn)
3719 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3720 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3721 // give it SSA names and value numbers?
3723 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3724 // end with an instance of the structure below, whose fields are described in the declaration.
3725 struct IndirectAssignmentAnnotation
3727 unsigned m_lclNum; // The local num that is being indirectly assigned.
3728 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3729 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3730 // be the singleton field sequence "g". The individual assignments would
3731 // further append the fields of "s.g" to that.
3732 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3733 // structure has a single field).
3734 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3735 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3738 IndirectAssignmentAnnotation(unsigned lclNum,
3739 FieldSeqNode* fldSeq,
3741 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3742 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3743 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3747 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3748 NodeToIndirAssignMap;
3749 NodeToIndirAssignMap* m_indirAssignMap;
3750 NodeToIndirAssignMap* GetIndirAssignMap()
3752 if (m_indirAssignMap == nullptr)
3754 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3755 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3756 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3758 return m_indirAssignMap;
3761 // Performs SSA conversion.
3764 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3765 void fgResetForSsa();
3767 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3769 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3770 inline bool fgExcludeFromSsa(unsigned lclNum);
3772 // The value numbers for this compilation.
3773 ValueNumStore* vnStore;
3776 ValueNumStore* GetValueNumStore()
3781 // Do value numbering (assign a value number to each
3783 void fgValueNumber();
3785 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3786 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3787 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3788 // match the element type of the array or fldSeq. When this type doesn't match
3789 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3791 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3794 FieldSeqNode* fldSeq,
3798 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3799 // has been parsed to yield the other input arguments. If evaluation of the address
3800 // can raise exceptions, those should be captured in the exception set "excVN."
3801 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3802 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3803 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3804 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3805 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3807 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3808 CORINFO_CLASS_HANDLE elemTypeEq,
3812 FieldSeqNode* fldSeq);
3814 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3815 // by evaluating the array index expression "tree". Returns the value number resulting from
3816 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3817 // "GT_IND" that does the dereference, and it is given the returned value number.
3818 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3820 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3822 // Utility functions for fgValueNumber.
3824 // Perform value-numbering for the trees in "blk".
3825 void fgValueNumberBlock(BasicBlock* blk);
3827 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3828 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3829 // assumed for the heap at the start "entryBlk".
3830 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3832 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3833 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3835 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3836 // value in that SSA #.
3837 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3839 // The input 'tree' is a leaf node that is a constant
3840 // Assign the proper value number to the tree
3841 void fgValueNumberTreeConst(GenTreePtr tree);
3843 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3844 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3846 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3848 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3850 // Does value-numbering for a block assignment.
3851 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3853 // Does value-numbering for a cast tree.
3854 void fgValueNumberCastTree(GenTreePtr tree);
3856 // Does value-numbering for an intrinsic tree.
3857 void fgValueNumberIntrinsic(GenTreePtr tree);
3859 // Does value-numbering for a call. We interpret some helper calls.
3860 void fgValueNumberCall(GenTreeCall* call);
3862 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3863 void fgUpdateArgListVNs(GenTreeArgList* args);
3865 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3866 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3868 // Requires "helpCall" to be a helper call. Assigns it a value number;
3869 // we understand the semantics of some of the calls. Returns "true" if
3870 // the call may modify the heap (we assume arbitrary memory side effects if so).
3871 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3873 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3874 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3876 // This is the current value number for the "Heap" implicit variable while
3877 // doing value numbering. This is the value number under the "liberal" interpretation
3878 // of heap values; the "conservative" interpretation needs no VN, since every access of
3879 // the heap yields an unknown value.
3880 ValueNum fgCurHeapVN;
3882 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3883 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3884 // is 1, and the rest is an encoding of "elemTyp".
3885 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3887 if (elemStructType != nullptr)
3889 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3890 varTypeIsIntegral(elemTyp));
3891 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3892 return elemStructType;
3896 elemTyp = varTypeUnsignedToSigned(elemTyp);
3897 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3900 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3901 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3902 // the struct type of the element).
3903 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3905 size_t clsHndVal = size_t(clsHnd);
3906 if (clsHndVal & 0x1)
3908 return var_types(clsHndVal >> 1);
3916 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3917 var_types getJitGCType(BYTE gcType);
3919 enum structPassingKind
3921 SPK_Unknown, // Invalid value, never returned
3922 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3923 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3924 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3925 // parameters registers are used, then the stack will be used)
3926 // for X86 passed on the stack, for ARM32 passed in registers
3927 // or the stack or split between registers and the stack.
3928 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3930 }; // The struct is passed/returned by reference to a copy/buffer.
3932 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3933 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3934 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3935 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3937 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3939 // Get the type that is used to pass values of the given struct type.
3940 // If you have already retrieved the struct size then pass it as the optional third argument
3942 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3943 structPassingKind* wbPassStruct,
3944 unsigned structSize = 0);
3946 // Get the type that is used to return values of the given struct type.
3947 // If you have already retrieved the struct size then pass it as the optional third argument
3949 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3950 structPassingKind* wbPassStruct = nullptr,
3951 unsigned structSize = 0);
3954 // Print a representation of "vnp" or "vn" on standard output.
3955 // If "level" is non-zero, we also print out a partial expansion of the value.
3956 void vnpPrint(ValueNumPair vnp, unsigned level);
3957 void vnPrint(ValueNum vn, unsigned level);
3960 // Dominator computation member functions
3961 // Not exposed outside Compiler
3963 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3965 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3967 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3968 // flow graph. We first assume the fields bbIDom on each
3969 // basic block are invalid. This computation is needed later
3970 // by fgBuildDomTree to build the dominance tree structure.
3971 // Based on: A Simple, Fast Dominance Algorithm
3972 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3974 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3975 // Note: this is relatively slow compared to calling fgDominate(),
3976 // especially if dealing with a single block versus block check.
3978 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3980 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3982 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3984 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3986 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3988 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3989 // processed in topological sort, this function takes care of that.
3991 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3993 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3994 // Returns this as a set.
3996 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3997 // root nodes. Returns this as a set.
4000 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4003 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4004 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4007 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4008 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4009 // && postOrder(A) >= postOrder(B) making the computation O(1).
4010 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4012 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4014 void fgUpdateChangedFlowGraph();
4017 // Compute the predecessors of the blocks in the control flow graph.
4018 void fgComputePreds();
4020 // Remove all predecessor information.
4021 void fgRemovePreds();
4023 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4024 // before the full predecessors lists are computed.
4025 void fgComputeCheapPreds();
4028 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4030 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4040 // Initialize the per-block variable sets (used for liveness analysis).
4041 void fgInitBlockVarSets();
4043 // true if we've gone through and created GC Poll calls.
4044 bool fgGCPollsCreated;
4045 void fgMarkGCPollBlocks();
4046 void fgCreateGCPolls();
4047 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4049 // Requires that "block" is a block that returns from
4050 // a finally. Returns the number of successors (jump targets of
4051 // of blocks in the covered "try" that did a "LEAVE".)
4052 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4054 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4055 // a finally. Returns its "i"th successor (jump targets of
4056 // of blocks in the covered "try" that did a "LEAVE".)
4057 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4058 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4061 // Factor out common portions of the impls of the methods above.
4062 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4065 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4066 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4067 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4068 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4069 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4070 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4071 // we leave the entry associated with the block, but it will no longer be accessed.)
4072 struct SwitchUniqueSuccSet
4074 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4075 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4078 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4079 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4080 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4081 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4084 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4085 BlockToSwitchDescMap;
4088 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4089 // iteration over only the distinct successors.
4090 BlockToSwitchDescMap* m_switchDescMap;
4093 BlockToSwitchDescMap* GetSwitchDescMap()
4095 if (m_switchDescMap == nullptr)
4097 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4099 return m_switchDescMap;
4102 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4103 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4104 // we don't accidentally look up and return the wrong switch data.
4105 void InvalidateUniqueSwitchSuccMap()
4107 m_switchDescMap = nullptr;
4110 // Requires "switchBlock" to be a block that ends in a switch. Returns
4111 // the corresponding SwitchUniqueSuccSet.
4112 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4114 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4115 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4116 // remove it from "this", and ensure that "to" is a member.
4117 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4119 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4120 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4122 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4124 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4126 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4128 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4130 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4132 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4134 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4136 void fgRemoveBlockAsPred(BasicBlock* block);
4138 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4140 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4142 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4144 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4146 flowList* fgAddRefPred(BasicBlock* block,
4147 BasicBlock* blockPred,
4148 flowList* oldEdge = nullptr,
4149 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4152 void fgFindBasicBlocks();
4154 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4156 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4158 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4159 bool putInTryRegion,
4160 BasicBlock* startBlk,
4162 BasicBlock* nearBlk,
4163 BasicBlock* jumpBlk,
4166 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4168 void fgRemoveEmptyBlocks();
4170 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4172 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4174 void fgCreateLoopPreHeader(unsigned lnum);
4176 void fgUnreachableBlock(BasicBlock* block);
4178 void fgRemoveConditionalJump(BasicBlock* block);
4180 BasicBlock* fgLastBBInMainFunction();
4182 BasicBlock* fgEndBBAfterMainFunction();
4184 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4186 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4188 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4190 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4192 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4194 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4196 bool fgRenumberBlocks();
4198 bool fgExpandRarelyRunBlocks();
4200 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4202 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4204 enum FG_RELOCATE_TYPE
4206 FG_RELOCATE_TRY, // relocate the 'try' region
4207 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4209 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4211 #if FEATURE_EH_FUNCLETS
4212 #if defined(_TARGET_ARM_)
4213 void fgClearFinallyTargetBit(BasicBlock* block);
4214 #endif // defined(_TARGET_ARM_)
4215 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4216 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4217 void fgInsertFuncletPrologBlock(BasicBlock* block);
4218 void fgCreateFuncletPrologBlocks();
4219 void fgCreateFunclets();
4220 #else // !FEATURE_EH_FUNCLETS
4221 bool fgRelocateEHRegions();
4222 #endif // !FEATURE_EH_FUNCLETS
4224 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4226 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4228 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4230 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4232 bool fgOptimizeEmptyBlock(BasicBlock* block);
4234 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4236 bool fgOptimizeBranch(BasicBlock* bJump);
4238 bool fgOptimizeSwitchBranches(BasicBlock* block);
4240 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4242 bool fgOptimizeSwitchJumps();
4244 void fgPrintEdgeWeights();
4246 void fgComputeEdgeWeights();
4248 void fgReorderBlocks();
4250 void fgDetermineFirstColdBlock();
4252 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4254 bool fgUpdateFlowGraph(bool doTailDup = false);
4256 void fgFindOperOrder();
4258 // method that returns if you should split here
4259 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4261 void fgSetBlockOrder();
4263 void fgRemoveReturnBlock(BasicBlock* block);
4265 /* Helper code that has been factored out */
4266 inline void fgConvertBBToThrowBB(BasicBlock* block);
4268 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4269 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4270 GenTreePtr fgMakeTmpArgNode(
4271 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4273 // The following check for loops that don't execute calls
4274 bool fgLoopCallMarked;
4276 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4277 void fgLoopCallMark();
4279 void fgMarkLoopHead(BasicBlock* block);
4281 unsigned fgGetCodeEstimate(BasicBlock* block);
4284 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4285 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4286 bool fgDumpFlowGraph(Phases phase);
4288 #endif // DUMP_FLOWGRAPHS
4293 void fgDispBBLiveness(BasicBlock* block);
4294 void fgDispBBLiveness();
4295 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4296 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4297 void fgDispBasicBlocks(bool dumpTrees = false);
4298 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4299 void fgDumpBlock(BasicBlock* block);
4300 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4302 static fgWalkPreFn fgStress64RsltMulCB;
4303 void fgStress64RsltMul();
4304 void fgDebugCheckUpdate();
4305 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4306 void fgDebugCheckBlockLinks();
4307 void fgDebugCheckLinks(bool morphTrees = false);
4308 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4309 void fgDebugCheckFlags(GenTreePtr tree);
4310 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4311 void fgDebugCheckTryFinallyExits();
4314 #ifdef LEGACY_BACKEND
4315 static void fgOrderBlockOps(GenTreePtr tree,
4319 GenTreePtr* opsPtr, // OUT
4320 regMaskTP* regsPtr); // OUT
4321 #endif // LEGACY_BACKEND
4323 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4324 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4326 inline bool fgIsInlining()
4328 return fgExpandInline;
4331 void fgTraverseRPO();
4333 //--------------------- Walking the trees in the IR -----------------------
4338 fgWalkPreFn* wtprVisitorFn;
4339 fgWalkPostFn* wtpoVisitorFn;
4340 void* pCallbackData; // user-provided data
4341 bool wtprLclsOnly; // whether to only visit lclvar nodes
4342 GenTreePtr parent; // parent of current node, provided to callback
4343 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4345 bool printModified; // callback can use this
4349 template <bool computeStack>
4350 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4352 // general purpose tree-walker that is capable of doing pre- and post- order
4353 // callbacks at the same time
4354 template <bool doPreOrder, bool doPostOrder>
4355 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4357 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4358 fgWalkPreFn* visitor,
4359 void* pCallBackData = nullptr,
4360 bool lclVarsOnly = false,
4361 bool computeStack = false);
4363 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4364 fgWalkPreFn* preVisitor,
4365 fgWalkPostFn* postVisitor,
4366 void* pCallBackData = nullptr);
4368 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4372 template <bool computeStack>
4373 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4375 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4376 fgWalkPostFn* visitor,
4377 void* pCallBackData = nullptr,
4378 bool computeStack = false);
4380 // An fgWalkPreFn that looks for expressions that have inline throws in
4381 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4382 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4383 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4384 // properly propagated to parent trees). It returns WALK_CONTINUE
4386 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4387 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4388 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4390 /**************************************************************************
4392 *************************************************************************/
4395 friend class SsaBuilder;
4396 friend struct ValueNumberState;
4398 //--------------------- Detect the basic blocks ---------------------------
4400 BasicBlock** fgBBs; // Table of pointers to the BBs
4402 void fgInitBBLookup();
4403 BasicBlock* fgLookupBB(unsigned addr);
4405 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4407 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4409 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4411 void fgLinkBasicBlocks();
4413 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4415 void fgCheckBasicBlockControlFlow();
4417 void fgControlFlowPermitted(BasicBlock* blkSrc,
4418 BasicBlock* blkDest,
4419 BOOL IsLeave = false /* is the src a leave block */);
4421 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4423 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4425 void fgAdjustForAddressExposedOrWrittenThis();
4427 bool fgProfileData_ILSizeMismatch;
4428 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4429 ULONG fgProfileBufferCount;
4430 ULONG fgNumProfileRuns;
4432 unsigned fgStressBBProf()
4435 unsigned result = JitConfig.JitStressBBProf();
4438 if (compStressCompile(STRESS_BB_PROFILE, 15))
4449 bool fgHaveProfileData();
4450 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4452 bool fgIsUsingProfileWeights()
4454 return (fgHaveProfileData() || fgStressBBProf());
4456 void fgInstrumentMethod();
4458 //-------- Insert a statement at the start or end of a basic block --------
4462 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4466 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4468 public: // Used by linear scan register allocation
4469 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4472 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4473 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4475 public: // Used by linear scan register allocation
4476 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4479 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4481 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4483 // Create a new temporary variable to hold the result of *ppTree,
4484 // and transform the graph accordingly.
