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, measureIR) 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[];
955 static bool PhaseReportsIRSize[];
957 unsigned m_byteCodeBytes;
958 unsigned __int64 m_totalCycles;
959 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
960 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
961 #if MEASURE_CLRAPI_CALLS
962 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
963 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
966 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
968 // For better documentation, we call EndPhase on
969 // non-leaf phases. We should also call EndPhase on the
970 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
971 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
972 // We add all such "redundant end phase" intervals to this variable below; we print
973 // it out in a report, so we can verify that it is, indeed, very small. If it ever
974 // isn't, this means that we're doing something significant between the end of the last
975 // declared subphase and the end of its parent.
976 unsigned __int64 m_parentPhaseEndSlop;
979 #if MEASURE_CLRAPI_CALLS
980 // The following measures the time spent inside each individual CLR API call.
981 unsigned m_allClrAPIcalls;
982 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
983 unsigned __int64 m_allClrAPIcycles;
984 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
985 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
986 #endif // MEASURE_CLRAPI_CALLS
988 CompTimeInfo(unsigned byteCodeBytes);
992 #ifdef FEATURE_JIT_METHOD_PERF
994 #if MEASURE_CLRAPI_CALLS
995 struct WrapICorJitInfo;
998 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
999 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1000 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1001 // threads, so it is protected by a lock.
1002 // This class is intended to be used as a singleton type, with only a single instance.
1003 class CompTimeSummaryInfo
1005 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1006 static CritSecObject s_compTimeSummaryLock;
1010 CompTimeInfo m_total;
1011 CompTimeInfo m_maximum;
1013 int m_numFilteredMethods;
1014 CompTimeInfo m_filtered;
1016 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1017 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1018 // If any OS interaction fails, returns 0.0.
1019 double CyclesPerSecond();
1021 // This can use what ever data you want to determine if the value to be added
1022 // belongs in the filtered section (it's always included in the unfiltered section)
1023 bool IncludedInFilteredData(CompTimeInfo& info);
1026 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1027 static CompTimeSummaryInfo s_compTimeSummary;
1029 CompTimeSummaryInfo()
1030 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1034 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1035 // This is thread safe.
1036 void AddInfo(CompTimeInfo& info, bool includePhases);
1038 // Print the summary information to "f".
1039 // This is not thread-safe; assumed to be called by only one thread.
1040 void Print(FILE* f);
1043 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1044 // and when the current phase started. This is intended to be part of a Compilation object. This is
1045 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1049 unsigned __int64 m_start; // Start of the compilation.
1050 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1051 #if MEASURE_CLRAPI_CALLS
1052 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1053 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1054 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1055 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1056 static double s_cyclesPerSec; // Cached for speedier measurements
1059 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1061 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1063 static CritSecObject s_csvLock; // Lock to protect the time log file.
1064 void PrintCsvMethodStats(Compiler* comp);
1067 void* operator new(size_t);
1068 void* operator new[](size_t);
1069 void operator delete(void*);
1070 void operator delete[](void*);
1073 // Initialized the timer instance
1074 JitTimer(unsigned byteCodeSize);
1076 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1078 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1081 static void PrintCsvHeader();
1083 // Ends the current phase (argument is for a redundant check).
1084 void EndPhase(Compiler* compiler, Phases phase);
1086 #if MEASURE_CLRAPI_CALLS
1087 // Start and end a timed CLR API call.
1088 void CLRApiCallEnter(unsigned apix);
1089 void CLRApiCallLeave(unsigned apix);
1090 #endif // MEASURE_CLRAPI_CALLS
1092 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1093 // and adds it to "sum".
1094 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1096 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1097 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1098 // "m_info" to true.
1099 bool GetThreadCycles(unsigned __int64* cycles)
1101 bool res = CycleTimer::GetThreadCyclesS(cycles);
1104 m_info.m_timerFailure = true;
1109 #endif // FEATURE_JIT_METHOD_PERF
1111 //------------------- Function/Funclet info -------------------------------
1112 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1114 FUNC_ROOT, // The main/root function (always id==0)
1115 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1116 FUNC_FILTER, // a funclet associated with an EH filter
1119 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1126 BYTE funFlags; // Currently unused, just here for padding
1127 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1128 // funclet. It is only valid if funKind field indicates this is a
1129 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1131 #if defined(_TARGET_AMD64_)
1133 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1134 emitLocation* startLoc;
1135 emitLocation* endLoc;
1136 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1137 emitLocation* coldEndLoc;
1138 UNWIND_INFO unwindHeader;
1139 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1140 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1141 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1142 unsigned unwindCodeSlot;
1144 #ifdef UNIX_AMD64_ABI
1145 jitstd::vector<CFI_CODE>* cfiCodes;
1146 #endif // UNIX_AMD64_ABI
1148 #elif defined(_TARGET_ARMARCH_)
1150 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1151 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1152 // Note: we only have a pointer here instead of the actual object,
1153 // to save memory in the JIT case (compared to the NGEN case),
1154 // where we don't have any cold section.
1155 // Note 2: we currently don't support hot/cold splitting in functions
1156 // with EH, so uwiCold will be NULL for all funclets.
1158 #endif // _TARGET_ARMARCH_
1160 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1161 // that isn't shared between the main function body and funclets.
1164 struct fgArgTabEntry
1167 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1170 otherRegNum = REG_NA;
1171 isStruct = false; // is this a struct arg
1173 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1175 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1177 // it will point at the actual argument in the gtCallLateArgs list.
1178 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1180 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1182 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1184 unsigned numRegs; // Count of number of registers that this argument uses
1186 // A slot is a pointer sized region in the OutArg area.
1187 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1188 unsigned numSlots; // Count of number of slots that this argument uses
1190 unsigned alignment; // 1 or 2 (slots/registers)
1191 unsigned lateArgInx; // index into gtCallLateArgs list
1192 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1193 #if defined(UNIX_X86_ABI)
1194 unsigned padStkAlign; // Count of number of padding slots for stack alignment. For each Call, only the first
1195 // argument may have a value to emit "sub esp, n" to adjust the stack before pushing
1199 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1200 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1201 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1202 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1203 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1204 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1205 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1206 // previous arguments.
1207 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1208 // to be on the stack despite its arg list position.
1210 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1211 bool isStruct : 1; // True if this is a struct arg
1213 regNumber otherRegNum; // The (second) register to use when passing this argument.
1215 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1216 #elif defined(_TARGET_X86_)
1217 __declspec(property(get = getIsStruct)) bool isStruct;
1220 return varTypeIsStruct(node);
1222 #endif // _TARGET_X86_
1225 void SetIsHfaRegArg(bool hfaRegArg)
1227 isHfaRegArg = hfaRegArg;
1230 void SetIsBackFilled(bool backFilled)
1232 isBackFilled = backFilled;
1235 bool IsBackFilled() const
1237 return isBackFilled;
1239 #else // !_TARGET_ARM_
1240 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1242 void SetIsHfaRegArg(bool hfaRegArg)
1246 void SetIsBackFilled(bool backFilled)
1250 bool IsBackFilled() const
1254 #endif // !_TARGET_ARM_
1260 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1262 //-------------------------------------------------------------------------
1264 // The class fgArgInfo is used to handle the arguments
1265 // when morphing a GT_CALL node.
1270 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1271 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1272 unsigned argCount; // Updatable arg count value
1273 unsigned nextSlotNum; // Updatable slot count value
1274 unsigned stkLevel; // Stack depth when we make this call (for x86)
1275 #if defined(UNIX_X86_ABI)
1276 unsigned padStkAlign; // Count of number of padding slots for stack alignment. This value is used to turn back
1277 // stack pointer before it was adjusted after each Call
1280 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1281 bool hasRegArgs; // true if we have one or more register arguments
1282 bool hasStackArgs; // true if we have one or more stack arguments
1283 bool argsComplete; // marker for state
1284 bool argsSorted; // marker for state
1285 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1288 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1291 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1292 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1294 fgArgTabEntryPtr AddRegArg(
1295 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1297 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1298 fgArgTabEntryPtr AddRegArg(
1305 const bool isStruct,
1306 const regNumber otherRegNum = REG_NA,
1307 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1308 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1310 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1314 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1316 void RemorphReset();
1317 fgArgTabEntryPtr RemorphRegArg(
1318 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1320 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1322 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1324 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1326 void ArgsComplete();
1328 #if defined(UNIX_X86_ABI)
1329 void ArgsAlignPadding();
1334 void EvalArgsToTemps();
1336 void RecordStkLevel(unsigned stkLvl);
1337 unsigned RetrieveStkLevel();
1343 fgArgTabEntryPtr* ArgTable()
1347 unsigned GetNextSlotNum()
1351 #if defined(UNIX_X86_ABI)
1352 unsigned GetPadStackAlign()
1363 return hasStackArgs;
1365 bool AreArgsComplete() const
1367 return argsComplete;
1370 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1371 GenTreePtr GetLateArg(unsigned argIndex);
1375 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1376 // We have the ability to mark source expressions with "Test Labels."