4485 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4486 GenTree* fgMakeMultiUse(GenTree** ppTree);
4489 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4490 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4491 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4493 //-------- Determine the order in which the trees will be evaluated -------
4495 unsigned fgTreeSeqNum;
4496 GenTree* fgTreeSeqLst;
4497 GenTree* fgTreeSeqBeg;
4499 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4500 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4501 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4502 void fgSetStmtSeq(GenTree* tree);
4503 void fgSetBlockOrder(BasicBlock* block);
4505 //------------------------- Morphing --------------------------------------
4507 unsigned fgPtrArgCntCur;
4508 unsigned fgPtrArgCntMax;
4509 hashBv* fgOutgoingArgTemps;
4510 hashBv* fgCurrentlyInUseArgTemps;
4512 bool compCanEncodePtrArgCntMax();
4514 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4517 void fgMoveOpsLeft(GenTreePtr tree);
4520 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4522 bool fgIsThrow(GenTreePtr tree);
4524 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4525 bool fgIsBlockCold(BasicBlock* block);
4527 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4529 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4531 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4533 bool fgMorphRelopToQmark(GenTreePtr tree);
4535 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4536 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4537 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4538 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4539 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4540 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4541 // small; hence the other fields of MorphAddrContext.
4542 enum MorphAddrContextKind
4547 struct MorphAddrContext
4549 MorphAddrContextKind m_kind;
4550 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4551 // top-level indirection and here have been constants.
4552 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4553 // In that case, is the sum of those constant offsets.
4555 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4560 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4561 static MorphAddrContext s_CopyBlockMAC;
4564 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4565 var_types* baseTypeOut,
4567 unsigned* simdSizeOut,
4568 bool ignoreUsedInSIMDIntrinsic = false);
4569 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4570 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4571 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4572 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4574 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4575 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4576 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4578 #endif // FEATURE_SIMD
4579 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4580 GenTreePtr fgMorphCast(GenTreePtr tree);
4581 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4582 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4584 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4587 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4588 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4590 void fgFixupStructReturn(GenTreePtr call);
4591 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4592 bool fgAddrCouldBeNull(GenTreePtr addr);
4593 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4594 bool fgCanFastTailCall(GenTreeCall* call);
4595 void fgMorphTailCall(GenTreeCall* call);
4596 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4597 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4598 fgArgTabEntryPtr argTabEntry,
4600 IL_OFFSETX callILOffset,
4601 GenTreePtr tmpAssignmentInsertionPoint,
4602 GenTreePtr paramAssignmentInsertionPoint);
4603 static int fgEstimateCallStackSize(GenTreeCall* call);
4604 GenTreePtr fgMorphCall(GenTreeCall* call);
4605 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4606 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4608 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4609 static fgWalkPreFn fgFindNonInlineCandidate;
4611 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4612 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4613 void fgAssignSetVarDef(GenTreePtr tree);
4614 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4615 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4616 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4617 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4618 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4619 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4620 void fgMorphUnsafeBlk(GenTreeObj* obj);
4621 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4622 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4623 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4624 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4625 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4626 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4627 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4629 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4630 GenTreePtr fgMorphConst(GenTreePtr tree);
4633 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4636 #if LOCAL_ASSERTION_PROP
4637 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4638 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4640 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4642 GenTreeStmt* fgMorphStmt;
4644 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4645 // used when morphing big offset.
4647 //----------------------- Liveness analysis -------------------------------
4649 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4650 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4652 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4653 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4654 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4656 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4658 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4659 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4661 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4662 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4664 void fgExtendDbgScopes();
4665 void fgExtendDbgLifetimes();
4668 void fgDispDebugScopes();
4671 //-------------------------------------------------------------------------
4673 // The following keeps track of any code we've added for things like array
4674 // range checking or explicit calls to enable GC, and so on.
4679 AddCodeDsc* acdNext;
4680 BasicBlock* acdDstBlk; // block to which we jump
4682 SpecialCodeKind acdKind; // what kind of a special block is this?
4683 unsigned short acdStkLvl;
4687 static unsigned acdHelper(SpecialCodeKind codeKind);
4689 AddCodeDsc* fgAddCodeList;
4691 bool fgRngChkThrowAdded;
4692 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4694 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4696 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4699 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4702 bool fgIsCodeAdded();
4704 bool fgIsThrowHlpBlk(BasicBlock* block);
4705 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4707 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4709 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4710 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4711 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4712 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4713 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4715 #if FEATURE_MULTIREG_RET
4716 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4717 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4718 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4719 #endif // FEATURE_MULTIREG_RET
4721 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4724 static fgWalkPreFn fgDebugCheckInlineCandidates;
4727 void fgPromoteStructs();
4728 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4729 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4730 void fgMarkImplicitByRefArgs();
4731 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4732 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4733 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4734 void fgMarkAddressExposedLocals();
4735 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4737 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4739 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4741 // The given local variable, required to be a struct variable, is being assigned via
4742 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4743 // the variable is not enregistered, and is therefore not promoted independently.
4744 void fgLclFldAssign(unsigned lclNum);
4746 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4747 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4748 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4749 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4752 bool fgPrintInlinedMethods;
4755 bool fgIsBigOffset(size_t offset);
4757 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4758 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4759 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4760 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4761 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4764 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4765 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4769 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4770 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4777 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4780 void optRemoveRangeCheck(
4781 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4782 bool optIsRangeCheckRemovable(GenTreePtr tree);
4785 static fgWalkPreFn optValidRangeCheckIndex;
4786 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4789 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4791 /**************************************************************************
4793 *************************************************************************/
4796 // Do hoisting for all loops.
4797 void optHoistLoopCode();
4799 // To represent sets of VN's that have already been hoisted in outer loops.
4800 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4801 typedef VNToBoolMap VNSet;
4803 struct LoopHoistContext
4806 // The set of variables hoisted in the current loop (or nullptr if there are none).
4807 VNSet* m_pHoistedInCurLoop;
4810 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4811 VNSet m_hoistedInParentLoops;
4812 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4813 // Previous decisions on loop-invariance of value numbers in the current loop.
4814 VNToBoolMap m_curLoopVnInvariantCache;
4816 VNSet* GetHoistedInCurLoop(Compiler* comp)
4818 if (m_pHoistedInCurLoop == nullptr)
4820 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4822 return m_pHoistedInCurLoop;
4825 VNSet* ExtractHoistedInCurLoop()
4827 VNSet* res = m_pHoistedInCurLoop;
4828 m_pHoistedInCurLoop = nullptr;
4832 LoopHoistContext(Compiler* comp)
4833 : m_pHoistedInCurLoop(nullptr)
4834 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4835 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4840 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4841 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4842 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4843 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4845 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4846 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4847 // "m_hoistedInParentLoops".
4849 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4851 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4852 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4853 // expressions to "hoistInLoop".
4854 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4856 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4857 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4859 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4860 // that are invariant in loop "lnum" (an index into the optLoopTable)
4861 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4862 // expressions to "hoistInLoop".
4863 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4864 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4865 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4866 bool optHoistLoopExprsForTree(GenTreePtr tree,
4868 LoopHoistContext* hoistCtxt,
4869 bool* firstBlockAndBeforeSideEffect,
4872 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4873 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4875 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4876 // Constants and init values are always loop invariant.
4877 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4878 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4880 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4881 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4882 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4883 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4884 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4886 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4887 // in the loop table.
4888 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4890 // Records the set of "side effects" of all loops: fields (object instance and static)
4891 // written to, and SZ-array element type equivalence classes updated.
4892 void optComputeLoopSideEffects();
4895 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4896 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4897 // static) written to, and SZ-array element type equivalence classes updated.
4898 void optComputeLoopNestSideEffects(unsigned lnum);
4900 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4901 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4903 // Hoist the expression "expr" out of loop "lnum".
4904 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4907 void optOptimizeBools();
4910 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4912 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4915 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4917 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4918 // the loop into a "do-while" loop
4919 // Also finds all natural loops and records them in the loop table
4921 // Optionally clone loops in the loop table.
4922 void optCloneLoops();
4924 // Clone loop "loopInd" in the loop table.
4925 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4927 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4928 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4929 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4931 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4933 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4936 // This enumeration describes what is killed by a call.
4940 CALLINT_NONE, // no interference (most helpers)
4941 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4942 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4943 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4944 CALLINT_ALL, // kills everything (normal method call)
4948 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4949 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4950 // in bbNext order; we use comparisons on the bbNum to decide order.)
4951 // The blocks that define the body are
4952 // first <= top <= entry <= bottom .
4953 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4954 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4955 // Compiler::optFindNaturalLoops().
4958 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4959 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4960 // loop, but not the outer loop.)
4961 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4963 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4964 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4965 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4967 callInterf lpAsgCall; // "callInterf" for calls in the loop
4968 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4969 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4971 unsigned short lpFlags; // Mask of the LPFLG_* constants
4973 unsigned char lpExitCnt; // number of exits from the loop
4975 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4976 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4977 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4978 // (Actually, an "immediately" nested loop --
4979 // no other child of this loop is a parent of lpChild.)
4980 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4981 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4982 // by following "lpChild" then "lpSibling" links.
4984 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4985 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4987 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4988 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4989 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4991 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4992 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4994 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4995 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4996 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4997 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4999 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5000 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5001 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5003 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5004 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5005 // type are assigned to.
5007 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
5008 // If this is set, the fields below may not be accurate (since they become irrelevant.)
5009 bool lpContainsCall; // True if executing the loop body *may* execute a call
5011 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5012 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5014 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5016 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5017 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5019 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5021 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5022 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5024 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5025 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5027 JitSimplerHashBehavior>
5029 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5030 // instance fields modified
5033 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5034 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5036 JitSimplerHashBehavior>
5038 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5039 // arrays of that type are modified
5042 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5043 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5045 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5046 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5047 // (shifted left, with a low-order bit set to distinguish.)
5048 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5049 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5051 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5053 GenTreePtr lpIterTree; // The "i <op>= const" tree
5054 unsigned lpIterVar(); // iterator variable #
5055 int lpIterConst(); // the constant with which the iterator is incremented
5056 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5057 void VERIFY_lpIterTree();
5059 var_types lpIterOperType(); // For overflow instructions
5062 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5063 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5067 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5069 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5070 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5071 void VERIFY_lpTestTree();
5073 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5074 GenTreePtr lpIterator(); // the iterator node in the loop test
5075 GenTreePtr lpLimit(); // the limit node in the loop test
5077 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5078 // LPFLG_CONST_LIMIT
5079 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5081 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5082 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5083 // LPFLG_ARRLEN_LIMIT
5085 // Returns "true" iff "*this" contains the blk.
5086 bool lpContains(BasicBlock* blk)
5088 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5090 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5091 // to be equal, but requiring bottoms to be different.)
5092 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5094 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5097 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5098 // bottoms to be different.)
5099 bool lpContains(const LoopDsc& lp2)
5101 return lpContains(lp2.lpFirst, lp2.lpBottom);
5104 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5105 // (allowing firsts to be equal, but requiring bottoms to be different.)
5106 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5108 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5111 // Returns "true" iff "*this" is (properly) contained by "lp2"
5112 // (allowing firsts to be equal, but requiring bottoms to be different.)
5113 bool lpContainedBy(const LoopDsc& lp2)
5115 return lpContains(lp2.lpFirst, lp2.lpBottom);
5118 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5119 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5121 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5123 // Returns "true" iff "*this" is disjoint from "lp2".
5124 bool lpDisjoint(const LoopDsc& lp2)
5126 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5128 // Returns "true" iff the loop is well-formed (see code for defn).
5131 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5132 lpEntry->bbNum <= lpBottom->bbNum &&
5133 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5138 bool fgMightHaveLoop(); // returns true if there are any backedges
5139 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5142 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5143 unsigned char optLoopCount; // number of tracked loops
5146 unsigned optCallCount; // number of calls made in the method
5147 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5148 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5149 unsigned optLoopsCloned; // number of loops cloned in the current method.
5152 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5153 void optPrintLoopInfo(unsigned loopNum,
5155 BasicBlock* lpFirst,
5157 BasicBlock* lpEntry,
5158 BasicBlock* lpBottom,
5159 unsigned char lpExitCnt,
5161 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5162 void optPrintLoopInfo(unsigned lnum);
5163 void optPrintLoopRecording(unsigned lnum);
5165 void optCheckPreds();
5168 void optSetBlockWeights();
5170 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5172 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5174 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5176 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5177 unsigned optIsLoopIncrTree(GenTreePtr incr);
5178 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5179 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5180 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5181 bool optExtractInitTestIncr(BasicBlock* head,
5186 GenTreePtr* ppIncr);
5188 void optRecordLoop(BasicBlock* head,
5194 unsigned char exitCnt);
5196 void optFindNaturalLoops();
5198 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5199 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5200 bool optCanonicalizeLoopNest(unsigned char loopInd);
5202 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5203 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5204 bool optCanonicalizeLoop(unsigned char loopInd);
5206 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5207 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5208 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5209 bool optLoopContains(unsigned l1, unsigned l2);
5211 // Requires "loopInd" to be a valid index into the loop table.
5212 // Updates the loop table by changing loop "loopInd", whose head is required
5213 // to be "from", to be "to". Also performs this transformation for any
5214 // loop nested in "loopInd" that shares the same head as "loopInd".
5215 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5217 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5218 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5219 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5221 // Marks the containsCall information to "lnum" and any parent loops.
5222 void AddContainsCallAllContainingLoops(unsigned lnum);
5223 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5224 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5225 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5226 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5227 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5228 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5230 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5231 // of "from".) Copies the jump destination from "from" to "to".
5232 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5234 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5235 unsigned optLoopDepth(unsigned lnum)
5237 unsigned par = optLoopTable[lnum].lpParent;
5238 if (par == BasicBlock::NOT_IN_LOOP)
5244 return 1 + optLoopDepth(par);
5248 void fgOptWhileLoop(BasicBlock* block);
5250 bool optComputeLoopRep(int constInit,
5253 genTreeOps iterOper,
5255 genTreeOps testOper,
5258 unsigned* iterCount);
5259 #if FEATURE_STACK_FP_X87
5262 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5263 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5264 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5265 #endif // FEATURE_STACK_FP_X87
5268 static fgWalkPreFn optIsVarAssgCB;
5271 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5273 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5275 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5277 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5279 /**************************************************************************
5280 * Optimization conditions
5281 *************************************************************************/
5283 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5284 bool optPentium4(void);
5285 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5286 bool optAvoidIntMult(void);
5291 // The following is the upper limit on how many expressions we'll keep track
5292 // of for the CSE analysis.
5294 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5296 static const int MIN_CSE_COST = 2;
5298 // Keeps tracked cse indices
5299 BitVecTraits* cseTraits;
5303 /* Generic list of nodes - used by the CSE logic */
5311 typedef struct treeLst* treeLstPtr;
5315 treeStmtLst* tslNext;
5316 GenTreePtr tslTree; // tree node
5317 GenTreePtr tslStmt; // statement containing the tree
5318 BasicBlock* tslBlock; // block containing the statement
5321 typedef struct treeStmtLst* treeStmtLstPtr;
5323 // The following logic keeps track of expressions via a simple hash table.
5327 CSEdsc* csdNextInBucket; // used by the hash table
5329 unsigned csdHashValue; // the orginal hashkey
5331 unsigned csdIndex; // 1..optCSECandidateCount
5332 char csdLiveAcrossCall; // 0 or 1
5334 unsigned short csdDefCount; // definition count
5335 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5337 unsigned csdDefWtCnt; // weighted def count
5338 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5340 GenTreePtr csdTree; // treenode containing the 1st occurance
5341 GenTreePtr csdStmt; // stmt containing the 1st occurance
5342 BasicBlock* csdBlock; // block containing the 1st occurance
5344 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5345 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5348 static const size_t s_optCSEhashSize;
5349 CSEdsc** optCSEhash;
5354 CSEdsc* optCSEfindDsc(unsigned index);
5355 void optUnmarkCSE(GenTreePtr tree);
5357 // user defined callback data for the tree walk function optCSE_MaskHelper()
5358 struct optCSE_MaskData
5360 EXPSET_TP CSE_defMask;
5361 EXPSET_TP CSE_useMask;
5364 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5365 static fgWalkPreFn optCSE_MaskHelper;
5367 // This function walks all the node for an given tree
5368 // and return the mask of CSE definitions and uses for the tree
5370 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5372 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5373 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5374 bool optCSE_canSwap(GenTree* tree);
5376 static fgWalkPostFn optPropagateNonCSE;
5377 static fgWalkPreFn optHasNonCSEChild;
5379 static fgWalkPreFn optUnmarkCSEs;
5381 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5382 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5384 void optCleanupCSEs();
5387 void optEnsureClearCSEInfo();
5390 #endif // FEATURE_ANYCSE
5392 #if FEATURE_VALNUM_CSE
5393 /**************************************************************************
5394 * Value Number based CSEs
5395 *************************************************************************/
5398 void optOptimizeValnumCSEs();
5401 void optValnumCSE_Init();
5402 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5403 unsigned optValnumCSE_Locate();
5404 void optValnumCSE_InitDataFlow();
5405 void optValnumCSE_DataFlow();
5406 void optValnumCSE_Availablity();
5407 void optValnumCSE_Heuristic();
5408 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5410 #endif // FEATURE_VALNUM_CSE
5413 bool optDoCSE; // True when we have found a duplicate CSE tree
5414 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5415 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5416 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5417 unsigned optCSEstart; // The first local variable number that is a CSE
5418 unsigned optCSEcount; // The total count of CSE's introduced.
5419 unsigned optCSEweight; // The weight of the current block when we are
5420 // scanning for CSE expressions
5422 bool optIsCSEcandidate(GenTreePtr tree);
5424 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5426 bool lclNumIsTrueCSE(unsigned lclNum) const
5428 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5431 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5433 bool lclNumIsCSE(unsigned lclNum) const
5435 return lvaTable[lclNum].lvIsCSE;
5439 bool optConfigDisableCSE();
5440 bool optConfigDisableCSE2();
5442 void optOptimizeCSEs();
5444 #endif // FEATURE_ANYCSE
5452 unsigned ivaVar; // Variable we are interested in, or -1
5453 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5454 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5455 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5456 callInterf ivaMaskCall; // What kind of calls are there?
5459 static callInterf optCallInterf(GenTreePtr call);
5462 // VN based copy propagation.
5463 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5464 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5465 LclNumToGenTreePtrStack;
5467 // Kill set to track variables with intervening definitions.
5468 VARSET_TP optCopyPropKillSet;
5470 // Copy propagation functions.
5471 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5472 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5473 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5474 bool optIsSsaLocal(GenTreePtr tree);
5475 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5476 void optVnCopyProp();
5478 /**************************************************************************
5479 * Early value propagation
5480 *************************************************************************/
5486 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5490 static unsigned GetHashCode(SSAName ssaNm)
5492 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5495 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5497 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5501 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5502 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5503 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5504 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5505 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5507 unsigned optMethodFlags;
5509 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5510 // No throughput diff was found with backward walk bound between 3-8.
5511 static const int optEarlyPropRecurBound = 5;
5513 enum class optPropKind
5521 bool gtIsVtableRef(GenTreePtr tree);
5522 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5523 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5524 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5525 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5526 bool optEarlyPropRewriteTree(GenTreePtr tree);
5527 bool optDoEarlyPropForBlock(BasicBlock* block);
5528 bool optDoEarlyPropForFunc();
5529 void optEarlyProp();
5530 void optFoldNullCheck(GenTreePtr tree);
5531 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5534 /**************************************************************************
5535 * Value/Assertion propagation
5536 *************************************************************************/
5538 // Data structures for assertion prop
5539 BitVecTraits* apTraits;
5543 enum optAssertionKind
5558 O1K_ARRLEN_OPER_BND,
5559 O1K_ARRLEN_LOOP_BND,
5560 O1K_CONSTANT_LOOP_BND,
5581 optAssertionKind assertionKind;
5584 unsigned lclNum; // assigned to or property of this local var number
5592 struct AssertionDscOp1
5594 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5601 struct AssertionDscOp2
5603 optOp2Kind kind; // a const or copy assignment
5607 ssize_t iconVal; // integer
5608 unsigned iconFlags; // gtFlags
5610 struct Range // integer subrange
5624 bool IsArrLenArithBound()
5626 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5628 bool IsArrLenBound()
5630 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5632 bool IsConstantBound()
5634 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5635 op1.kind == O1K_CONSTANT_LOOP_BND);
5637 bool IsBoundsCheckNoThrow()
5639 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5642 bool IsCopyAssertion()
5644 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5647 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5649 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5650 a1->op2.kind == a2->op2.kind;
5653 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5655 if (kind == OAK_EQUAL)
5657 return kind2 == OAK_NOT_EQUAL;
5659 else if (kind == OAK_NOT_EQUAL)
5661 return kind2 == OAK_EQUAL;
5666 static ssize_t GetLowerBoundForIntegralType(var_types type)
5686 static ssize_t GetUpperBoundForIntegralType(var_types type)
5710 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5712 return (op1.kind == that->op1.kind) &&
5713 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5716 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5718 if (op2.kind != that->op2.kind)
5724 case O2K_IND_CNS_INT:
5726 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5728 case O2K_CONST_LONG:
5729 return (op2.lconVal == that->op2.lconVal);
5731 case O2K_CONST_DOUBLE:
5732 // exact match because of positive and negative zero.
5733 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5735 case O2K_LCLVAR_COPY:
5737 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5738 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5741 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5744 // we will return false
5748 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5754 bool Complementary(AssertionDsc* that, bool vnBased)
5756 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5757 HasSameOp2(that, vnBased);
5760 bool Equals(AssertionDsc* that, bool vnBased)
5762 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5766 typedef unsigned short AssertionIndex;
5769 static fgWalkPreFn optAddCopiesCallback;
5770 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5771 unsigned optAddCopyLclNum;
5772 GenTreePtr optAddCopyAsgnNode;
5774 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5775 bool optAssertionPropagated; // set to true if we modified the trees
5776 bool optAssertionPropagatedCurrentStmt;
5778 GenTreePtr optAssertionPropCurrentTree;
5780 AssertionIndex* optComplementaryAssertionMap;
5781 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5782 // using the value of a local var) for each local var
5783 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5784 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5785 AssertionIndex optMaxAssertionCount;
5788 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5789 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5790 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5791 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5792 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5793 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5795 AssertionIndex GetAssertionCount()
5797 return optAssertionCount;
5799 ASSERT_TP* bbJtrueAssertionOut;
5800 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5801 ValueNumToAssertsMap;
5802 ValueNumToAssertsMap* optValueNumToAsserts;
5804 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5806 // Assertion prop helpers.
5807 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5808 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5809 void optAssertionInit(bool isLocalProp);
5810 void optAssertionTraitsInit(AssertionIndex assertionCount);
5811 #if LOCAL_ASSERTION_PROP
5812 void optAssertionReset(AssertionIndex limit);
5813 void optAssertionRemove(AssertionIndex index);
5816 // Assertion prop data flow functions.
5817 void optAssertionPropMain();
5818 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5819 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5820 ASSERT_TP* optInitAssertionDataflowFlags();
5821 ASSERT_TP* optComputeAssertionGen();
5823 // Assertion Gen functions.
5824 void optAssertionGen(GenTreePtr tree);
5825 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5826 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5827 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5828 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5829 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5830 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5832 // Assertion creation functions.
5833 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5834 AssertionIndex optCreateAssertion(GenTreePtr op1,
5836 optAssertionKind assertionKind,
5837 AssertionDsc* assertion);
5838 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5840 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5841 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5842 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5844 void optPrintVnAssertionMapping();
5846 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5848 // Used for respective assertion propagations.
5849 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5850 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5851 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5852 bool optAssertionIsNonNull(GenTreePtr op,
5853 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5855 // Used for Relop propagation.
5856 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5857 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5858 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5860 // Assertion prop for lcl var functions.
5861 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5862 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5864 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5865 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5866 const GenTreePtr tree,
5867 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5868 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5870 // Assertion propagation functions.
5871 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5872 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5873 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5874 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5875 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5876 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5877 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5878 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5879 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5880 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5881 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5882 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5884 // Implied assertion functions.
5885 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5886 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5887 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5888 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5890 ASSERT_VALRET_TP optNewFullAssertSet();
5891 ASSERT_VALRET_TP optNewEmptyAssertSet();
5894 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5895 void optDebugCheckAssertion(AssertionDsc* assertion);
5896 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5898 void optAddCopies();
5899 #endif // ASSERTION_PROP
5901 /**************************************************************************
5903 *************************************************************************/
5906 struct LoopCloneVisitorInfo
5908 LoopCloneContext* context;
5911 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5912 : context(context), loopNum(loopNum), stmt(nullptr)
5917 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5918 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5919 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5920 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5921 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5922 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5923 void optObtainLoopCloningOpts(LoopCloneContext* context);
5924 bool optIsLoopClonable(unsigned loopInd);
5926 bool optCanCloneLoops();
5929 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5931 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5932 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5933 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5934 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5938 void optInsertLoopCloningStress(BasicBlock* head);
5940 #if COUNT_RANGECHECKS
5941 static unsigned optRangeChkRmv;
5942 static unsigned optRangeChkAll;
5951 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5956 RngChkDsc* rcdNextInBucket; // used by the hash table
5958 unsigned short rcdHashValue; // to make matching faster
5959 unsigned short rcdIndex; // 0..optRngChkCount-1
5961 GenTreePtr rcdTree; // the array index tree
5964 unsigned optRngChkCount;
5965 static const size_t optRngChkHashSize;
5967 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5968 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5970 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5973 bool optLoopsMarked;
5976 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5977 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5981 XX Does the register allocation and puts the remaining lclVars on the stack XX
5983 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5984 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5988 #ifndef LEGACY_BACKEND
5993 #else // LEGACY_BACKEND
5998 #endif // LEGACY_BACKEND
6000 #ifdef LEGACY_BACKEND
6002 void raAssignVars(); // register allocation
6003 #endif // LEGACY_BACKEND
6005 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6007 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6009 void raMarkStkVars();
6012 // Some things are used by both LSRA and regpredict allocators.