1377 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1378 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1380 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1383 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1384 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1385 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1386 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1387 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1390 struct TestLabelAndNum
1395 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1400 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1402 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1405 // This class implements the "IAllocator" interface, so that we can use
1406 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1408 class CompAllocator : public IAllocator
1411 #if MEASURE_MEM_ALLOC
1415 CompAllocator(Compiler* comp, CompMemKind cmk)
1417 #if MEASURE_MEM_ALLOC
1423 inline void* Alloc(size_t sz);
1425 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1427 // For the compiler's no-release allocator, free operations are no-ops.
1434 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1435 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1437 XX The big guy. The sections are currently organized as : XX
1439 XX o GenTree and BasicBlock XX
1451 XX o PrologScopeInfo XX
1452 XX o CodeGenerator XX
1457 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1458 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1463 friend class emitter;
1464 friend class UnwindInfo;
1465 friend class UnwindFragmentInfo;
1466 friend class UnwindEpilogInfo;
1467 friend class JitTimer;
1468 friend class LinearScan;
1469 friend class fgArgInfo;
1470 friend class Rationalizer;
1472 friend class Lowering;
1473 friend class CSE_DataFlow;
1474 friend class CSE_Heuristic;
1475 friend class CodeGenInterface;
1476 friend class CodeGen;
1477 friend class LclVarDsc;
1478 friend class TempDsc;
1480 friend class ObjectAllocator;
1482 #ifndef _TARGET_64BIT_
1483 friend class DecomposeLongs;
1484 #endif // !_TARGET_64BIT_
1487 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1488 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1490 XX Misc structs definitions XX
1492 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1493 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1497 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1516 bool dumpIRDataflow;
1517 bool dumpIRBlockHeaders;
1519 LPCWSTR dumpIRPhase;
1520 LPCWSTR dumpIRFormat;
1522 bool shouldUseVerboseTrees();
1523 bool asciiTrees; // If true, dump trees using only ASCII characters
1524 bool shouldDumpASCIITrees();
1525 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1526 bool shouldUseVerboseSsa();
1527 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1528 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1530 const char* VarNameToStr(VarName name)
1535 DWORD expensiveDebugCheckLevel;
1538 #if FEATURE_MULTIREG_RET
1539 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1540 #endif // FEATURE_MULTIREG_RET
1543 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1544 #endif // ARM_SOFTFP
1546 //-------------------------------------------------------------------------
1547 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1548 // HFAs are one to four element structs where each element is the same
1549 // type, either all float or all double. They are treated specially
1550 // in the ARM Procedure Call Standard, specifically, they are passed in
1551 // floating-point registers instead of the general purpose registers.
1554 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1555 bool IsHfa(GenTreePtr tree);
1557 var_types GetHfaType(GenTreePtr tree);
1558 unsigned GetHfaCount(GenTreePtr tree);
1560 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1561 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1563 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1564 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1566 //-------------------------------------------------------------------------
1567 // The following is used for validating format of EH table
1571 typedef struct EHNodeDsc* pEHNodeDsc;
1573 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1574 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1587 EHBlockType ehnBlockType; // kind of EH block
1588 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1589 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1590 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1592 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1593 pEHNodeDsc ehnChild; // leftmost nested block
1595 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1596 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1598 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1599 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1601 inline void ehnSetTryNodeType()
1603 ehnBlockType = TryNode;
1605 inline void ehnSetFilterNodeType()
1607 ehnBlockType = FilterNode;
1609 inline void ehnSetHandlerNodeType()
1611 ehnBlockType = HandlerNode;
1613 inline void ehnSetFinallyNodeType()
1615 ehnBlockType = FinallyNode;
1617 inline void ehnSetFaultNodeType()
1619 ehnBlockType = FaultNode;
1622 inline BOOL ehnIsTryBlock()
1624 return ehnBlockType == TryNode;
1626 inline BOOL ehnIsFilterBlock()
1628 return ehnBlockType == FilterNode;
1630 inline BOOL ehnIsHandlerBlock()
1632 return ehnBlockType == HandlerNode;
1634 inline BOOL ehnIsFinallyBlock()
1636 return ehnBlockType == FinallyNode;
1638 inline BOOL ehnIsFaultBlock()
1640 return ehnBlockType == FaultNode;
1643 // returns true if there is any overlap between the two nodes
1644 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1646 if (node1->ehnStartOffset < node2->ehnStartOffset)
1648 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1652 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1656 // fails with BADCODE if inner is not completely nested inside outer
1657 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1659 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1663 //-------------------------------------------------------------------------
1664 // Exception handling functions
1667 #if !FEATURE_EH_FUNCLETS
1669 bool ehNeedsShadowSPslots()
1671 return (info.compXcptnsCount || opts.compDbgEnC);
1674 // 0 for methods with no EH
1675 // 1 for methods with non-nested EH, or where only the try blocks are nested
1676 // 2 for a method with a catch within a catch
1678 unsigned ehMaxHndNestingCount;
1680 #endif // !FEATURE_EH_FUNCLETS
1682 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1683 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1685 bool bbInCatchHandlerILRange(BasicBlock* blk);
1686 bool bbInFilterILRange(BasicBlock* blk);
1687 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1688 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1689 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1690 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1691 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1693 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1694 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1696 // Returns true if "block" is the start of a try region.
1697 bool bbIsTryBeg(BasicBlock* block);
1699 // Returns true if "block" is the start of a handler or filter region.
1700 bool bbIsHandlerBeg(BasicBlock* block);
1702 // Returns true iff "block" is where control flows if an exception is raised in the
1703 // try region, and sets "*regionIndex" to the index of the try for the handler.
1704 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1705 // block of the filter, but not for the filter's handler.
1706 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1708 bool ehHasCallableHandlers();
1710 // Return the EH descriptor for the given region index.
1711 EHblkDsc* ehGetDsc(unsigned regionIndex);
1713 // Return the EH index given a region descriptor.
1714 unsigned ehGetIndex(EHblkDsc* ehDsc);
1716 // Return the EH descriptor index of the enclosing try, for the given region index.
1717 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1719 // Return the EH descriptor index of the enclosing handler, for the given region index.
1720 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1722 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1723 // block is not in a 'try' region).
1724 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1726 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1727 // if this block is not in a filter or handler region).
1728 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1730 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1731 // nullptr if this block's exceptions propagate to caller).
1732 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1734 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1735 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1736 bool ehIsBlockEHLast(BasicBlock* block);
1738 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1740 // Return the region index of the most nested EH region this block is in.
1741 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1743 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1744 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1746 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1747 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1748 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1749 // (It can never be a filter.)
1750 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1752 // A block has been deleted. Update the EH table appropriately.
1753 void ehUpdateForDeletedBlock(BasicBlock* block);
1755 // Determine whether a block can be deleted while preserving the EH normalization rules.
1756 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1758 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1759 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1761 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1762 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1763 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1764 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1765 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1766 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1767 // lives in a filter.)
1768 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1770 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1771 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1772 // (nullptr if the last block is the last block in the program).
1773 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1774 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1777 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1778 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1779 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1782 #if FEATURE_EH_FUNCLETS
1783 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1784 // if there is a filter that protects a region with a nested EH clause (such as a
1785 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1786 // genFuncletProlog() for more details. However, the VM seems to use it for more
1787 // purposes, maybe including debugging. Until we are sure otherwise, always create
1788 // a PSPSym for functions with any EH.
1789 bool ehNeedsPSPSym() const
1793 #else // _TARGET_X86_
1794 return compHndBBtabCount > 0;
1795 #endif // _TARGET_X86_
1798 bool ehAnyFunclets(); // Are there any funclets in this function?
1799 unsigned ehFuncletCount(); // Return the count of funclets in the function
1801 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1802 #else // !FEATURE_EH_FUNCLETS
1803 bool ehAnyFunclets()
1807 unsigned ehFuncletCount()
1812 unsigned bbThrowIndex(BasicBlock* blk)
1814 return blk->bbTryIndex;
1815 } // Get the index to use as the cache key for sharing throw blocks
1816 #endif // !FEATURE_EH_FUNCLETS
1818 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1819 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1820 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1821 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1822 // convenient to also consider it a predecessor.)