6014 FrameType rpFrameType;
6015 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6017 #ifdef LEGACY_BACKEND
6018 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6020 #endif // LEGACY_BACKEND
6022 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6024 #if FEATURE_FP_REGALLOC
6025 enum enumConfigRegisterFP
6027 CONFIG_REGISTER_FP_NONE = 0x0,
6028 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6029 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6030 CONFIG_REGISTER_FP_FULL = 0x3,
6032 enumConfigRegisterFP raConfigRegisterFP();
6033 #endif // FEATURE_FP_REGALLOC
6036 regMaskTP raConfigRestrictMaskFP();
6039 #ifndef LEGACY_BACKEND
6040 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6041 #else // LEGACY_BACKEND
6042 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6043 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6044 bool raNewBlocks; // True is we added killing blocks for FPU registers
6045 unsigned rpPasses; // Number of passes made by the register predicter
6046 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6047 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6048 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6049 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6050 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6051 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6052 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6053 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6054 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6055 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6056 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6057 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6059 bool rpRegAllocDone; // Set to true after we have completed register allocation
6061 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6063 void raSetupArgMasks(RegState* r);
6065 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6067 void raDumpVarIntf(); // Dump the variable to variable interference graph
6068 void raDumpRegIntf(); // Dump the variable to register interference graph
6070 void raAdjustVarIntf();
6072 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6074 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6076 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6077 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6079 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6081 static fgWalkPreFn rpMarkRegIntf;
6083 regMaskTP rpPredictAddressMode(
6084 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6086 void rpPredictRefAssign(unsigned lclNum);
6088 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6090 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6092 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6094 void rpPredictRegUse(); // Entry point
6096 unsigned raPredictTreeRegUse(GenTreePtr tree);
6097 unsigned raPredictListRegUse(GenTreePtr list);
6099 void raSetRegVarOrder(var_types regType,
6100 regNumber* customVarOrder,
6101 unsigned* customVarOrderSize,
6103 regMaskTP avoidReg);
6105 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6106 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6107 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6108 void raAddToStkPredict(unsigned val)
6110 unsigned newStkPredict = rpStkPredict + val;
6111 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6112 rpStkPredict = UINT_MAX - 1;
6114 rpStkPredict = newStkPredict;
6118 #if !FEATURE_FP_REGALLOC
6119 void raDispFPlifeInfo();
6123 regMaskTP genReturnRegForTree(GenTreePtr tree);
6124 #endif // LEGACY_BACKEND
6126 /* raIsVarargsStackArg is called by raMaskStkVars and by
6127 lvaSortByRefCount. It identifies the special case
6128 where a varargs function has a parameter passed on the
6129 stack, other than the special varargs handle. Such parameters
6130 require special treatment, because they cannot be tracked
6131 by the GC (their offsets in the stack are not known
6135 bool raIsVarargsStackArg(unsigned lclNum)
6139 LclVarDsc* varDsc = &lvaTable[lclNum];
6141 assert(varDsc->lvIsParam);
6143 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6145 #else // _TARGET_X86_
6149 #endif // _TARGET_X86_
6152 #ifdef LEGACY_BACKEND
6153 // Records the current prediction, if it's better than any previous recorded prediction.
6154 void rpRecordPrediction();
6155 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6156 void rpUseRecordedPredictionIfBetter();
6158 // Data members used in the methods above.
6159 unsigned rpBestRecordedStkPredict;
6160 struct VarRegPrediction
6162 bool m_isEnregistered;
6163 regNumberSmall m_regNum;
6164 regNumberSmall m_otherReg;
6166 VarRegPrediction* rpBestRecordedPrediction;
6167 #endif // LEGACY_BACKEND
6170 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6171 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6175 XX Get to the class and method info from the Execution Engine given XX
6176 XX tokens for the class and method XX
6178 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6179 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6183 /* These are the different addressing modes used to access a local var.
6184 * The JIT has to report the location of the locals back to the EE
6185 * for debugging purposes.
6191 VLT_REG_BYREF, // this type is currently only used for value types on X64
6194 VLT_STK_BYREF, // this type is currently only used for value types on X64
6208 siVarLocType vlType;
6211 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6213 // VLT_REG_BYREF -- the specified register contains the address of the variable
6221 // VLT_STK -- Any 32 bit value which is on the stack
6222 // eg. [ESP+0x20], or [EBP-0x28]
6223 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6224 // eg. mov EAX, [ESP+0x20]; [EAX]
6228 regNumber vlsBaseReg;
6229 NATIVE_OFFSET vlsOffset;
6232 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6241 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6242 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6250 regNumber vlrssBaseReg;
6251 NATIVE_OFFSET vlrssOffset;
6255 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6256 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6262 regNumber vlsrsBaseReg;
6263 NATIVE_OFFSET vlsrsOffset;
6269 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6270 // eg 2 DWords at [ESP+0x10]
6274 regNumber vls2BaseReg;
6275 NATIVE_OFFSET vls2Offset;
6278 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6279 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6286 // VLT_FIXED_VA -- fixed argument of a varargs function.
6287 // The argument location depends on the size of the variable
6288 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6289 // location of the first arg. This argument can then be accessed
6290 // relative to the position of the first arg
6294 unsigned vlfvOffset;
6301 void* rpValue; // pointer to the in-process
6302 // location of the value.
6308 bool vlIsInReg(regNumber reg);
6309 bool vlIsOnStk(regNumber reg, signed offset);
6312 /*************************************************************************/
6317 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6318 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6319 CORINFO_CALLINFO_FLAGS flags,
6320 CORINFO_CALL_INFO* pResult);
6321 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6323 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6324 CORINFO_ACCESS_FLAGS flags,
6325 CORINFO_FIELD_INFO* pResult);
6329 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6331 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6333 bool IsSuperPMIException(unsigned code)
6335 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6337 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6338 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6339 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6340 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6341 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6342 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6343 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6344 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6348 case EXCEPTIONCODE_DebugBreakorAV:
6349 case EXCEPTIONCODE_MC:
6350 case EXCEPTIONCODE_LWM:
6351 case EXCEPTIONCODE_SASM:
6352 case EXCEPTIONCODE_SSYM:
6353 case EXCEPTIONCODE_CALLUTILS:
6354 case EXCEPTIONCODE_TYPEUTILS:
6355 case EXCEPTIONCODE_ASSERT:
6362 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6363 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6365 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6366 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6369 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6370 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6371 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6373 // VOM info, method sigs
6375 void eeGetSig(unsigned sigTok,
6376 CORINFO_MODULE_HANDLE scope,
6377 CORINFO_CONTEXT_HANDLE context,
6378 CORINFO_SIG_INFO* retSig);
6380 void eeGetCallSiteSig(unsigned sigTok,
6381 CORINFO_MODULE_HANDLE scope,
6382 CORINFO_CONTEXT_HANDLE context,
6383 CORINFO_SIG_INFO* retSig);
6385 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6387 // Method entry-points, instrs
6389 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6391 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6393 CORINFO_EE_INFO eeInfo;
6394 bool eeInfoInitialized;
6396 CORINFO_EE_INFO* eeGetEEInfo();
6398 // Gets the offset of a SDArray's first element
6399 unsigned eeGetArrayDataOffset(var_types type);
6400 // Gets the offset of a MDArray's first element
6401 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6403 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6405 // Returns the page size for the target machine as reported by the EE.
6406 inline size_t eeGetPageSize()
6408 #if COR_JIT_EE_VERSION > 460
6409 return eeGetEEInfo()->osPageSize;
6410 #else // COR_JIT_EE_VERSION <= 460
6411 return CORINFO_PAGE_SIZE;
6412 #endif // COR_JIT_EE_VERSION > 460
6415 // Returns the frame size at which we will generate a loop to probe the stack.
6416 inline size_t getVeryLargeFrameSize()
6419 // The looping probe code is 40 bytes, whereas the straight-line probing for
6420 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6421 // or greater, to generate smaller code.
6422 return 2 * eeGetPageSize();
6424 return 3 * eeGetPageSize();
6428 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6430 #if COR_JIT_EE_VERSION > 460
6431 return eeGetEEInfo()->targetAbi == abi;
6433 return CORINFO_DESKTOP_ABI == abi;
6437 inline bool generateCFIUnwindCodes()
6439 #ifdef UNIX_AMD64_ABI
6440 return IsTargetAbi(CORINFO_CORERT_ABI);
6448 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6450 // Debugging support - Line number info
6452 void eeGetStmtOffsets();
6454 unsigned eeBoundariesCount;
6456 struct boundariesDsc
6458 UNATIVE_OFFSET nativeIP;
6460 unsigned sourceReason;
6461 } * eeBoundaries; // Boundaries to report to EE
6462 void eeSetLIcount(unsigned count);
6463 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6467 static void eeDispILOffs(IL_OFFSET offs);
6468 static void eeDispLineInfo(const boundariesDsc* line);
6469 void eeDispLineInfos();
6472 // Debugging support - Local var info
6476 unsigned eeVarsCount;
6478 struct VarResultInfo
6480 UNATIVE_OFFSET startOffset;
6481 UNATIVE_OFFSET endOffset;
6485 void eeSetLVcount(unsigned count);
6486 void eeSetLVinfo(unsigned which,
6487 UNATIVE_OFFSET startOffs,
6488 UNATIVE_OFFSET length,
6493 const siVarLoc& loc);
6497 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6498 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6501 // ICorJitInfo wrappers
6503 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6505 void eeAllocUnwindInfo(BYTE* pHotCode,
6511 CorJitFuncKind funcKind);
6513 void eeSetEHcount(unsigned cEH);
6515 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6517 WORD eeGetRelocTypeHint(void* target);
6519 // ICorStaticInfo wrapper functions
6521 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6523 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6525 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6528 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6529 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6530 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6531 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6533 template <typename ParamType>
6534 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6536 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6539 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6541 // Utility functions
6543 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6546 const wchar_t* eeGetCPString(size_t stringHandle);
6549 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6551 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6552 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6554 static fgWalkPreFn CountSharedStaticHelper;
6555 static bool IsSharedStaticHelper(GenTreePtr tree);
6556 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6558 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6559 // returns true/false if 'field' is a Jit Data offset
6560 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6561 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6562 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6564 /*****************************************************************************/
6569 enum TEMP_USAGE_TYPE
6575 static var_types tmpNormalizeType(var_types type);
6576 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6577 void tmpRlsTemp(TempDsc* temp);
6578 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6581 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6582 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6586 bool tmpAllFree() const;
6589 #ifndef LEGACY_BACKEND
6590 void tmpPreAllocateTemps(var_types type, unsigned count);
6591 #endif // !LEGACY_BACKEND
6594 #ifdef LEGACY_BACKEND
6595 unsigned tmpIntSpillMax; // number of int-sized spill temps
6596 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6597 #endif // LEGACY_BACKEND
6599 unsigned tmpCount; // Number of temps
6600 unsigned tmpSize; // Size of all the temps
6603 // Used by RegSet::rsSpillChk()
6604 unsigned tmpGetCount; // Temps which haven't been released yet
6607 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6609 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6610 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6613 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6614 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6618 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6619 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6623 CodeGenInterface* codeGen;
6625 // The following holds information about instr offsets in terms of generated code.
6629 IPmappingDsc* ipmdNext; // next line# record
6630 IL_OFFSETX ipmdILoffsx; // the instr offset
6631 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6632 bool ipmdIsLabel; // Can this code be a branch label?
6635 // Record the instr offset mapping to the generated code
6637 IPmappingDsc* genIPmappingList;
6638 IPmappingDsc* genIPmappingLast;
6640 // Managed RetVal - A side hash table meant to record the mapping from a
6641 // GT_CALL node to its IL offset. This info is used to emit sequence points
6642 // that can be used by debugger to determine the native offset at which the
6643 // managed RetVal will be available.
6645 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6646 // favor of a side table for two reasons: 1) We need IL offset for only those
6647 // GT_CALL nodes (created during importation) that correspond to an IL call and
6648 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6649 // structure and IL offset is needed only when generating debuggable code. Therefore
6650 // it is desirable to avoid memory size penalty in retail scenarios.
6651 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6652 CallSiteILOffsetTable;
6653 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6655 unsigned genReturnLocal; // Local number for the return value when applicable.
6656 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6658 // The following properties are part of CodeGenContext. Getters are provided here for
6659 // convenience and backward compatibility, but the properties can only be set by invoking
6660 // the setter on CodeGenContext directly.
6662 __declspec(property(get = getEmitter)) emitter* genEmitter;
6663 emitter* getEmitter()
6665 return codeGen->getEmitter();
6668 const bool isFramePointerUsed()
6670 return codeGen->isFramePointerUsed();
6673 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6674 bool getInterruptible()
6676 return codeGen->genInterruptible;
6678 void setInterruptible(bool value)
6680 codeGen->setInterruptible(value);
6684 const bool genDoubleAlign()
6686 return codeGen->doDoubleAlign();
6688 DWORD getCanDoubleAlign();
6689 bool shouldDoubleAlign(unsigned refCntStk,
6691 unsigned refCntWtdReg,
6692 unsigned refCntStkParam,
6693 unsigned refCntWtdStkDbl);
6694 #endif // DOUBLE_ALIGN
6696 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6697 bool getFullPtrRegMap()
6699 return codeGen->genFullPtrRegMap;
6701 void setFullPtrRegMap(bool value)
6703 codeGen->setFullPtrRegMap(value);
6706 // Things that MAY belong either in CodeGen or CodeGenContext
6708 #if FEATURE_EH_FUNCLETS
6709 FuncInfoDsc* compFuncInfos;
6710 unsigned short compCurrFuncIdx;
6711 unsigned short compFuncInfoCount;
6713 unsigned short compFuncCount()
6715 assert(fgFuncletsCreated);
6716 return compFuncInfoCount;
6719 #else // !FEATURE_EH_FUNCLETS
6721 // This is a no-op when there are no funclets!
6722 void genUpdateCurrentFunclet(BasicBlock* block)
6727 FuncInfoDsc compFuncInfoRoot;
6729 static const unsigned compCurrFuncIdx = 0;
6731 unsigned short compFuncCount()
6736 #endif // !FEATURE_EH_FUNCLETS
6738 FuncInfoDsc* funCurrentFunc();
6739 void funSetCurrentFunc(unsigned funcIdx);
6740 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6741 unsigned int funGetFuncIdx(BasicBlock* block);
6745 VARSET_TP compCurLife; // current live variables
6746 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6748 template <bool ForCodeGen>
6749 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6751 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6753 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6756 template <bool ForCodeGen>
6757 void compUpdateLife(GenTreePtr tree);
6759 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6760 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6761 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6762 template <bool ForCodeGen>
6763 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6765 template <bool ForCodeGen>
6766 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6768 // Gets a register mask that represent the kill set for a helper call since
6769 // not all JIT Helper calls follow the standard ABI on the target architecture.
6770 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6772 // Gets a register mask that represent the kill set for a NoGC helper call.
6773 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6776 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6777 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6778 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6779 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6780 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6781 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6782 #endif // _TARGET_ARM_
6784 // 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
6786 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6788 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6789 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6790 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6791 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6792 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6793 // for the tracked var indices of the field vars, as in a live var set).
6794 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6796 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6798 if (m_promotedStructDeathVars == nullptr)
6800 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6802 return m_promotedStructDeathVars;
6806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6807 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6811 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6812 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6815 #if !defined(__GNUC__)
6816 #pragma region Unwind information
6821 // Infrastructure functions: start/stop/reserve/emit.
6824 void unwindBegProlog();
6825 void unwindEndProlog();
6826 void unwindBegEpilog();
6827 void unwindEndEpilog();
6828 void unwindReserve();
6829 void unwindEmit(void* pHotCode, void* pColdCode);
6832 // Specific unwind information functions: called by code generation to indicate a particular
6833 // prolog or epilog unwindable instruction has been generated.
6836 void unwindPush(regNumber reg);
6837 void unwindAllocStack(unsigned size);
6838 void unwindSetFrameReg(regNumber reg, unsigned offset);
6839 void unwindSaveReg(regNumber reg, unsigned offset);
6841 #if defined(_TARGET_ARM_)
6842 void unwindPushMaskInt(regMaskTP mask);
6843 void unwindPushMaskFloat(regMaskTP mask);
6844 void unwindPopMaskInt(regMaskTP mask);
6845 void unwindPopMaskFloat(regMaskTP mask);
6846 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6847 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6848 // called via unwindPadding().
6849 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6850 // instruction and the current location.
6851 #endif // _TARGET_ARM_
6853 #if defined(_TARGET_ARM64_)
6855 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6856 // instruction and the current location.
6857 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6858 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6859 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6860 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6861 void unwindSaveNext(); // unwind code: save_next
6862 void unwindReturn(regNumber reg); // ret lr
6863 #endif // defined(_TARGET_ARM64_)
6866 // Private "helper" functions for the unwind implementation.
6870 #if FEATURE_EH_FUNCLETS
6871 void unwindGetFuncLocations(FuncInfoDsc* func,
6872 bool getHotSectionData,
6873 /* OUT */ emitLocation** ppStartLoc,
6874 /* OUT */ emitLocation** ppEndLoc);
6875 #endif // FEATURE_EH_FUNCLETS
6877 void unwindReserveFunc(FuncInfoDsc* func);
6878 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6880 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6882 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6883 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6885 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6887 #if defined(_TARGET_AMD64_)
6889 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6891 void unwindBegPrologWindows();
6892 void unwindPushWindows(regNumber reg);
6893 void unwindAllocStackWindows(unsigned size);
6894 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6895 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6897 #ifdef UNIX_AMD64_ABI
6898 void unwindBegPrologCFI();
6899 void unwindPushCFI(regNumber reg);
6900 void unwindAllocStackCFI(unsigned size);
6901 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6902 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6903 int mapRegNumToDwarfReg(regNumber reg);
6904 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6905 #endif // UNIX_AMD64_ABI
6906 #elif defined(_TARGET_ARM_)
6908 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6909 void unwindPushPopMaskFloat(regMaskTP mask);
6910 void unwindSplit(FuncInfoDsc* func);
6912 #endif // _TARGET_ARM_
6914 #if !defined(__GNUC__)
6915 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6919 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6920 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6924 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6925 XX that contains the distinguished, well-known SIMD type definitions). XX
6927 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6928 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6931 // Get highest available instruction set for floating point codegen
6932 InstructionSet getFloatingPointInstructionSet()
6934 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6937 return InstructionSet_AVX;
6942 return InstructionSet_SSE3_4;
6946 assert(canUseSSE2());
6947 return InstructionSet_SSE2;
6949 assert(!"getFPInstructionSet() is not implemented for target arch");
6951 return InstructionSet_NONE;
6955 // Get highest available instruction set for SIMD codegen
6956 InstructionSet getSIMDInstructionSet()
6958 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6959 return getFloatingPointInstructionSet();
6961 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6963 return InstructionSet_NONE;
6969 // Should we support SIMD intrinsics?
6972 // Have we identified any SIMD types?
6973 // This is currently used by struct promotion to avoid getting type information for a struct
6974 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6976 bool _usesSIMDTypes;
6977 bool usesSIMDTypes()
6979 return _usesSIMDTypes;
6981 void setUsesSIMDTypes(bool value)
6983 _usesSIMDTypes = value;
6986 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6987 // that require indexed access to the individual fields of the vector, which is not well supported
6988 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6989 unsigned lvaSIMDInitTempVarNum;
6992 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6993 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6994 CORINFO_CLASS_HANDLE SIMDIntHandle;
6995 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6996 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6997 CORINFO_CLASS_HANDLE SIMDShortHandle;
6998 CORINFO_CLASS_HANDLE SIMDByteHandle;
6999 CORINFO_CLASS_HANDLE SIMDLongHandle;
7000 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7001 CORINFO_CLASS_HANDLE SIMDULongHandle;
7002 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7003 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7004 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7005 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7007 // Get the handle for a SIMD type.
7008 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7010 if (simdBaseType == TYP_FLOAT)
7015 return SIMDVector2Handle;
7017 return SIMDVector3Handle;
7019 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7021 return SIMDVector4Handle;
7030 assert(simdType == getSIMDVectorType());
7031 switch (simdBaseType)
7034 return SIMDFloatHandle;
7036 return SIMDDoubleHandle;
7038 return SIMDIntHandle;
7040 return SIMDUShortHandle;
7042 return SIMDUShortHandle;
7044 return SIMDUByteHandle;
7046 return SIMDShortHandle;
7048 return SIMDByteHandle;
7050 return SIMDLongHandle;
7052 return SIMDUIntHandle;
7054 return SIMDULongHandle;
7056 assert(!"Didn't find a class handle for simdType");
7058 return NO_CLASS_HANDLE;
7062 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7063 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7064 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7066 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7067 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7068 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7069 bool isSIMDTypeLocal(GenTree* tree)
7071 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7074 // Returns true if the type of the tree is a byref of TYP_SIMD
7075 bool isAddrOfSIMDType(GenTree* tree)
7077 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7079 switch (tree->OperGet())
7082 return varTypeIsSIMD(tree->gtGetOp1());
7084 case GT_LCL_VAR_ADDR:
7085 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7088 return isSIMDTypeLocal(tree);
7095 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7097 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7098 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7099 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7102 // Returns base type of a TYP_SIMD local.
7103 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7104 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7106 if (isSIMDTypeLocal(tree))
7108 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7114 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7116 return info.compCompHnd->isInSIMDModule(clsHnd);
7119 bool isSIMDClass(typeInfo* pTypeInfo)
7121 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7124 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7125 // if it is not a SIMD type or is an unsupported base type.
7126 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7128 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7130 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7133 // Get SIMD Intrinsic info given the method handle.
7134 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7135 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7136 CORINFO_METHOD_HANDLE methodHnd,
7137 CORINFO_SIG_INFO* sig,
7140 var_types* baseType,
7141 unsigned* sizeBytes);
7143 // Pops and returns GenTree node from importers type stack.
7144 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7145 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7147 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7148 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7150 // Creates a GT_SIMD tree for Select operation
7151 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7153 unsigned simdVectorSize,
7158 // Creates a GT_SIMD tree for Min/Max operation
7159 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7160 CORINFO_CLASS_HANDLE typeHnd,
7162 unsigned simdVectorSize,
7166 // Transforms operands and returns the SIMD intrinsic to be applied on
7167 // transformed operands to obtain given relop result.
7168 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7169 CORINFO_CLASS_HANDLE typeHnd,
7170 unsigned simdVectorSize,
7171 var_types* baseType,
7175 // Creates a GT_SIMD tree for Abs intrinsic.
7176 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7178 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7179 // Transforms operands and returns the SIMD intrinsic to be applied on
7180 // transformed operands to obtain == comparison result.
7181 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7182 unsigned simdVectorSize,
7186 // Transforms operands and returns the SIMD intrinsic to be applied on
7187 // transformed operands to obtain > comparison result.
7188 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7189 unsigned simdVectorSize,
7193 // Transforms operands and returns the SIMD intrinsic to be applied on
7194 // transformed operands to obtain >= comparison result.
7195 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7196 unsigned simdVectorSize,
7200 // Transforms operands and returns the SIMD intrinsic to be applied on
7201 // transformed operands to obtain >= comparison result in case of int32
7202 // and small int base type vectors.
7203 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7204 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7205 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7207 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7208 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7209 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7210 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7211 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7213 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7214 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7215 GenTreePtr newobjThis,
7216 CORINFO_CLASS_HANDLE clsHnd,
7217 CORINFO_METHOD_HANDLE method,
7218 CORINFO_SIG_INFO* sig,
7221 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7223 // Whether SIMD vector occupies part of SIMD register.
7224 // SSE2: vector2f/3f are considered sub register SIMD types.
7225 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7226 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7228 unsigned sizeBytes = 0;
7229 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7230 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7233 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7235 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7238 // Get the type for the hardware SIMD vector.
7239 // This is the maximum SIMD type supported for this target.
7240 var_types getSIMDVectorType()
7242 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7249 assert(canUseSSE2());
7253 assert(!"getSIMDVectorType() unimplemented on target arch");
7258 // Get the size of the SIMD type in bytes
7259 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7261 unsigned sizeBytes = 0;
7262 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7266 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7267 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7269 // Get the the number of elements of basetype of SIMD vector given by its type handle
7270 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7272 // Get preferred alignment of SIMD type.
7273 int getSIMDTypeAlignment(var_types simdType);
7275 // Get the number of bytes in a SIMD Vector for the current compilation.
7276 unsigned getSIMDVectorRegisterByteLength()
7278 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7281 return YMM_REGSIZE_BYTES;
7285 assert(canUseSSE2());
7286 return XMM_REGSIZE_BYTES;
7289 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7294 // The minimum and maximum possible number of bytes in a SIMD vector.
7295 unsigned int maxSIMDStructBytes()
7297 return getSIMDVectorRegisterByteLength();
7299 unsigned int minSIMDStructBytes()
7301 return emitTypeSize(TYP_SIMD8);
7304 #ifdef FEATURE_AVX_SUPPORT
7305 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7306 static const unsigned maxPossibleSIMDStructBytes = 32;
7307 #else // !FEATURE_AVX_SUPPORT
7308 static const unsigned maxPossibleSIMDStructBytes = 16;
7309 #endif // !FEATURE_AVX_SUPPORT
7311 // Returns the codegen type for a given SIMD size.
7312 var_types getSIMDTypeForSize(unsigned size)
7314 var_types simdType = TYP_UNDEF;
7317 simdType = TYP_SIMD8;
7319 else if (size == 12)
7321 simdType = TYP_SIMD12;
7323 else if (size == 16)
7325 simdType = TYP_SIMD16;
7327 #ifdef FEATURE_AVX_SUPPORT
7328 else if (size == 32)
7330 simdType = TYP_SIMD32;
7332 #endif // FEATURE_AVX_SUPPORT
7335 noway_assert(!"Unexpected size for SIMD type");
7340 unsigned getSIMDInitTempVarNum()
7342 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7344 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7345 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7347 return lvaSIMDInitTempVarNum;
7350 #endif // FEATURE_SIMD
7353 //------------------------------------------------------------------------
7354 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7356 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7357 // candidate for enregistration.
7359 unsigned largestEnregisterableStructSize()
7362 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7363 if (vectorRegSize > TARGET_POINTER_SIZE)
7365 return vectorRegSize;
7368 #endif // FEATURE_SIMD
7370 return TARGET_POINTER_SIZE;
7375 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7376 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7377 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7379 // Is this var is of type simd struct?
7380 bool lclVarIsSIMDType(unsigned varNum)
7382 LclVarDsc* varDsc = lvaTable + varNum;
7383 return varDsc->lvIsSIMDType();
7386 // Is this Local node a SIMD local?
7387 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7389 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7392 // Returns true if the TYP_SIMD locals on stack are aligned at their
7393 // preferred byte boundary specified by getSIMDTypeAlignment().
7395 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7396 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7397 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7398 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7399 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7400 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7401 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7404 bool isSIMDTypeLocalAligned(unsigned varNum)
7406 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7407 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7410 int off = lvaFrameAddress(varNum, &ebpBased);
7411 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7412 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7413 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7416 #endif // FEATURE_SIMD
7421 // Whether SSE2 is available
7422 bool canUseSSE2() const
7424 #ifdef _TARGET_XARCH_
7425 return opts.compCanUseSSE2;
7431 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7432 bool CanUseSSE3_4() const
7434 #ifdef _TARGET_XARCH_
7435 return opts.compCanUseSSE3_4;
7441 bool canUseAVX() const
7443 #ifdef FEATURE_AVX_SUPPORT
7444 return opts.compCanUseAVX;
7451 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7452 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7456 XX Generic info about the compilation and the method being compiled. XX
7457 XX It is responsible for driving the other phases. XX
7458 XX It is also responsible for all the memory management. XX
7460 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7461 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7465 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7467 InlineResult* compInlineResult; // The result of importing the inlinee method.
7469 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7470 bool compJmpOpUsed; // Does the method do a JMP
7471 bool compLongUsed; // Does the method use TYP_LONG
7472 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7473 bool compTailCallUsed; // Does the method do a tailcall
7474 bool compLocallocUsed; // Does the method use localloc.
7475 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7476 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7477 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7479 // NOTE: These values are only reliable after
7480 // the importing is completely finished.
7482 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7483 // we can iterate over these efficiently.
7485 #if CPU_USES_BLOCK_MOVE
7486 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7490 // State information - which phases have completed?
7491 // These are kept together for easy discoverability
7493 bool bRangeAllowStress;
7494 bool compCodeGenDone;
7495 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7496 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7497 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7498 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7501 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7502 bool fgLocalVarLivenessChanged;
7504 bool compStackProbePrologDone;
7506 #ifndef LEGACY_BACKEND
7508 #endif // !LEGACY_BACKEND
7509 bool compRationalIRForm;
7511 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7513 bool compGeneratingProlog;
7514 bool compGeneratingEpilog;
7515 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7516 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7517 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7518 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7519 bool getNeedsGSSecurityCookie() const
7521 return compNeedsGSSecurityCookie;
7523 void setNeedsGSSecurityCookie()
7525 compNeedsGSSecurityCookie = true;
7528 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7529 // frame layout calculations, this is the level we are currently
7532 //---------------------------- JITing options -----------------------------
7545 JitFlags* jitFlags; // all flags passed from the EE
7546 unsigned compFlags; // method attributes
7548 codeOptimize compCodeOpt; // what type of code optimizations
7552 #ifdef _TARGET_XARCH_
7553 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7554 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7556 #ifdef FEATURE_AVX_SUPPORT
7557 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7558 #endif // FEATURE_AVX_SUPPORT
7559 #endif // _TARGET_XARCH_
7561 // optimize maximally and/or favor speed over size?