1823 flowList* BlockPredsWithEH(BasicBlock* blk);
1825 // This table is useful for memoization of the method above.
1826 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1828 BlockToFlowListMap* m_blockToEHPreds;
1829 BlockToFlowListMap* GetBlockToEHPreds()
1831 if (m_blockToEHPreds == nullptr)
1833 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1835 return m_blockToEHPreds;
1838 void* ehEmitCookie(BasicBlock* block);
1839 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1841 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1843 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1845 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1847 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1849 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1851 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1853 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1855 void fgAllocEHTable();
1857 void fgRemoveEHTableEntry(unsigned XTnum);
1859 #if FEATURE_EH_FUNCLETS
1861 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1863 #endif // FEATURE_EH_FUNCLETS
1867 #endif // !FEATURE_EH
1869 void fgSortEHTable();
1871 // Causes the EH table to obey some well-formedness conditions, by inserting
1872 // empty BB's when necessary:
1873 // * No block is both the first block of a handler and the first block of a try.
1874 // * No block is the first block of multiple 'try' regions.
1875 // * No block is the last block of multiple EH regions.
1876 void fgNormalizeEH();
1877 bool fgNormalizeEHCase1();
1878 bool fgNormalizeEHCase2();
1879 bool fgNormalizeEHCase3();
1882 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1883 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1884 void fgVerifyHandlerTab();
1885 void fgDispHandlerTab();
1888 bool fgNeedToSortEHTable;
1890 void verInitEHTree(unsigned numEHClauses);
1891 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1892 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1893 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1894 void verCheckNestingLevel(EHNodeDsc* initRoot);
1897 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1898 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1900 XX GenTree and BasicBlock XX
1902 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1904 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1905 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1908 // Functions to create nodes
1909 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1912 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1914 // For binary opers.
1915 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1917 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1919 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1920 var_types type = TYP_I_IMPL,
1921 GenTreePtr op1 = nullptr,
1922 GenTreePtr op2 = nullptr);
1924 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1926 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1928 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1930 GenTreePtr gtNewJmpTableNode();
1931 GenTreePtr gtNewIconHandleNode(
1932 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1934 unsigned gtTokenToIconFlags(unsigned token);
1936 GenTreePtr gtNewIconEmbHndNode(void* value,
1939 unsigned handle1 = 0,
1940 void* handle2 = nullptr,
1941 void* compileTimeHandle = nullptr);
1943 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1944 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1945 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1946 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1948 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1950 GenTreePtr gtNewLconNode(__int64 value);
1952 GenTreePtr gtNewDconNode(double value);
1954 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1956 GenTreePtr gtNewZeroConNode(var_types type);
1958 GenTreePtr gtNewOneConNode(var_types type);
1961 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1962 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1965 GenTreeBlk* gtNewBlkOpNode(
1966 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1968 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1971 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1974 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1975 void gtSetObjGcInfo(GenTreeObj* objNode);
1976 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1977 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1979 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1981 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1983 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1984 CORINFO_METHOD_HANDLE handle,
1986 GenTreeArgList* args,
1987 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1989 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1991 GenTreeArgList* args,
1992 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1994 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1997 GenTreeArgList* args = nullptr);
1999 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2002 GenTreeSIMD* gtNewSIMDNode(
2003 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2004 GenTreeSIMD* gtNewSIMDNode(var_types type,
2007 SIMDIntrinsicID simdIntrinsicID,
2010 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2013 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2014 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2015 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2017 GenTreePtr gtNewCodeRef(BasicBlock* block);
2019 GenTreePtr gtNewFieldRef(
2020 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2022 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2024 GenTreeArgList* gtNewArgList(GenTreePtr op);
2025 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2026 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2028 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2029 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2030 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2031 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2033 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2035 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2037 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2038 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2039 CORINFO_ACCESS_FLAGS access,
2040 CORINFO_FIELD_INFO* pFieldInfo,
2042 CORINFO_CLASS_HANDLE structType,
2045 GenTreePtr gtNewNothingNode();
2047 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2049 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2051 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2053 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2055 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2057 //------------------------------------------------------------------------
2058 // Other GenTree functions
2060 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2062 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2063 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2064 // IntCnses with value `deepVarVal`.
2065 GenTreePtr gtCloneExpr(
2066 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2068 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2069 // `varNum` to int constants with value `varVal`.
2070 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2072 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2075 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2077 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2079 // Returns "true" iff the complexity (not formally defined, but first interpretation
2080 // is #of nodes in subtree) of "tree" is greater than "limit".
2081 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2082 // before they have been set.)
2083 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2085 bool gtCompareTree(GenTree* op1, GenTree* op2);
2087 GenTreePtr gtReverseCond(GenTree* tree);
2089 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2091 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2093 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2095 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2098 unsigned gtHashValue(GenTree* tree);
2100 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2103 void gtPrepareCost(GenTree* tree);
2104 bool gtIsLikelyRegVar(GenTree* tree);
2106 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2108 // Returns true iff the secondNode can be swapped with firstNode.
2109 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2111 unsigned gtSetEvalOrder(GenTree* tree);
2113 #if FEATURE_STACK_FP_X87
2115 void gtComputeFPlvls(GenTreePtr tree);
2116 #endif // FEATURE_STACK_FP_X87
2118 void gtSetStmtInfo(GenTree* stmt);
2120 // Returns "true" iff "node" has any of the side effects in "flags".
2121 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2123 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2124 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2126 // Appends 'expr' in front of 'list'
2127 // 'list' will typically start off as 'nullptr'
2128 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2129 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2131 void gtExtractSideEffList(GenTreePtr expr,
2133 unsigned flags = GTF_SIDE_EFFECT,
2134 bool ignoreRoot = false);
2136 GenTreePtr gtGetThisArg(GenTreePtr call);
2138 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2139 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2140 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2141 // the given "fldHnd", is such an object pointer.
2142 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2144 // Return true if call is a recursive call; return false otherwise.
2145 // Note when inlining, this looks for calls back to the root method.
2146 bool gtIsRecursiveCall(GenTreeCall* call)
2148 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2151 //-------------------------------------------------------------------------
2153 GenTreePtr gtFoldExpr(GenTreePtr tree);
2156 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2157 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2158 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2159 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2160 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2161 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2162 // optimizations for now.
2163 __attribute__((optnone))
2165 gtFoldExprConst(GenTreePtr tree);
2166 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2167 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2169 //-------------------------------------------------------------------------
2170 // Get the handle, if any.
2171 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2172 // Get the handle, and assert if not found.
2173 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2175 //-------------------------------------------------------------------------
2176 // Functions to display the trees
2179 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2181 void gtDispVN(GenTreePtr tree);
2182 void gtDispConst(GenTreePtr tree);
2183 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2184 void gtDispNodeName(GenTreePtr tree);
2185 void gtDispRegVal(GenTreePtr tree);
2197 void gtDispChild(GenTreePtr child,
2198 IndentStack* indentStack,
2200 __in_opt const char* msg = nullptr,
2201 bool topOnly = false);
2202 void gtDispTree(GenTreePtr tree,
2203 IndentStack* indentStack = nullptr,
2204 __in_opt const char* msg = nullptr,
2205 bool topOnly = false,
2206 bool isLIR = false);
2207 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2208 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2209 char* gtGetLclVarName(unsigned lclNum);
2210 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2211 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2212 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2213 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2214 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2215 void gtDispFieldSeq(FieldSeqNode* pfsn);
2217 void gtDispRange(LIR::ReadOnlyRange const& range);
2219 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2221 void gtDispLIRNode(GenTree* node);
2233 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2234 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2237 static fgWalkPreFn gtAssertColonCond;
2239 static fgWalkPreFn gtMarkColonCond;
2240 static fgWalkPreFn gtClearColonCond;
2242 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2243 bool gtHasCatchArg(GenTreePtr tree);
2244 bool gtHasUnmanagedCall(GenTreePtr tree);
2246 typedef ArrayStack<GenTree*> GenTreeStack;
2248 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2249 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2251 //=========================================================================
2252 // BasicBlock functions
2254 // This is a debug flag we will use to assert when creating block during codegen
2255 // as this interferes with procedure splitting. If you know what you're doing, set
2256 // it to true before creating the block. (DEBUG only)
2257 bool fgSafeBasicBlockCreation;
2260 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2263 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2264 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2268 XX The variables to be used by the code generator. XX
2270 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2271 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2275 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2276 // be placed in the stack frame and it's fields must be laid out sequentially.
2278 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2279 // a local variable that can be enregistered or placed in the stack frame.