7563 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7564 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7565 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7566 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7567 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7569 // Maximun number of locals before turning off the inlining
7570 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7573 unsigned instrCount;
7574 unsigned lvRefCount;
7575 bool compMinOptsIsSet;
7577 bool compMinOptsIsUsed;
7579 inline bool MinOpts()
7581 assert(compMinOptsIsSet);
7582 compMinOptsIsUsed = true;
7585 inline bool IsMinOptsSet()
7587 return compMinOptsIsSet;
7590 inline bool MinOpts()
7594 inline bool IsMinOptsSet()
7596 return compMinOptsIsSet;
7599 inline void SetMinOpts(bool val)
7601 assert(!compMinOptsIsUsed);
7602 assert(!compMinOptsIsSet || (compMinOpts == val));
7604 compMinOptsIsSet = true;
7607 // true if the CLFLG_* for an optimization is set.
7608 inline bool OptEnabled(unsigned optFlag)
7610 return !!(compFlags & optFlag);
7613 #ifdef FEATURE_READYTORUN_COMPILER
7614 inline bool IsReadyToRun()
7616 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7619 inline bool IsReadyToRun()
7625 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7626 // PInvoke transitions inline (e.g. when targeting CoreRT).
7627 inline bool ShouldUsePInvokeHelpers()
7629 #if COR_JIT_EE_VERSION > 460
7630 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7636 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7638 inline bool IsReversePInvoke()
7640 #if COR_JIT_EE_VERSION > 460
7641 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7647 // true if we must generate code compatible with JIT32 quirks
7648 inline bool IsJit32Compat()
7650 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7651 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7657 // true if we must generate code compatible with Jit64 quirks
7658 inline bool IsJit64Compat()
7660 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7661 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7662 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7669 bool compScopeInfo; // Generate the LocalVar info ?
7670 bool compDbgCode; // Generate debugger-friendly code?
7671 bool compDbgInfo; // Gather debugging info?
7674 #ifdef PROFILING_SUPPORTED
7675 bool compNoPInvokeInlineCB;
7677 static const bool compNoPInvokeInlineCB;
7681 bool compGcChecks; // Check arguments and return values to ensure they are sane
7682 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7683 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7687 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7688 // to be allocated on the stack.
7689 // It will be set to true in the following cases:
7690 // 1. When the method being compiled has a declarative security
7691 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7692 // This is also the case when we inject a prolog and epilog in the method.
7694 // 2. When the method being compiled has imperative security (i.e. the method
7695 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7697 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7699 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7700 // which gets reported as a GC root to stackwalker.
7701 // (See also ICodeManager::GetAddrOfSecurityObject.)
7708 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7709 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7713 #ifdef UNIX_AMD64_ABI
7714 // This flag is indicating if there is a need to align the frame.
7715 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7716 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7717 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7718 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7719 // there are calls and making sure the frame alignment logic is executed.
7720 bool compNeedToAlignFrame;
7721 #endif // UNIX_AMD64_ABI
7723 bool compProcedureSplitting; // Separate cold code from hot code
7725 bool genFPorder; // Preserve FP order (operations are non-commutative)
7726 bool genFPopt; // Can we do frame-pointer-omission optimization?
7727 bool altJit; // True if we are an altjit and are compiling this method
7730 bool optRepeat; // Repeat optimizer phases k times
7731 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7732 bool dspCode; // Display native code generated
7733 bool dspEHTable; // Display the EH table reported to the VM
7734 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7735 bool dspEmit; // Display emitter output
7736 bool dspLines; // Display source-code lines intermixed with native code output
7737 bool dmpHex; // Display raw bytes in hex of native code output
7738 bool varNames; // Display variables names in native code output
7739 bool disAsm; // Display native code as it is generated
7740 bool disAsmSpilled; // Display native code when any register spilling occurs
7741 bool disDiffable; // Makes the Disassembly code 'diff-able'
7742 bool disAsm2; // Display native code after it is generated using external disassembler
7743 bool dspOrder; // Display names of each of the methods that we ngen/jit
7744 bool dspUnwind; // Display the unwind info output
7745 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7746 bool compLongAddress; // Force using large pseudo instructions for long address
7747 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7748 bool dspGCtbls; // Display the GC tables
7752 bool doLateDisasm; // Run the late disassembler
7753 #endif // LATE_DISASM
7755 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7756 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7757 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7758 static const bool dspGCtbls = true;
7761 // We need stack probes to guarantee that we won't trigger a stack overflow
7762 // when calling unmanaged code until they get a chance to set up a frame, because
7763 // the EE will have no idea where it is.
7765 // We will only be doing this currently for hosted environments. Unfortunately
7766 // we need to take care of stubs, so potentially, we will have to do the probes
7767 // for any call. We have a plan for not needing for stubs though
7768 bool compNeedStackProbes;
7770 #ifdef PROFILING_SUPPORTED
7771 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7772 // This option helps make the JIT behave as if it is running under a profiler.
7773 bool compJitELTHookEnabled;
7774 #endif // PROFILING_SUPPORTED
7776 #if FEATURE_TAILCALL_OPT
7777 // Whether opportunistic or implicit tail call optimization is enabled.
7778 bool compTailCallOpt;
7779 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7780 bool compTailCallLoopOpt;
7784 static const bool compUseSoftFP = true;
7785 #else // !ARM_SOFTFP
7786 static const bool compUseSoftFP = false;
7789 GCPollType compGCPollType;
7793 static bool s_pAltJitExcludeAssembliesListInitialized;
7794 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7799 template <typename T>
7802 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7805 template <typename T>
7808 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7811 static int dspTreeID(GenTree* tree)
7813 return tree->gtTreeID;
7815 static void printTreeID(GenTree* tree)
7817 if (tree == nullptr)
7823 printf("[%06d]", dspTreeID(tree));
7830 #define STRESS_MODES \
7834 /* "Variations" stress areas which we try to mix up with each other. */ \
7835 /* These should not be exhaustively used as they might */ \
7836 /* hide/trivialize other areas */ \
7838 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7839 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7840 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7841 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7842 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7843 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7844 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7845 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7846 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7847 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7848 STRESS_MODE(NULL_OBJECT_CHECK) \
7849 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7850 STRESS_MODE(RANDOM_INLINE) \
7852 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7854 /* "Check" stress areas that can be exhaustively used if we */ \
7855 /* dont care about performance at all */ \
7857 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7858 STRESS_MODE(CHK_FLOW_UPDATE) \
7859 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7861 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7865 #define STRESS_MODE(mode) STRESS_##mode,
7872 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7873 BYTE compActiveStressModes[STRESS_COUNT];
7876 #define MAX_STRESS_WEIGHT 100
7878 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7882 bool compInlineStress()
7884 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7887 bool compRandomInlineStress()
7889 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7894 bool compTailCallStress()
7897 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7903 codeOptimize compCodeOpt()
7906 // Switching between size & speed has measurable throughput impact
7907 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7908 // DEBUG, but should generate identical code between CHK & RET builds,
7909 // so that's not acceptable.
7910 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7911 // Investigate the cause of the throughput regression.
7913 return opts.compCodeOpt;
7915 return BLENDED_CODE;
7919 //--------------------- Info about the procedure --------------------------
7923 COMP_HANDLE compCompHnd;
7924 CORINFO_MODULE_HANDLE compScopeHnd;
7925 CORINFO_CLASS_HANDLE compClassHnd;
7926 CORINFO_METHOD_HANDLE compMethodHnd;
7927 CORINFO_METHOD_INFO* compMethodInfo;
7929 BOOL hasCircularClassConstraints;
7930 BOOL hasCircularMethodConstraints;
7932 #if defined(DEBUG) || defined(LATE_DISASM)
7933 const char* compMethodName;
7934 const char* compClassName;
7935 const char* compFullName;
7936 #endif // defined(DEBUG) || defined(LATE_DISASM)
7938 #if defined(DEBUG) || defined(INLINE_DATA)
7939 // Method hash is logcally const, but computed
7941 mutable unsigned compMethodHashPrivate;
7942 unsigned compMethodHash() const;
7943 #endif // defined(DEBUG) || defined(INLINE_DATA)
7945 #ifdef PSEUDORANDOM_NOP_INSERTION
7946 // things for pseudorandom nop insertion
7947 unsigned compChecksum;
7951 // The following holds the FLG_xxxx flags for the method we're compiling.
7954 // The following holds the class attributes for the method we're compiling.
7955 unsigned compClassAttr;
7957 const BYTE* compCode;
7958 IL_OFFSET compILCodeSize; // The IL code size
7959 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7960 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7961 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7962 // (2) the code is hot/cold split, and we issued less code than we expected
7963 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7965 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7966 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7967 bool compIsContextful : 1; // contextful method
7968 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7969 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7970 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7971 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7972 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7974 var_types compRetType; // Return type of the method as declared in IL
7975 var_types compRetNativeType; // Normalized return type as per target arch ABI
7976 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7977 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7978 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7979 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7980 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7981 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7982 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7983 unsigned compMaxStack;
7984 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7985 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7987 unsigned compCallUnmanaged; // count of unmanaged calls
7988 unsigned compLvFrameListRoot; // lclNum for the Frame root
7989 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7990 // You should generally use compHndBBtabCount instead: it is the
7991 // current number of EH clauses (after additions like synchronized
7992 // methods and funclets, and removals like unreachable code deletion).
7994 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7995 // and the VM expects that, or the JIT is a "self-host" compiler
7996 // (e.g., x86 hosted targeting x86) and the VM expects that.
7998 /* The following holds IL scope information about local variables.
8001 unsigned compVarScopesCount;
8002 VarScopeDsc* compVarScopes;
8004 /* The following holds information about instr offsets for
8005 * which we need to report IP-mappings
8008 IL_OFFSET* compStmtOffsets; // sorted
8009 unsigned compStmtOffsetsCount;
8010 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8012 #define CPU_X86 0x0100 // The generic X86 CPU
8013 #define CPU_X86_PENTIUM_4 0x0110
8015 #define CPU_X64 0x0200 // The generic x64 CPU
8016 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8017 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8019 #define CPU_ARM 0x0300 // The generic ARM CPU
8021 unsigned genCPU; // What CPU are we running on
8024 // Returns true if the method being compiled returns a non-void and non-struct value.
8025 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8026 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8027 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8028 // Methods returning such structs are considered to return non-struct return value and
8029 // this method returns true in that case.
8030 bool compMethodReturnsNativeScalarType()
8032 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8035 // Returns true if the method being compiled returns RetBuf addr as its return value
8036 bool compMethodReturnsRetBufAddr()
8038 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8039 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8041 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8042 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8043 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8044 // methods with hidden RetBufArg.
8046 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8047 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8048 // returning the address of RetBuf.
8050 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8051 // to be returned in RAX.
8052 CLANG_FORMAT_COMMENT_ANCHOR;
8054 #ifdef _TARGET_AMD64_
8055 return (info.compRetBuffArg != BAD_VAR_NUM);
8056 #else // !_TARGET_AMD64_
8057 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8058 #endif // !_TARGET_AMD64_
8061 // Returns true if the method returns a value in more than one return register
8062 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8063 // TODO-ARM64: Does this apply for ARM64 too?
8064 bool compMethodReturnsMultiRegRetType()
8066 #if FEATURE_MULTIREG_RET
8067 #if defined(_TARGET_X86_)
8068 // On x86 only 64-bit longs are returned in multiple registers
8069 return varTypeIsLong(info.compRetNativeType);
8070 #else // targets: X64-UNIX, ARM64 or ARM32
8071 // On all other targets that support multireg return values:
8072 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8073 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8074 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8075 #endif // TARGET_XXX
8077 #else // not FEATURE_MULTIREG_RET
8079 // For this architecture there are no multireg returns
8082 #endif // FEATURE_MULTIREG_RET
8085 #if FEATURE_MULTIREG_ARGS
8086 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8087 // return the gcPtr layout for the pointers sized fields
8088 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8089 #endif // FEATURE_MULTIREG_ARGS
8091 // Returns true if the method being compiled returns a value
8092 bool compMethodHasRetVal()
8094 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8095 compMethodReturnsMultiRegRetType();
8100 void compDispLocalVars();
8104 //-------------------------- Global Compiler Data ------------------------------------
8107 static unsigned s_compMethodsCount; // to produce unique label names
8108 unsigned compGenTreeID;
8111 BasicBlock* compCurBB; // the current basic block in process
8112 GenTreePtr compCurStmt; // the current statement in process
8114 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8117 // The following is used to create the 'method JIT info' block.
8118 size_t compInfoBlkSize;
8119 BYTE* compInfoBlkAddr;
8121 EHblkDsc* compHndBBtab; // array of EH data
8122 unsigned compHndBBtabCount; // element count of used elements in EH data array
8123 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8125 #if defined(_TARGET_X86_)
8127 //-------------------------------------------------------------------------
8128 // Tracking of region covered by the monitor in synchronized methods
8129 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8130 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8132 #endif // !_TARGET_X86_
8134 Phases previousCompletedPhase; // the most recently completed phase
8136 //-------------------------------------------------------------------------
8137 // The following keeps track of how many bytes of local frame space we've
8138 // grabbed so far in the current function, and how many argument bytes we
8139 // need to pop when we return.
8142 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8144 // Count of callee-saved regs we pushed in the prolog.
8145 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8146 // In case of Amd64 this doesn't include float regs saved on stack.
8147 unsigned compCalleeRegsPushed;
8149 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8150 // Mask of callee saved float regs on stack.
8151 regMaskTP compCalleeFPRegsSavedMask;
8153 #ifdef _TARGET_AMD64_
8154 // Quirk for VS debug-launch scenario to work:
8155 // Bytes of padding between save-reg area and locals.
8156 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8157 unsigned compVSQuirkStackPaddingNeeded;
8158 bool compQuirkForPPPflag;
8161 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8163 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8164 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8165 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8167 //-------------------------------------------------------------------------
8169 static void compStartup(); // One-time initialization
8170 static void compShutdown(); // One-time finalization
8172 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8175 static void compDisplayStaticSizes(FILE* fout);
8177 //------------ Some utility functions --------------
8179 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8180 void** ppIndirection); /* OUT */
8182 // Several JIT/EE interface functions return a CorInfoType, and also return a
8183 // class handle as an out parameter if the type is a value class. Returns the
8184 // size of the type these describe.
8185 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8188 // Components used by the compiler may write unit test suites, and
8189 // have them run within this method. They will be run only once per process, and only
8190 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8191 // These should fail by asserting.