2280 // The fields do not need to be laid out sequentially
2282 enum lvaPromotionType
2284 PROMOTION_TYPE_NONE, // The struct local is not promoted
2285 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2286 // and its field locals are independent of its parent struct local.
2287 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2288 // but its field locals depend on its parent struct local.
2291 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2292 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2294 /*****************************************************************************/
2296 enum FrameLayoutState
2299 INITIAL_FRAME_LAYOUT,
2300 PRE_REGALLOC_FRAME_LAYOUT,
2301 REGALLOC_FRAME_LAYOUT,
2302 TENTATIVE_FRAME_LAYOUT,
2307 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2308 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2309 bool lvaSortAgain; // true: We need to sort the lvaTable
2310 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2311 unsigned lvaCount; // total number of locals
2313 unsigned lvaRefCount; // total number of references to locals
2314 LclVarDsc* lvaTable; // variable descriptor table
2315 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2317 LclVarDsc** lvaRefSorted; // table sorted by refcount
2319 unsigned short lvaTrackedCount; // actual # of locals being tracked
2320 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2322 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2323 // Only for AMD64 System V cache the first caller stack homed argument.
2324 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2325 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2328 VARSET_TP lvaTrackedVars; // set of tracked variables
2330 #ifndef _TARGET_64BIT_
2331 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2333 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2335 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2336 // It that changes, this changes. VarSets from different epochs
2337 // cannot be meaningfully combined.
2339 unsigned GetCurLVEpoch()
2344 // reverse map of tracked number to var number
2345 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2347 #ifdef LEGACY_BACKEND
2348 // variable interference graph
2349 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2352 // variable preference graph
2353 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2357 // # of procs compiled a with double-aligned stack
2358 static unsigned s_lvaDoubleAlignedProcsCount;
2362 // Getters and setters for address-exposed and do-not-enregister local var properties.
2363 bool lvaVarAddrExposed(unsigned varNum);
2364 void lvaSetVarAddrExposed(unsigned varNum);
2365 bool lvaVarDoNotEnregister(unsigned varNum);
2367 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2368 enum DoNotEnregisterReason
2373 DNER_VMNeedsStackAddr,
2374 DNER_LiveInOutOfHandler,
2375 DNER_LiveAcrossUnmanagedCall,
2376 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2377 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2378 #ifdef JIT32_GCENCODER
2383 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2385 unsigned lvaVarargsHandleArg;
2387 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2389 #endif // _TARGET_X86_
2391 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2392 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2393 #if FEATURE_FIXED_OUT_ARGS
2394 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2396 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2397 // that tracks whether the lock has been taken
2399 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2400 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2401 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2403 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2404 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2406 #if FEATURE_FIXED_OUT_ARGS
2407 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2408 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2409 #endif // FEATURE_FIXED_OUT_ARGS
2412 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2413 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2414 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2415 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2416 // this variable to be this scratch word whenever struct promotion occurs.
2417 unsigned lvaPromotedStructAssemblyScratchVar;
2418 #endif // _TARGET_ARM_
2421 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2422 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2425 bool lvaGenericsContextUsed;
2427 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2428 // CORINFO_GENERICS_CTXT_FROM_THIS?
2429 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2431 //-------------------------------------------------------------------------
2432 // All these frame offsets are inter-related and must be kept in sync
2434 #if !FEATURE_EH_FUNCLETS
2435 // This is used for the callable handlers
2436 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2437 #endif // FEATURE_EH_FUNCLETS
2439 unsigned lvaCachedGenericContextArgOffs;
2440 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2443 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2445 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2447 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2448 // after the reg predict we will use a computed maxTmpSize
2449 // which is based upon the number of spill temps predicted by reg predict
2450 // All this is necessary because if we under-estimate the size of the spill
2451 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2453 // Pre codegen max spill temp size.
2454 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2456 //-------------------------------------------------------------------------
2458 unsigned lvaGetMaxSpillTempSize();
2460 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2461 #endif // _TARGET_ARM_
2462 void lvaAssignFrameOffsets(FrameLayoutState curState);
2463 void lvaFixVirtualFrameOffsets();
2465 #ifndef LEGACY_BACKEND
2466 void lvaUpdateArgsWithInitialReg();
2467 #endif // !LEGACY_BACKEND
2469 void lvaAssignVirtualFrameOffsetsToArgs();
2470 #ifdef UNIX_AMD64_ABI
2471 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2472 #else // !UNIX_AMD64_ABI
2473 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2474 #endif // !UNIX_AMD64_ABI
2475 void lvaAssignVirtualFrameOffsetsToLocals();
2476 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2477 #ifdef _TARGET_AMD64_
2478 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2479 bool lvaIsCalleeSavedIntRegCountEven();
2481 void lvaAlignFrame();
2482 void lvaAssignFrameOffsetsToPromotedStructs();
2483 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2486 void lvaDumpRegLocation(unsigned lclNum);
2487 void lvaDumpFrameLocation(unsigned lclNum);
2488 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2489 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2490 // layout state defined by lvaDoneFrameLayout
2493 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2494 // to avoid bugs from borderline cases.
2495 #define MAX_FrameSize 0x3FFFFFFF
2496 void lvaIncrementFrameSize(unsigned size);
2498 unsigned lvaFrameSize(FrameLayoutState curState);
2500 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2501 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2503 // Returns the caller-SP-relative offset for the local variable "varNum."
2504 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2506 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2507 int lvaGetSPRelativeOffset(unsigned varNum);
2509 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2510 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2512 //------------------------ For splitting types ----------------------------
2514 void lvaInitTypeRef();
2516 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2517 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2518 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2519 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2520 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2521 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2523 void lvaInitVarDsc(LclVarDsc* varDsc,
2525 CorInfoType corInfoType,
2526 CORINFO_CLASS_HANDLE typeHnd,
2527 CORINFO_ARG_LIST_HANDLE varList,
2528 CORINFO_SIG_INFO* varSig);
2530 static unsigned lvaTypeRefMask(var_types type);
2532 var_types lvaGetActualType(unsigned lclNum);
2533 var_types lvaGetRealType(unsigned lclNum);
2535 //-------------------------------------------------------------------------
2539 unsigned lvaLclSize(unsigned varNum);
2540 unsigned lvaLclExactSize(unsigned varNum);
2542 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2544 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2545 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2546 // the return result.
2547 bool lvaLclVarRefsAccum(
2548 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2550 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2551 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2552 // and (destructively) unions "trkedVars" into "*result".
2553 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2555 ALLVARSET_VALARG_TP allVars,
2556 VARSET_VALARG_TP trkdVars);
2558 bool lvaHaveManyLocals() const;
2560 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2561 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2562 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2565 void lvaSortByRefCount();
2566 void lvaDumpRefCounts();
2568 void lvaMarkLocalVars(BasicBlock* block);
2570 void lvaMarkLocalVars(); // Local variable ref-counting
2572 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2574 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2576 static fgWalkPreFn lvaIncRefCntsCB;
2577 void lvaIncRefCnts(GenTreePtr tree);
2579 static fgWalkPreFn lvaDecRefCntsCB;
2580 void lvaDecRefCnts(GenTreePtr tree);
2581 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2582 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2583 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2586 struct lvaStressLclFldArgs
2588 Compiler* m_pCompiler;
2592 static fgWalkPreFn lvaStressLclFldCB;
2593 void lvaStressLclFld();
2595 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2596 void lvaDispVarSet(VARSET_VALARG_TP set);
2601 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2603 int lvaFrameAddress(int varNum, bool* pFPbased);
2606 bool lvaIsParameter(unsigned varNum);
2607 bool lvaIsRegArgument(unsigned varNum);
2608 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2609 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2610 // that writes to arg0
2612 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2613 // (this is an overload of lvIsTemp because there are no temp parameters).
2614 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2615 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2616 bool lvaIsImplicitByRefLocal(unsigned varNum)
2618 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2619 LclVarDsc* varDsc = &(lvaTable[varNum]);
2620 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2622 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2625 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2629 // Returns true if this local var is a multireg struct
2630 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2632 // If the class is a TYP_STRUCT, get/set a class handle describing it
2634 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2635 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2637 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2639 // Info about struct fields
2640 struct lvaStructFieldInfo
2642 CORINFO_FIELD_HANDLE fldHnd;
2643 unsigned char fldOffset;
2644 unsigned char fldOrdinal;
2647 CORINFO_CLASS_HANDLE fldTypeHnd;
2650 // Info about struct to be promoted.