8192 void compDoComponentUnitTestsOnce();
8195 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8196 CORINFO_MODULE_HANDLE classPtr,
8197 COMP_HANDLE compHnd,
8198 CORINFO_METHOD_INFO* methodInfo,
8199 void** methodCodePtr,
8200 ULONG* methodCodeSize,
8201 JitFlags* compileFlags);
8202 void compCompileFinish();
8203 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8204 COMP_HANDLE compHnd,
8205 CORINFO_METHOD_INFO* methodInfo,
8206 void** methodCodePtr,
8207 ULONG* methodCodeSize,
8208 JitFlags* compileFlags,
8209 CorInfoInstantiationVerification instVerInfo);
8211 ArenaAllocator* compGetAllocator();
8213 #if MEASURE_MEM_ALLOC
8215 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8219 unsigned allocCnt; // # of allocs
8220 UINT64 allocSz; // total size of those alloc.
8221 UINT64 allocSzMax; // Maximum single allocation.
8222 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8223 UINT64 nraTotalSizeAlloc;
8224 UINT64 nraTotalSizeUsed;
8226 static const char* s_CompMemKindNames[]; // Names of the kinds.
8228 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8230 for (int i = 0; i < CMK_Count; i++)
8232 allocSzByKind[i] = 0;
8235 MemStats(const MemStats& ms)
8236 : allocCnt(ms.allocCnt)
8237 , allocSz(ms.allocSz)
8238 , allocSzMax(ms.allocSzMax)
8239 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8240 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8242 for (int i = 0; i < CMK_Count; i++)
8244 allocSzByKind[i] = ms.allocSzByKind[i];
8248 // Until we have ubiquitous constructors.
8251 this->MemStats::MemStats();
8254 void AddAlloc(size_t sz, CompMemKind cmk)
8258 if (sz > allocSzMax)
8262 allocSzByKind[cmk] += sz;
8265 void Print(FILE* f); // Print these stats to f.
8266 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8268 MemStats genMemStats;
8270 struct AggregateMemStats : public MemStats
8274 AggregateMemStats() : MemStats(), nMethods(0)
8278 void Add(const MemStats& ms)
8281 allocCnt += ms.allocCnt;
8282 allocSz += ms.allocSz;
8283 allocSzMax = max(allocSzMax, ms.allocSzMax);
8284 for (int i = 0; i < CMK_Count; i++)
8286 allocSzByKind[i] += ms.allocSzByKind[i];
8288 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8289 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8292 void Print(FILE* f); // Print these stats to jitstdout.
8295 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8296 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8297 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8299 #endif // MEASURE_MEM_ALLOC
8301 #if LOOP_HOIST_STATS
8302 unsigned m_loopsConsidered;
8303 bool m_curLoopHasHoistedExpression;
8304 unsigned m_loopsWithHoistedExpressions;
8305 unsigned m_totalHoistedExpressions;
8307 void AddLoopHoistStats();
8308 void PrintPerMethodLoopHoistStats();
8310 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8311 static unsigned s_loopsConsidered;
8312 static unsigned s_loopsWithHoistedExpressions;
8313 static unsigned s_totalHoistedExpressions;
8315 static void PrintAggregateLoopHoistStats(FILE* f);
8316 #endif // LOOP_HOIST_STATS
8318 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8319 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8320 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8321 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8322 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8323 void compFreeMem(void*);
8325 bool compIsForImportOnly();
8326 bool compIsForInlining();
8327 bool compDonotInline();
8330 const char* compLocalVarName(unsigned varNum, unsigned offs);
8331 VarName compVarName(regNumber reg, bool isFloatReg = false);
8332 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8333 const char* compRegPairName(regPairNo regPair);
8334 const char* compRegNameForSize(regNumber reg, size_t size);
8335 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8336 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8337 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8340 //-------------------------------------------------------------------------
8342 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8344 struct VarScopeMapInfo
8346 VarScopeListNode* head;
8347 VarScopeListNode* tail;
8348 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8350 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8357 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8358 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8360 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8361 VarNumToScopeDscMap;
8363 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8364 VarNumToScopeDscMap* compVarScopeMap;
8366 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8368 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8370 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8372 void compInitVarScopeMap();
8374 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8375 // enter scope, sorted by instr offset
8376 unsigned compNextEnterScope;
8378 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8379 // go out of scope, sorted by instr offset
8380 unsigned compNextExitScope;
8382 void compInitScopeLists();
8384 void compResetScopeLists();
8386 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8388 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8390 void compProcessScopesUntil(unsigned offset,
8392 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8393 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8396 void compDispScopeLists();
8399 bool compIsProfilerHookNeeded();
8401 //-------------------------------------------------------------------------
8402 /* Statistical Data Gathering */
8404 void compJitStats(); // call this function and enable
8405 // various ifdef's below for statistical data
8408 void compCallArgStats();
8409 static void compDispCallArgStats(FILE* fout);
8412 //-------------------------------------------------------------------------
8419 ArenaAllocator* compAllocator;
8422 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8423 // suitable for use by utilcode collection types.
8424 IAllocator* compAsIAllocator;
8426 #if MEASURE_MEM_ALLOC
8427 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8428 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8429 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8431 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8433 #endif // MEASURE_MEM_ALLOC
8435 void compFunctionTraceStart();
8436 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8439 size_t compMaxUncheckedOffsetForNullObject;
8441 void compInitOptions(JitFlags* compileFlags);
8443 void compSetProcessor();
8444 void compInitDebuggingInfo();
8445 void compSetOptimizationLevel();
8446 #ifdef _TARGET_ARMARCH_
8447 bool compRsvdRegCheck(FrameLayoutState curState);
8449 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8451 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8452 void ResetOptAnnotations();
8454 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8455 void RecomputeLoopInfo();
8457 #ifdef PROFILING_SUPPORTED
8458 // Data required for generating profiler Enter/Leave/TailCall hooks
8460 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8461 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8462 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8465 #ifdef _TARGET_AMD64_
8466 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8469 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8470 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8472 IAllocator* getAllocator()
8474 return compAsIAllocator;
8477 #if MEASURE_MEM_ALLOC
8478 IAllocator* getAllocatorBitset()
8480 return compAsIAllocatorBitset;
8482 IAllocator* getAllocatorGC()
8484 return compAsIAllocatorGC;
8486 IAllocator* getAllocatorLoopHoist()
8488 return compAsIAllocatorLoopHoist;
8490 #else // !MEASURE_MEM_ALLOC
8491 IAllocator* getAllocatorBitset()
8493 return compAsIAllocator;
8495 IAllocator* getAllocatorGC()
8497 return compAsIAllocator;
8499 IAllocator* getAllocatorLoopHoist()
8501 return compAsIAllocator;
8503 #endif // !MEASURE_MEM_ALLOC
8506 IAllocator* getAllocatorDebugOnly()
8508 #if MEASURE_MEM_ALLOC
8509 return compAsIAllocatorDebugOnly;
8510 #else // !MEASURE_MEM_ALLOC
8511 return compAsIAllocator;
8512 #endif // !MEASURE_MEM_ALLOC
8517 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8518 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8522 XX Checks for type compatibility and merges types XX
8524 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8525 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8529 // Set to TRUE if verification cannot be skipped for this method
8530 // If we detect unverifiable code, we will lazily check
8531 // canSkipMethodVerification() to see if verification is REALLY needed.
8532 BOOL tiVerificationNeeded;
8534 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8535 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8536 BOOL tiIsVerifiableCode;
8538 // Set to TRUE if runtime callout is needed for this method
8539 BOOL tiRuntimeCalloutNeeded;
8541 // Set to TRUE if security prolog/epilog callout is needed for this method
8542 // Note: This flag is different than compNeedSecurityCheck.
8543 // compNeedSecurityCheck means whether or not a security object needs
8544 // to be allocated on the stack, which is currently true for EnC as well.
8545 // tiSecurityCalloutNeeded means whether or not security callouts need
8546 // to be inserted in the jitted code.
8547 BOOL tiSecurityCalloutNeeded;
8549 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8550 // This support is necessary to suport attributes that are not described in
8551 // for example, signatures. For example, the permanent home byref (byref that
8552 // points to the gc heap), isn't a property of method signatures, therefore,
8553 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8554 // but when deciding if we need to reimport a block, we need to take these
8556 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8558 // Returns TRUE if child is equal to or a subtype of parent.
8559 // normalisedForStack indicates that both types are normalised for the stack
8560 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8562 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8563 // *pDest is modified to represent the merged type. Sets "*changed" to true
8564 // if this changes "*pDest".
8565 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8567 // Set pDest from the primitive value type.
8568 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8570 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8573 // <BUGNUM> VSW 471305
8574 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8575 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8576 // We use a "short" as we need to push/pop this scope.
8578 short compRegSetCheckLevel;
8582 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8583 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8585 XX IL verification stuff XX
8588 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8589 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8593 // The following is used to track liveness of local variables, initialization
8594 // of valueclass constructors, and type safe use of IL instructions.
8596 // dynamic state info needed for verification
8597 EntryState verCurrentState;
8599 // this ptr of object type .ctors are considered intited only after
8600 // the base class ctor is called, or an alternate ctor is called.
8601 // An uninited this ptr can be used to access fields, but cannot
8602 // be used to call a member function.
8603 BOOL verTrackObjCtorInitState;
8605 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8607 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8608 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8609 void verInitCurrentState();
8610 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8612 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8613 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8614 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8616 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8617 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8618 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8619 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8620 typeInfo verMakeTypeInfo(CorInfoType ciType,
8621 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8622 BOOL verIsSDArray(typeInfo ti);
8623 typeInfo verGetArrayElemType(typeInfo ti);
8625 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8626 BOOL verNeedsVerification();
8627 BOOL verIsByRefLike(const typeInfo& ti);
8628 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8630 // generic type variables range over types that satisfy IsBoxable
8631 BOOL verIsBoxable(const typeInfo& ti);
8633 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8634 DEBUGARG(unsigned line));
8635 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8636 DEBUGARG(unsigned line));
8637 bool verCheckTailCallConstraint(OPCODE opcode,
8638 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8639 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8640 // on a type parameter?
8641 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8642 // return false to the caller.
8643 // If false, it will throw.
8645 bool verIsBoxedValueType(typeInfo ti);
8647 void verVerifyCall(OPCODE opcode,
8648 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8649 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8651 bool readonlyCall, // is this a "readonly." call?
8652 const BYTE* delegateCreateStart,
8653 const BYTE* codeAddr,
8654 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8656 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8658 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8659 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8660 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8661 const CORINFO_FIELD_INFO& fieldInfo,
8662 const typeInfo* tiThis,
8664 BOOL allowPlainStructAsThis = FALSE);
8665 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8666 void verVerifyThisPtrInitialised();
8667 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8669 // Register allocator
8670 void raInitStackFP();
8671 void raEnregisterVarsPrePassStackFP();
8672 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8673 void raEnregisterVarsPostPassStackFP();
8674 void raGenerateFPRefCounts();
8675 void raEnregisterVarsStackFP();
8676 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8678 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8679 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8681 // returns true if enregistering v1 would save more mem accesses than v2
8682 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8685 void raDumpHeightsStackFP();
8686 void raDumpVariableRegIntfFloat();
8689 #if FEATURE_STACK_FP_X87
8691 // Currently, we use FP transition blocks in only 2 situations:
8693 // -conditional jump on longs where FP stack differs with target: it's not strictly
8694 // necessary, but its low frequency and the code would get complicated if we try to
8695 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8696 // minimize the way we generate the jump code.
8697 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8698 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8700 // However, transition blocks have 2 problems
8702 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8703 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8704 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8705 // in the right place without preordering them), this causes us to have to generate the transition
8706 // blocks in the cold area if we want procedure splitting.
8709 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8710 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8711 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8712 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8713 // a big change in the exception.
8715 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8716 // optimizations. For these 2 cases:
8718 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8719 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8720 // a switch statement.
8722 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8723 // current procedure splitting and exception code have.
8724 bool compMayHaveTransitionBlocks;
8726 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8728 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8730 unsigned raCntStkStackFP;
8731 unsigned raCntWtdStkDblStackFP;
8732 unsigned raCntStkParamDblStackFP;
8734 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8735 // TODO: Do we want to put this in LclVarDsc?
8736 unsigned raPayloadStackFP[lclMAX_TRACKED];
8737 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8739 // Useful for debugging
8740 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8742 #endif // FEATURE_STACK_FP_X87
8745 // One line log function. Default level is 0. Increasing it gives you
8746 // more log information
8748 // levels are currently unused: #define JITDUMP(level,...) ();
8749 void JitLogEE(unsigned level, const char* fmt, ...);
8751 bool compDebugBreak;
8753 bool compJitHaltMethod();
8758 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8759 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8761 XX GS Security checks for unsafe buffers XX
8763 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8764 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8767 struct ShadowParamVarInfo
8769 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8770 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8772 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8774 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8775 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8776 // slots and update all trees to refer to shadow slots is done immediately after
8777 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8778 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8779 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8780 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8781 // creating a shadow slot even though this routine returns true.
8783 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8784 // required. There are two cases under which a reg arg could potentially be used from its
8786 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8787 // b) LSRA spills it
8789 // Possible solution to address case (a)
8790 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8791 // in this routine. Note that live out of exception handler is something we may not be
8792 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8793 // Therefore, for methods with exception handling and need GS cookie check we might have
8794 // to take conservative approach.
8796 // Possible solution to address case (b)
8797 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8798 // create a new spill temp if the method needs GS cookie check.
8799 return varDsc->lvIsParam;
8800 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8801 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8808 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8813 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8814 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8815 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8817 void gsGSChecksInitCookie(); // Grabs cookie variable
8818 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8819 bool gsFindVulnerableParams(); // Shadow param analysis code
8820 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8822 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8823 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8825 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8826 // This can be overwritten by setting complus_JITInlineSize env variable.
8828 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8831 #ifdef FEATURE_JIT_METHOD_PERF
8832 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8833 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8835 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8836 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8838 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8840 #if MEASURE_CLRAPI_CALLS
8841 // Thin wrappers that call into JitTimer (if present).
8842 inline void CLRApiCallEnter(unsigned apix);
8843 inline void CLRApiCallLeave(unsigned apix);
8846 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8847 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8852 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8853 // These variables are associated with maintaining SQM data about compile time.
8854 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8855 // in the current compilation.
8856 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8857 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8858 // the inlining phase in the current compilation.
8859 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8861 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8862 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8863 // type-loading and class initialization).
8864 void RecordStateAtEndOfInlining();
8865 // Assumes being called at the end of compilation. Update the SQM state.
8866 void RecordStateAtEndOfCompilation();
8868 #ifdef FEATURE_CLRSQM
8869 // Does anything SQM related necessary at process shutdown time.