2651 struct lvaStructPromotionInfo
2653 CORINFO_CLASS_HANDLE typeHnd;
2655 bool requiresScratchVar;
2658 unsigned char fieldCnt;
2659 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2661 lvaStructPromotionInfo()
2662 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2667 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2668 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2669 lvaStructPromotionInfo* StructPromotionInfo,
2671 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2672 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2673 #if !defined(_TARGET_64BIT_)
2674 void lvaPromoteLongVars();
2675 #endif // !defined(_TARGET_64BIT_)
2676 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2677 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2678 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2679 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2680 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2681 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2682 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2684 #if defined(FEATURE_SIMD)
2685 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2687 assert(varDsc->lvType == TYP_SIMD12);
2688 assert(varDsc->lvExactSize == 12);
2690 #if defined(_TARGET_64BIT_)
2691 assert(varDsc->lvSize() == 16);
2693 #else // !defined(_TARGET_64BIT_)
2695 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2696 // already does this calculation. However, we also need to prevent mapping types if the var is a
2697 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2698 // However, we don't know this until late, so we may have already pretended the field is bigger
2700 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2709 #endif // !defined(_TARGET_64BIT_)
2711 #endif // defined(FEATURE_SIMD)
2713 BYTE* lvaGetGcLayout(unsigned varNum);
2714 bool lvaTypeIsGC(unsigned varNum);
2715 unsigned lvaGSSecurityCookie; // LclVar number
2716 bool lvaTempsHaveLargerOffsetThanVars();
2718 unsigned lvaSecurityObject; // variable representing the security object on the stack
2719 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2721 #if FEATURE_EH_FUNCLETS
2722 unsigned lvaPSPSym; // variable representing the PSPSym
2725 InlineInfo* impInlineInfo;
2726 InlineStrategy* m_inlineStrategy;
2728 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2729 Compiler* impInlineRoot();
2731 #if defined(DEBUG) || defined(INLINE_DATA)
2732 unsigned __int64 getInlineCycleCount()
2734 return m_compCycles;
2736 #endif // defined(DEBUG) || defined(INLINE_DATA)
2738 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2739 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2741 //=========================================================================
2743 //=========================================================================
2746 //---------------- Local variable ref-counting ----------------------------
2749 BasicBlock* lvaMarkRefsCurBlock;
2750 GenTreePtr lvaMarkRefsCurStmt;
2752 BasicBlock::weight_t lvaMarkRefsWeight;
2754 static fgWalkPreFn lvaMarkLclRefsCallback;
2755 void lvaMarkLclRefs(GenTreePtr tree);
2757 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2758 PerSsaArray lvMemoryPerSsaData;
2759 unsigned lvMemoryNumSsaNames;
2762 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2763 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2764 // not an SSA variable).
2765 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2767 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2768 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2770 assert(ssaNum < lvMemoryNumSsaNames);
2771 return &lvMemoryPerSsaData.GetRef(ssaNum);
2775 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2776 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2780 XX Imports the given method and converts it to semantic trees XX
2782 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2783 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2789 void impImport(BasicBlock* method);
2791 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2792 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2793 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2794 CORINFO_CLASS_HANDLE impGetStringClass();
2795 CORINFO_CLASS_HANDLE impGetObjectClass();
2797 //=========================================================================
2799 //=========================================================================
2802 //-------------------- Stack manipulation ---------------------------------
2804 unsigned impStkSize; // Size of the full stack
2806 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2808 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2810 struct SavedStack // used to save/restore stack contents.
2812 unsigned ssDepth; // number of values on stack
2813 StackEntry* ssTrees; // saved tree values
2816 bool impIsPrimitive(CorInfoType type);
2817 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2819 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2820 void impPushOnStackNoType(GenTreePtr tree);
2822 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2823 void impPushNullObjRefOnStack();
2824 StackEntry impPopStack();
2825 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2826 GenTreePtr impPopStack(typeInfo& ti);
2827 StackEntry& impStackTop(unsigned n = 0);
2829 void impSaveStackState(SavedStack* savePtr, bool copy);
2830 void impRestoreStackState(SavedStack* savePtr);
2832 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2833 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2834 CORINFO_CALL_INFO* pCallInfo);
2836 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2838 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2840 bool impCanPInvokeInline();
2841 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2842 void impCheckForPInvokeCall(
2843 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2844 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2845 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2847 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2848 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2849 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2851 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2852 CORINFO_METHOD_HANDLE calleeMethodHnd,
2853 CORINFO_CLASS_HANDLE delegateTypeHnd);
2855 var_types impImportCall(OPCODE opcode,
2856 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2857 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2859 GenTreePtr newobjThis,
2861 CORINFO_CALL_INFO* callInfo,
2862 IL_OFFSET rawILOffset);
2864 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2866 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2868 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2871 var_types impImportJitTestLabelMark(int numArgs);
2874 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2876 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2878 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2879 CORINFO_ACCESS_FLAGS access,
2880 CORINFO_FIELD_INFO* pFieldInfo,
2883 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2885 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2887 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2889 void impImportLeave(BasicBlock* block);
2890 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2891 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2892 CORINFO_CLASS_HANDLE clsHnd,
2893 CORINFO_METHOD_HANDLE method,
2894 CORINFO_SIG_INFO* sig,
2898 CorInfoIntrinsics* pIntrinsicID);
2899 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2900 CORINFO_SIG_INFO* sig,
2903 CorInfoIntrinsics intrinsicID);
2904 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2906 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2908 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2909 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2910 CORINFO_THIS_TRANSFORM transform);
2912 //----------------- Manipulating the trees and stmts ----------------------
2914 GenTreePtr impTreeList; // Trees for the BB being imported
2915 GenTreePtr impTreeLast; // The last tree for the current BB
2919 CHECK_SPILL_ALL = -1,
2920 CHECK_SPILL_NONE = -2
2924 void impBeginTreeList();
2925 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2926 void impEndTreeList(BasicBlock* block);
2927 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2928 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2929 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2930 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2931 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2932 void impAssignTempGen(unsigned tmp,
2935 GenTreePtr* pAfterStmt = nullptr,
2936 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2937 BasicBlock* block = nullptr);
2938 void impAssignTempGen(unsigned tmpNum,
2940 CORINFO_CLASS_HANDLE structHnd,
2942 GenTreePtr* pAfterStmt = nullptr,
2943 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2944 BasicBlock* block = nullptr);
2945 GenTreePtr impCloneExpr(GenTreePtr tree,
2947 CORINFO_CLASS_HANDLE structHnd,
2949 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2950 GenTreePtr impAssignStruct(GenTreePtr dest,
2952 CORINFO_CLASS_HANDLE structHnd,
2954 GenTreePtr* pAfterStmt = nullptr,
2955 BasicBlock* block = nullptr);
2956 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2958 CORINFO_CLASS_HANDLE structHnd,
2960 GenTreePtr* pAfterStmt = nullptr,
2961 BasicBlock* block = nullptr);
2963 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2964 CORINFO_CLASS_HANDLE structHnd,
2968 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2969 BYTE* gcLayout = nullptr,
2970 unsigned* numGCVars = nullptr,
2971 var_types* simdBaseType = nullptr);
2973 GenTreePtr impNormStructVal(GenTreePtr structVal,
2974 CORINFO_CLASS_HANDLE structHnd,
2976 bool forceNormalization = false);
2978 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2979 BOOL* pRuntimeLookup = nullptr,
2980 BOOL mustRestoreHandle = FALSE,
2981 BOOL importParent = FALSE);
2983 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2984 BOOL* pRuntimeLookup = nullptr,
2985 BOOL mustRestoreHandle = FALSE)
2987 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2990 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2991 CORINFO_LOOKUP* pLookup,
2993 void* compileTimeHandle);
2995 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2997 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2998 CORINFO_LOOKUP* pLookup,
2999 void* compileTimeHandle);
3001 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3003 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3004 CorInfoHelpFunc helper,
3006 GenTreeArgList* arg = nullptr,
3007 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3009 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3011 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3014 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3015 CORINFO_CLASS_HANDLE typeClass,
3019 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3020 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3021 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3022 static bool IsMathIntrinsic(GenTreePtr tree);
3025 //----------------- Importing the method ----------------------------------
3027 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3030 unsigned impCurOpcOffs;
3031 const char* impCurOpcName;
3032 bool impNestedStackSpill;
3034 // For displaying instrs with generated native code (-n:B)
3035 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3036 void impNoteLastILoffs();
3039 /* IL offset of the stmt currently being imported. It gets set to
3040 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3041 updated at IL offsets for which we have to report mapping info.
3042 It also includes flag bits, so use jitGetILoffs()
3043 to get the actual IL offset value.
3046 IL_OFFSETX impCurStmtOffs;
3047 void impCurStmtOffsSet(IL_OFFSET offs);
3049 void impNoteBranchOffs();
3051 unsigned impInitBlockLineInfo();
3053 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3054 bool impIsThis(GenTreePtr obj);
3055 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3056 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3057 bool impIsAnySTLOC(OPCODE opcode)
3059 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3060 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3063 GenTreeArgList* impPopList(unsigned count,
3065 CORINFO_SIG_INFO* sig,
3066 GenTreeArgList* prefixTree = nullptr);
3068 GenTreeArgList* impPopRevList(unsigned count,
3070 CORINFO_SIG_INFO* sig,
3071 unsigned skipReverseCount = 0);
3074 * Get current IL offset with stack-empty info incoporated
3076 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3078 //---------------- Spilling the importer stack ----------------------------
3084 SavedStack pdSavedStack;
3085 ThisInitState pdThisPtrInit;
3088 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3089 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3091 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3092 ExpandArray<BYTE> impPendingBlockMembers;
3094 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3095 // Operates on the map in the top-level ancestor.
3096 BYTE impGetPendingBlockMember(BasicBlock* blk)
3098 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3101 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3102 // Operates on the map in the top-level ancestor.
3103 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3105 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3108 bool impCanReimport;
3110 bool impSpillStackEntry(unsigned level,
3114 bool bAssertOnRecursion,
3119 void impSpillStackEnsure(bool spillLeaves = false);
3120 void impEvalSideEffects();
3121 void impSpillSpecialSideEff();
3122 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3123 void impSpillValueClasses();
3124 void impSpillEvalStack();
3125 static fgWalkPreFn impFindValueClasses;
3126 void impSpillLclRefs(ssize_t lclNum);
3128 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3130 void impImportBlockCode(BasicBlock* block);
3132 void impReimportMarkBlock(BasicBlock* block);
3133 void impReimportMarkSuccessors(BasicBlock* block);
3135 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3137 void impImportBlockPending(BasicBlock* block);
3139 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3140 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3141 // for the block, but instead, just re-uses the block's existing EntryState.
3142 void impReimportBlockPending(BasicBlock* block);
3144 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3146 void impImportBlock(BasicBlock* block);
3148 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3149 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3150 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3151 // the variables that will be used -- and for all the predecessors of those successors, and the
3152 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3153 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3154 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3155 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3156 // of local variable numbers, so we represent them with the base local variable number), returns that.
3157 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3158 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3159 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3160 // on which kind of member of the clique the block is).
3161 unsigned impGetSpillTmpBase(BasicBlock* block);
3163 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3164 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3165 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3166 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3167 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3168 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3169 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3170 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3171 // successors receive a native int. Similarly float and double are unified to double.
3172 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3173 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3174 // predecessors, so they insert an upcast if needed).
3175 void impReimportSpillClique(BasicBlock* block);
3177 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3178 // block, and represent the predecessor and successor members of the clique currently being computed.
3179 // *** Access to these will need to be locked in a parallel compiler.
3180 ExpandArray<BYTE> impSpillCliquePredMembers;
3181 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3189 // Abstract class for receiving a callback while walking a spill clique
3190 class SpillCliqueWalker
3193 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3196 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3197 class SetSpillTempsBase : public SpillCliqueWalker
3202 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3205 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3208 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3209 class ReimportSpillClique : public SpillCliqueWalker
3214 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3217 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3220 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3221 // predecessor or successor within the spill clique
3222 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3224 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3225 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3226 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3227 void impRetypeEntryStateTemps(BasicBlock* blk);
3229 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3230 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3232 void impPushVar(GenTree* op, typeInfo tiRetVal);
3233 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3234 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3236 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3238 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3239 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3240 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3243 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3246 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3247 struct BlockListNode
3250 BlockListNode* m_next;
3251 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3254 void* operator new(size_t sz, Compiler* comp);
3256 BlockListNode* impBlockListNodeFreeList;
3258 BlockListNode* AllocBlockListNode();
3259 void FreeBlockListNode(BlockListNode* node);
3261 bool impIsValueType(typeInfo* pTypeInfo);
3262 var_types mangleVarArgsType(var_types type);
3265 regNumber getCallArgIntRegister(regNumber floatReg);
3266 regNumber getCallArgFloatRegister(regNumber intReg);
3267 #endif // FEATURE_VARARG
3270 static unsigned jitTotalMethodCompiled;
3274 static LONG jitNestingLevel;
3277 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3279 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3281 // STATIC inlining decision based on the IL code.
3282 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3283 CORINFO_METHOD_INFO* methInfo,
3285 InlineResult* inlineResult);
3287 void impCheckCanInline(GenTreePtr call,
3288 CORINFO_METHOD_HANDLE fncHandle,
3290 CORINFO_CONTEXT_HANDLE exactContextHnd,
3291 InlineCandidateInfo** ppInlineCandidateInfo,
3292 InlineResult* inlineResult);
3294 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3295 GenTreePtr curArgVal,
3297 InlineResult* inlineResult);
3299 void impInlineInitVars(InlineInfo* pInlineInfo);
3301 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3303 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3305 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3307 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3308 GenTreePtr variableBeingDereferenced,
3309 InlArgInfo* inlArgInfo);
3311 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3313 bool impTailCallRetTypeCompatible(var_types callerRetType,
3314 CORINFO_CLASS_HANDLE callerRetTypeClass,
3315 var_types calleeRetType,
3316 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3318 bool impIsTailCallILPattern(bool tailPrefixed,
3320 const BYTE* codeAddrOfNextOpcode,
3321 const BYTE* codeEnd,
3323 bool* IsCallPopRet = nullptr);
3325 bool impIsImplicitTailCallCandidate(
3326 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3329 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3330 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3334 XX Info about the basic-blocks, their contents and the flow analysis XX
3336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3337 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3341 BasicBlock* fgFirstBB; // Beginning of the basic block list
3342 BasicBlock* fgLastBB; // End of the basic block list
3343 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3344 #if FEATURE_EH_FUNCLETS
3345 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3347 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3349 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3350 unsigned fgEdgeCount; // # of control flow edges between the BBs
3351 unsigned fgBBcount; // # of BBs in the method
3353 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3355 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3356 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3357 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3358 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3360 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3361 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3362 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3363 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3364 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3365 // index). The arrays are of size fgBBNumMax + 1.
3366 unsigned* fgDomTreePreOrder;
3367 unsigned* fgDomTreePostOrder;
3369 bool fgBBVarSetsInited;
3371 // Allocate array like T* a = new T[fgBBNumMax + 1];
3372 // Using helper so we don't keep forgetting +1.
3373 template <typename T>
3374 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3376 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3379 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3380 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3381 // cannot be meaningfully combined. Note that new blocks can be created with higher
3382 // block numbers without changing the basic block epoch. These blocks *cannot*
3383 // participate in a block set until the blocks are all renumbered, causing the epoch
3384 // to change. This is useful if continuing to use previous block sets is valuable.
3385 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3386 unsigned fgCurBBEpoch;
3388 unsigned GetCurBasicBlockEpoch()
3390 return fgCurBBEpoch;
3393 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3394 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3395 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3396 unsigned fgCurBBEpochSize;
3398 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3399 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3400 unsigned fgBBSetCountInSizeTUnits;
3402 void NewBasicBlockEpoch()
3404 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3406 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3408 fgCurBBEpochSize = fgBBNumMax + 1;
3409 fgBBSetCountInSizeTUnits =
3410 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3413 // All BlockSet objects are now invalid!
3414 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3415 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3419 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3420 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3421 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3422 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3424 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3425 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3426 // array of size_t bitsets), then print that out.
3427 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3434 void EnsureBasicBlockEpoch()
3436 if (fgCurBBEpochSize != fgBBNumMax + 1)
3438 NewBasicBlockEpoch();
3442 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3443 void fgEnsureFirstBBisScratch();
3444 bool fgFirstBBisScratch();
3445 bool fgBBisScratch(BasicBlock* block);
3447 void fgExtendEHRegionBefore(BasicBlock* block);
3448 void fgExtendEHRegionAfter(BasicBlock* block);
3450 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3452 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3454 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3457 BasicBlock* nearBlk,
3458 bool putInFilter = false,
3459 bool runRarely = false,
3460 bool insertAtEnd = false);
3462 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3464 bool runRarely = false,
3465 bool insertAtEnd = false);
3467 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3469 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3470 BasicBlock* afterBlk,
3471 unsigned xcptnIndex,
3472 bool putInTryRegion);
3474 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3475 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3476 void fgUnlinkBlock(BasicBlock* block);
3478 unsigned fgMeasureIR();
3480 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3481 bool fgMultipleNots;
3484 bool fgModified; // True if the flow graph has been modified recently
3485 bool fgComputePredsDone; // Have we computed the bbPreds list
3486 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3487 bool fgDomsComputed; // Have we computed the dominator sets?