8870 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8871 #endif // FEATURE_CLRSQM
8874 #if FUNC_INFO_LOGGING
8875 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8876 // filename to write it to.
8877 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8878 #endif // FUNC_INFO_LOGGING
8880 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8882 // Is the compilation in a full trust context?
8883 bool compIsFullTrust();
8885 #ifndef FEATURE_TRACELOGGING
8886 // Should we actually fire the noway assert body and the exception handler?
8887 bool compShouldThrowOnNoway();
8888 #else // FEATURE_TRACELOGGING
8889 // Should we actually fire the noway assert body and the exception handler?
8890 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8892 // Telemetry instance to use per method compilation.
8893 JitTelemetry compJitTelemetry;
8895 // Get common parameters that have to be logged with most telemetry data.
8896 void compGetTelemetryDefaults(const char** assemblyName,
8897 const char** scopeName,
8898 const char** methodName,
8899 unsigned* methodHash);
8900 #endif // !FEATURE_TRACELOGGING
8904 NodeToTestDataMap* m_nodeTestData;
8906 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8907 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8908 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8909 // Current kept in this.
8911 NodeToTestDataMap* GetNodeTestData()
8913 Compiler* compRoot = impInlineRoot();
8914 if (compRoot->m_nodeTestData == nullptr)
8916 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8918 return compRoot->m_nodeTestData;
8921 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8923 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8924 // currently occur in the AST graph.
8925 NodeToIntMap* FindReachableNodesInNodeTestData();
8927 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8928 // test data, associate that data with "to".
8929 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8931 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8932 // have annotations, attach similar annotations to the corresponding nodes in "to".
8933 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8935 // These are the methods that test that the various conditions implied by the
8936 // test attributes are satisfied.
8937 void JitTestCheckSSA(); // SSA builder tests.
8938 void JitTestCheckVN(); // Value numbering tests.
8941 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8943 FieldSeqStore* m_fieldSeqStore;
8945 FieldSeqStore* GetFieldSeqStore()
8947 Compiler* compRoot = impInlineRoot();
8948 if (compRoot->m_fieldSeqStore == nullptr)
8950 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8951 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8952 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8954 return compRoot->m_fieldSeqStore;
8957 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8959 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8960 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8961 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8962 // attach the field sequence directly to the address node.
8963 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8965 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8967 // Don't need to worry about inlining here
8968 if (m_zeroOffsetFieldMap == nullptr)
8970 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8972 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8973 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8975 return m_zeroOffsetFieldMap;
8978 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8979 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8980 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8981 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8982 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8983 // record the the field sequence using the ZeroOffsetFieldMap described above.
8985 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8986 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8987 // CoreRT. Such case is handled same as the default case.
8988 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8990 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8992 NodeToArrayInfoMap* m_arrayInfoMap;
8994 NodeToArrayInfoMap* GetArrayInfoMap()
8996 Compiler* compRoot = impInlineRoot();
8997 if (compRoot->m_arrayInfoMap == nullptr)
8999 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9000 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9001 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9003 return compRoot->m_arrayInfoMap;
9006 NodeToUnsignedMap* m_heapSsaMap;
9008 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
9009 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
9010 // all the possible heap states are possible initial states of the corresponding catch block(s).)
9011 NodeToUnsignedMap* GetHeapSsaMap()
9013 Compiler* compRoot = impInlineRoot();
9014 if (compRoot->m_heapSsaMap == nullptr)
9016 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9017 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9018 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
9020 return compRoot->m_heapSsaMap;
9023 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9024 CORINFO_CLASS_HANDLE m_refAnyClass;
9025 CORINFO_FIELD_HANDLE GetRefanyDataField()
9027 if (m_refAnyClass == nullptr)
9029 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9031 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9033 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9035 if (m_refAnyClass == nullptr)
9037 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9039 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9043 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9045 #if ALLVARSET_COUNTOPS
9046 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9049 static HelperCallProperties s_helperCallProperties;
9051 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9052 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9053 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9055 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9058 unsigned __int8* offset0,
9059 unsigned __int8* offset1);
9060 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9061 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9063 void fgMorphMultiregStructArgs(GenTreeCall* call);
9064 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9066 }; // end of class Compiler
9068 // Inline methods of CompAllocator.
9069 void* CompAllocator::Alloc(size_t sz)
9071 #if MEASURE_MEM_ALLOC
9072 return m_comp->compGetMem(sz, m_cmk);
9074 return m_comp->compGetMem(sz);
9078 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9080 #if MEASURE_MEM_ALLOC
9081 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9083 return m_comp->compGetMemArray(elems, elemSize);
9087 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9088 inline LclVarDsc::LclVarDsc(Compiler* comp)
9089 : // Initialize the ArgRegs to REG_STK.
9090 // The morph will do the right thing to change
9091 // to the right register if passed in register.
9094 #if FEATURE_MULTIREG_ARGS
9095 _lvOtherArgReg(REG_STK)
9097 #endif // FEATURE_MULTIREG_ARGS
9099 lvRefBlks(BlockSetOps::UninitVal())
9101 #endif // ASSERTION_PROP
9102 lvPerSsaData(comp->getAllocator())
9107 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9108 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9110 XX Miscellaneous Compiler stuff XX
9112 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9113 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9116 // Values used to mark the types a stack slot is used for
9118 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9119 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9120 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9121 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9122 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9123 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9124 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9125 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9127 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9129 /*****************************************************************************
9131 * Variables to keep track of total code amounts.
9136 extern size_t grossVMsize;
9137 extern size_t grossNCsize;
9138 extern size_t totalNCsize;
9140 extern unsigned genMethodICnt;
9141 extern unsigned genMethodNCnt;
9142 extern size_t gcHeaderISize;
9143 extern size_t gcPtrMapISize;
9144 extern size_t gcHeaderNSize;
9145 extern size_t gcPtrMapNSize;
9147 #endif // DISPLAY_SIZES
9149 /*****************************************************************************
9151 * Variables to keep track of basic block counts (more data on 1 BB methods)
9154 #if COUNT_BASIC_BLOCKS
9155 extern Histogram bbCntTable;
9156 extern Histogram bbOneBBSizeTable;
9159 /*****************************************************************************
9161 * Used by optFindNaturalLoops to gather statistical information such as
9162 * - total number of natural loops
9163 * - number of loops with 1, 2, ... exit conditions
9164 * - number of loops that have an iterator (for like)
9165 * - number of loops that have a constant iterator
9170 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9171 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9172 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9173 extern unsigned totalLoopCount; // counts the total number of natural loops
9174 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9175 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9176 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9177 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9179 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9180 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9181 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9182 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9183 extern Histogram loopCountTable; // Histogram of loop counts
9184 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9186 #endif // COUNT_LOOPS
9188 /*****************************************************************************
9189 * variables to keep track of how many iterations we go in a dataflow pass
9194 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9195 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9197 #endif // DATAFLOW_ITER
9199 #if MEASURE_BLOCK_SIZE
9200 extern size_t genFlowNodeSize;
9201 extern size_t genFlowNodeCnt;
9202 #endif // MEASURE_BLOCK_SIZE
9204 #if MEASURE_NODE_SIZE
9205 struct NodeSizeStats
9210 genTreeNodeSize = 0;
9211 genTreeNodeActualSize = 0;
9214 size_t genTreeNodeCnt;
9215 size_t genTreeNodeSize; // The size we allocate
9216 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9217 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9218 // a smaller node to a larger one. TODO-Cleanup: add stats on
9219 // SetOper()/ChangeOper() usage to quanitfy this.
9221 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9222 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9223 extern Histogram genTreeNcntHist;
9224 extern Histogram genTreeNsizHist;
9225 #endif // MEASURE_NODE_SIZE
9227 /*****************************************************************************
9228 * Count fatal errors (including noway_asserts).
9232 extern unsigned fatal_badCode;
9233 extern unsigned fatal_noWay;
9234 extern unsigned fatal_NOMEM;
9235 extern unsigned fatal_noWayAssertBody;
9237 extern unsigned fatal_noWayAssertBodyArgs;
9239 extern unsigned fatal_NYI;
9240 #endif // MEASURE_FATAL
9242 /*****************************************************************************
9246 #ifdef _TARGET_XARCH_
9248 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9249 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9250 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9252 const instruction INS_AND = INS_and;
9253 const instruction INS_OR = INS_or;
9254 const instruction INS_XOR = INS_xor;
9255 const instruction INS_NEG = INS_neg;
9256 const instruction INS_TEST = INS_test;
9257 const instruction INS_MUL = INS_imul;
9258 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9259 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9260 const instruction INS_BREAKPOINT = INS_int3;
9261 const instruction INS_ADDC = INS_adc;
9262 const instruction INS_SUBC = INS_sbb;
9263 const instruction INS_NOT = INS_not;
9269 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9270 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9271 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9273 const instruction INS_AND = INS_and;
9274 const instruction INS_OR = INS_orr;
9275 const instruction INS_XOR = INS_eor;
9276 const instruction INS_NEG = INS_rsb;
9277 const instruction INS_TEST = INS_tst;
9278 const instruction INS_MUL = INS_mul;
9279 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9280 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9281 const instruction INS_BREAKPOINT = INS_bkpt;
9282 const instruction INS_ADDC = INS_adc;
9283 const instruction INS_SUBC = INS_sbc;
9284 const instruction INS_NOT = INS_mvn;
9288 #ifdef _TARGET_ARM64_
9290 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9291 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9292 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9294 const instruction INS_AND = INS_and;
9295 const instruction INS_OR = INS_orr;
9296 const instruction INS_XOR = INS_eor;
9297 const instruction INS_NEG = INS_neg;
9298 const instruction INS_TEST = INS_tst;
9299 const instruction INS_MUL = INS_mul;
9300 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9301 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9302 const instruction INS_BREAKPOINT = INS_bkpt;
9303 const instruction INS_ADDC = INS_adc;
9304 const instruction INS_SUBC = INS_sbc;
9305 const instruction INS_NOT = INS_mvn;
9309 /*****************************************************************************/
9311 extern const BYTE genTypeSizes[];
9312 extern const BYTE genTypeAlignments[];
9313 extern const BYTE genTypeStSzs[];
9314 extern const BYTE genActualTypes[];
9316 /*****************************************************************************/
9318 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9319 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9322 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9323 #elif defined(_TARGET_ARM64_)
9324 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9327 /*****************************************************************************/
9329 #define REG_CORRUPT regNumber(REG_NA + 1)
9330 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9331 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9333 /*****************************************************************************/
9335 extern BasicBlock dummyBB;
9337 /*****************************************************************************/
9338 /*****************************************************************************/
9340 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9341 // nodes of a statement in execution order.
9342 // __stmt: a GT_STMT type GenTree*
9343 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9345 #define foreach_treenode_execution_order(__node, __stmt) \
9346 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9348 // foreach_block: An iterator over all blocks in the function.
9349 // __compiler: the Compiler* object
9350 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9352 #define foreach_block(__compiler, __block) \
9353 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9355 /*****************************************************************************/
9356 /*****************************************************************************/
9360 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9362 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9363 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9365 XX Debugging helpers XX
9367 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9368 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9371 /*****************************************************************************/
9372 /* The following functions are intended to be called from the debugger, to dump
9373 * various data structures. The can be used in the debugger Watch or Quick Watch
9374 * windows. They are designed to be short to type and take as few arguments as
9375 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9376 * See the function definition comment for more details.
9379 void cBlock(Compiler* comp, BasicBlock* block);
9380 void cBlocks(Compiler* comp);
9381 void cBlocksV(Compiler* comp);
9382 void cTree(Compiler* comp, GenTree* tree);
9383 void cTrees(Compiler* comp);
9384 void cEH(Compiler* comp);
9385 void cVar(Compiler* comp, unsigned lclNum);
9386 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9387 void cVars(Compiler* comp);
9388 void cVarsFinal(Compiler* comp);
9389 void cBlockPreds(Compiler* comp, BasicBlock* block);
9390 void cReach(Compiler* comp);
9391 void cDoms(Compiler* comp);
9392 void cLiveness(Compiler* comp);
9393 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9395 void cFuncIR(Compiler* comp);
9396 void cBlockIR(Compiler* comp, BasicBlock* block);
9397 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9398 void cTreeIR(Compiler* comp, GenTree* tree);
9399 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9400 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9401 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9402 int cOperandIR(Compiler* comp, GenTree* operand);
9403 int cLeafIR(Compiler* comp, GenTree* tree);
9404 int cIndirIR(Compiler* comp, GenTree* tree);
9405 int cListIR(Compiler* comp, GenTree* list);
9406 int cSsaNumIR(Compiler* comp, GenTree* tree);
9407 int cValNumIR(Compiler* comp, GenTree* tree);
9408 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9410 void dBlock(BasicBlock* block);
9413 void dTree(GenTree* tree);
9416 void dVar(unsigned lclNum);
9417 void dVarDsc(LclVarDsc* varDsc);
9420 void dBlockPreds(BasicBlock* block);
9424 void dCVarSet(VARSET_VALARG_TP vars);
9426 void dVarSet(VARSET_VALARG_TP vars);
9427 void dRegMask(regMaskTP mask);
9430 void dBlockIR(BasicBlock* block);
9431 void dTreeIR(GenTree* tree);
9432 void dLoopIR(Compiler::LoopDsc* loop);
9433 void dLoopNumIR(unsigned loopNum);
9434 int dTabStopIR(int curr, int tabstop);
9435 int dTreeTypeIR(GenTree* tree);
9436 int dTreeKindsIR(GenTree* tree);
9437 int dTreeFlagsIR(GenTree* tree);
9438 int dOperandIR(GenTree* operand);
9439 int dLeafIR(GenTree* tree);
9440 int dIndirIR(GenTree* tree);
9441 int dListIR(GenTree* list);
9442 int dSsaNumIR(GenTree* tree);
9443 int dValNumIR(GenTree* tree);
9444 int dDependsIR(GenTree* comma);
9447 GenTree* dFindTree(GenTree* tree, unsigned id);
9448 GenTree* dFindTree(unsigned id);
9449 GenTreeStmt* dFindStmt(unsigned id);
9450 BasicBlock* dFindBlock(unsigned bbNum);
9454 #include "compiler.hpp" // All the shared inline functions
9456 /*****************************************************************************/
9457 #endif //_COMPILER_H_
9458 /*****************************************************************************/