3488 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3490 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3491 bool fgHasPostfix; // any postfix ++/-- found?
3492 unsigned fgIncrCount; // number of increment nodes found
3494 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3498 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3499 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3502 bool fgRemoveRestOfBlock; // true if we know that we will throw
3503 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3505 // There are two modes for ordering of the trees.
3506 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3507 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3508 // by traversing the tree according to the order of the operands.
3509 // - In FGOrderLinear, the dominant ordering is the linear order.
3516 FlowGraphOrder fgOrder;
3518 // The following are boolean flags that keep track of the state of internal data structures
3520 bool fgStmtListThreaded;
3521 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3522 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3523 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3524 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3525 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3526 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3527 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3528 // This is derived from the profile data
3529 // or is BB_UNITY_WEIGHT when we don't have profile data
3531 #if FEATURE_EH_FUNCLETS
3532 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3533 #endif // FEATURE_EH_FUNCLETS
3535 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3536 // since fgMorphTree can be called from several places
3537 bool fgExpandInline; // indicates that we are creating tree for the inliner
3539 bool impBoxTempInUse; // the temp below is valid and available
3540 unsigned impBoxTemp; // a temporary that is used for boxing
3543 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3544 // and we are trying to compile again in a "safer", minopts mode?
3548 unsigned impInlinedCodeSize;
3551 //-------------------------------------------------------------------------
3557 void fgTransformFatCalli();
3561 void fgRemoveEmptyTry();
3563 void fgRemoveEmptyFinally();
3565 void fgCloneFinally();
3567 void fgCleanupContinuation(BasicBlock* continuation);
3569 void fgUpdateFinallyTargetFlags();
3571 GenTreePtr fgGetCritSectOfStaticMethod();
3573 #if !defined(_TARGET_X86_)
3575 void fgAddSyncMethodEnterExit();
3577 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3579 void fgConvertSyncReturnToLeave(BasicBlock* block);
3581 #endif // !_TARGET_X86_
3583 void fgAddReversePInvokeEnterExit();
3585 bool fgMoreThanOneReturnBlock();
3587 // The number of separate return points in the method.
3588 unsigned fgReturnCount;
3590 void fgAddInternal();
3592 bool fgFoldConditional(BasicBlock* block);
3594 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3595 void fgMorphBlocks();
3597 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3599 void fgCheckArgCnt();
3600 void fgSetOptions();
3603 static fgWalkPreFn fgAssertNoQmark;
3604 void fgPreExpandQmarkChecks(GenTreePtr expr);
3605 void fgPostExpandQmarkChecks();
3606 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3609 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3611 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3612 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3613 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3614 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3615 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3617 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3618 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3619 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3620 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3622 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3623 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3624 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3625 void fgExpandQmarkNodes();
3629 // Do "simple lowering." This functionality is (conceptually) part of "general"
3630 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3631 void fgSimpleLowering();
3633 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3635 GenTreePtr fgInitThisClass();
3637 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3639 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3641 void fgLocalVarLiveness();
3643 void fgLocalVarLivenessInit();
3645 #ifdef LEGACY_BACKEND
3646 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3648 void fgPerNodeLocalVarLiveness(GenTree* node);
3650 void fgPerBlockLocalVarLiveness();
3652 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3654 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3656 // This is used in the liveness computation, as a temporary. When we use the
3657 // arbitrary-length VarSet representation, it is better not to allocate a new one
3659 VARSET_TP fgMarkIntfUnionVS;
3661 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3663 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3665 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3667 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3669 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3671 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3673 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3674 GenTreePtr startNode,
3676 VARSET_VALARG_TP volatileVars,
3677 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3679 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3681 bool fgRemoveDeadStore(GenTree** pTree,
3685 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3687 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3689 // For updating liveset during traversal AFTER fgComputeLife has completed
3690 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3691 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3693 // Returns the set of live variables after endTree,
3694 // assuming that liveSet is the set of live variables BEFORE tree.
3695 // Requires that fgComputeLife has completed, and that tree is in the same
3696 // statement as endTree, and that it comes before endTree in execution order
3698 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3700 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3701 while (tree != nullptr && tree != endTree->gtNext)
3703 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3704 tree = tree->gtNext;
3706 assert(tree == endTree->gtNext);
3710 void fgInterBlockLocalVarLiveness();
3712 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3713 // "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
3714 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3715 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3716 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3717 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3718 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3720 if (m_opAsgnVarDefSsaNums == nullptr)
3722 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3724 return m_opAsgnVarDefSsaNums;
3727 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3728 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3729 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3731 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3733 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3734 // Except: assumes that lcl is a def, and if it is
3735 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3736 // rather than the "use" SSA number recorded in the tree "lcl".
3737 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3739 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3740 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3741 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3742 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3743 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3745 // (byref addrS1 = &s1,
3746 // *(addrS1 * offsetof(f0)) = s2f0,
3748 // *(addrS1 * offsetof(fn)) = s2fn)
3750 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3751 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3752 // give it SSA names and value numbers?
3754 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3755 // end with an instance of the structure below, whose fields are described in the declaration.
3756 struct IndirectAssignmentAnnotation
3758 unsigned m_lclNum; // The local num that is being indirectly assigned.
3759 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3760 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3761 // be the singleton field sequence "g". The individual assignments would
3762 // further append the fields of "s.g" to that.
3763 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3764 // structure has a single field).
3765 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3766 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3769 IndirectAssignmentAnnotation(unsigned lclNum,
3770 FieldSeqNode* fldSeq,
3772 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3773 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3774 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3778 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3779 NodeToIndirAssignMap;
3780 NodeToIndirAssignMap* m_indirAssignMap;
3781 NodeToIndirAssignMap* GetIndirAssignMap()
3783 if (m_indirAssignMap == nullptr)
3785 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3786 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3787 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3789 return m_indirAssignMap;
3792 // Performs SSA conversion.
3795 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3796 void fgResetForSsa();
3798 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3800 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3801 inline bool fgExcludeFromSsa(unsigned lclNum);
3803 // The value numbers for this compilation.
3804 ValueNumStore* vnStore;
3807 ValueNumStore* GetValueNumStore()
3812 // Do value numbering (assign a value number to each
3814 void fgValueNumber();
3816 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3817 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3818 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3819 // match the element type of the array or fldSeq. When this type doesn't match
3820 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3822 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3825 FieldSeqNode* fldSeq,
3829 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3830 // has been parsed to yield the other input arguments. If evaluation of the address
3831 // can raise exceptions, those should be captured in the exception set "excVN."
3832 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3833 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3834 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3835 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3836 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3838 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3839 CORINFO_CLASS_HANDLE elemTypeEq,
3843 FieldSeqNode* fldSeq);
3845 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3846 // by evaluating the array index expression "tree". Returns the value number resulting from
3847 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3848 // "GT_IND" that does the dereference, and it is given the returned value number.
3849 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3851 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3852 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3854 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3856 // Utility functions for fgValueNumber.
3858 // Perform value-numbering for the trees in "blk".
3859 void fgValueNumberBlock(BasicBlock* blk);
3861 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3862 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3863 // assumed for the memoryKind at the start "entryBlk".
3864 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3866 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3867 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3868 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3870 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3872 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3874 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3875 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3876 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3878 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3879 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3881 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3882 // value in that SSA #.
3883 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3885 // The input 'tree' is a leaf node that is a constant
3886 // Assign the proper value number to the tree
3887 void fgValueNumberTreeConst(GenTreePtr tree);
3889 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3890 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3892 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3894 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3896 // Does value-numbering for a block assignment.
3897 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3899 // Does value-numbering for a cast tree.
3900 void fgValueNumberCastTree(GenTreePtr tree);
3902 // Does value-numbering for an intrinsic tree.
3903 void fgValueNumberIntrinsic(GenTreePtr tree);
3905 // Does value-numbering for a call. We interpret some helper calls.
3906 void fgValueNumberCall(GenTreeCall* call);
3908 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3909 void fgUpdateArgListVNs(GenTreeArgList* args);
3911 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3912 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3914 // Requires "helpCall" to be a helper call. Assigns it a value number;
3915 // we understand the semantics of some of the calls. Returns "true" if
3916 // the call may modify the heap (we assume arbitrary memory side effects if so).
3917 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3919 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3920 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3922 // These are the current value number for the memory implicit variables while
3923 // doing value numbering. These are the value numbers under the "liberal" interpretation
3924 // of memory values; the "conservative" interpretation needs no VN, since every access of
3925 // memory yields an unknown value.
3926 ValueNum fgCurMemoryVN[MemoryKindCount];
3928 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3929 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3930 // is 1, and the rest is an encoding of "elemTyp".
3931 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3933 if (elemStructType != nullptr)
3935 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3936 varTypeIsIntegral(elemTyp));
3937 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3938 return elemStructType;
3942 elemTyp = varTypeUnsignedToSigned(elemTyp);
3943 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3946 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3947 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3948 // the struct type of the element).
3949 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3951 size_t clsHndVal = size_t(clsHnd);
3952 if (clsHndVal & 0x1)
3954 return var_types(clsHndVal >> 1);
3962 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3963 var_types getJitGCType(BYTE gcType);
3965 enum structPassingKind
3967 SPK_Unknown, // Invalid value, never returned
3968 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3969 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3970 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3971 // parameters registers are used, then the stack will be used)
3972 // for X86 passed on the stack, for ARM32 passed in registers
3973 // or the stack or split between registers and the stack.
3974 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3976 }; // The struct is passed/returned by reference to a copy/buffer.
3978 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3979 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3980 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3981 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3983 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3985 // Get the type that is used to pass values of the given struct type.
3986 // If you have already retrieved the struct size then pass it as the optional third argument
3988 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3989 structPassingKind* wbPassStruct,
3990 unsigned structSize = 0);
3992 // Get the type that is used to return values of the given struct type.
3993 // If you have already retrieved the struct size then pass it as the optional third argument
3995 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3996 structPassingKind* wbPassStruct = nullptr,
3997 unsigned structSize = 0);
4000 // Print a representation of "vnp" or "vn" on standard output.
4001 // If "level" is non-zero, we also print out a partial expansion of the value.
4002 void vnpPrint(ValueNumPair vnp, unsigned level);
4003 void vnPrint(ValueNum vn, unsigned level);
4006 // Dominator computation member functions
4007 // Not exposed outside Compiler
4009 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4011 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4013 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4014 // flow graph. We first assume the fields bbIDom on each
4015 // basic block are invalid. This computation is needed later
4016 // by fgBuildDomTree to build the dominance tree structure.
4017 // Based on: A Simple, Fast Dominance Algorithm
4018 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4020 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4021 // Note: this is relatively slow compared to calling fgDominate(),
4022 // especially if dealing with a single block versus block check.
4024 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4026 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4028 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4030 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4032 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4034 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4035 // processed in topological sort, this function takes care of that.
4037 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4039 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4040 // Returns this as a set.
4042 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4043 // root nodes. Returns this as a set.
4046 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4049 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4050 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4053 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4054 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4055 // && postOrder(A) >= postOrder(B) making the computation O(1).
4056 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4058 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4060 void fgUpdateChangedFlowGraph();
4063 // Compute the predecessors of the blocks in the control flow graph.
4064 void fgComputePreds();
4066 // Remove all predecessor information.
4067 void fgRemovePreds();
4069 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4070 // before the full predecessors lists are computed.
4071 void fgComputeCheapPreds();
4074 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4076 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4086 // Initialize the per-block variable sets (used for liveness analysis).
4087 void fgInitBlockVarSets();
4089 // true if we've gone through and created GC Poll calls.
4090 bool fgGCPollsCreated;
4091 void fgMarkGCPollBlocks();
4092 void fgCreateGCPolls();
4093 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4095 // Requires that "block" is a block that returns from
4096 // a finally. Returns the number of successors (jump targets of
4097 // of blocks in the covered "try" that did a "LEAVE".)
4098 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4100 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4101 // a finally. Returns its "i"th successor (jump targets of
4102 // of blocks in the covered "try" that did a "LEAVE".)
4103 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4104 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4107 // Factor out common portions of the impls of the methods above.
4108 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4111 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4112 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4113 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4114 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4115 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4116 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4117 // we leave the entry associated with the block, but it will no longer be accessed.)
4118 struct SwitchUniqueSuccSet
4120 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4121 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4124 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4125 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4126 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4127 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4130 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4131 BlockToSwitchDescMap;
4134 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4135 // iteration over only the distinct successors.
4136 BlockToSwitchDescMap* m_switchDescMap;
4139 BlockToSwitchDescMap* GetSwitchDescMap()
4141 if (m_switchDescMap == nullptr)
4143 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4145 return m_switchDescMap;
4148 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4149 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4150 // we don't accidentally look up and return the wrong switch data.
4151 void InvalidateUniqueSwitchSuccMap()
4153 m_switchDescMap = nullptr;
4156 // Requires "switchBlock" to be a block that ends in a switch. Returns
4157 // the corresponding SwitchUniqueSuccSet.
4158 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4160 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4161 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4162 // remove it from "this", and ensure that "to" is a member.
4163 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4165 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4166 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4168 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4170 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4172 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4174 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4176 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4178 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4180 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4182 void fgRemoveBlockAsPred(BasicBlock* block);
4184 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4186 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4188 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4190 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4192 flowList* fgAddRefPred(BasicBlock* block,
4193 BasicBlock* blockPred,
4194 flowList* oldEdge = nullptr,
4195 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4198 void fgFindBasicBlocks();
4200 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4202 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4204 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4205 bool putInTryRegion,
4206 BasicBlock* startBlk,
4208 BasicBlock* nearBlk,
4209 BasicBlock* jumpBlk,
4212 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4214 void fgRemoveEmptyBlocks();
4216 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4218 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4220 void fgCreateLoopPreHeader(unsigned lnum);
4222 void fgUnreachableBlock(BasicBlock* block);
4224 void fgRemoveConditionalJump(BasicBlock* block);
4226 BasicBlock* fgLastBBInMainFunction();
4228 BasicBlock* fgEndBBAfterMainFunction();
4230 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4232 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4234 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4236 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4238 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4240 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4242 bool fgRenumberBlocks();
4244 bool fgExpandRarelyRunBlocks();
4246 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4248 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4250 enum FG_RELOCATE_TYPE
4252 FG_RELOCATE_TRY, // relocate the 'try' region
4253 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4255 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4257 #if FEATURE_EH_FUNCLETS
4258 #if defined(_TARGET_ARM_)
4259 void fgClearFinallyTargetBit(BasicBlock* block);
4260 #endif // defined(_TARGET_ARM_)
4261 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4262 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4263 void fgInsertFuncletPrologBlock(BasicBlock* block);
4264 void fgCreateFuncletPrologBlocks();
4265 void fgCreateFunclets();
4266 #else // !FEATURE_EH_FUNCLETS
4267 bool fgRelocateEHRegions();
4268 #endif // !FEATURE_EH_FUNCLETS
4270 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4272 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4274 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4276 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4278 bool fgOptimizeEmptyBlock(BasicBlock* block);
4280 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4282 bool fgOptimizeBranch(BasicBlock* bJump);
4284 bool fgOptimizeSwitchBranches(BasicBlock* block);
4286 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4288 bool fgOptimizeSwitchJumps();
4290 void fgPrintEdgeWeights();
4292 void fgComputeEdgeWeights();
4294 void fgReorderBlocks();
4296 void fgDetermineFirstColdBlock();
4298 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4300 bool fgUpdateFlowGraph(bool doTailDup = false);
4302 void fgFindOperOrder();
4304 // method that returns if you should split here
4305 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4307 void fgSetBlockOrder();
4309 void fgRemoveReturnBlock(BasicBlock* block);
4311 /* Helper code that has been factored out */
4312 inline void fgConvertBBToThrowBB(BasicBlock* block);
4314 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4315 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4316 GenTreePtr fgMakeTmpArgNode(
4317 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4319 // The following check for loops that don't execute calls
4320 bool fgLoopCallMarked;
4322 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4323 void fgLoopCallMark();
4325 void fgMarkLoopHead(BasicBlock* block);
4327 unsigned fgGetCodeEstimate(BasicBlock* block);
4330 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4331 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4332 bool fgDumpFlowGraph(Phases phase);
4334 #endif // DUMP_FLOWGRAPHS
4339 void fgDispBBLiveness(BasicBlock* block);
4340 void fgDispBBLiveness();
4341 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4342 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4343 void fgDispBasicBlocks(bool dumpTrees = false);
4344 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4345 void fgDumpBlock(BasicBlock* block);
4346 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4348 static fgWalkPreFn fgStress64RsltMulCB;
4349 void fgStress64RsltMul();
4350 void fgDebugCheckUpdate();
4351 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4352 void fgDebugCheckBlockLinks();
4353 void fgDebugCheckLinks(bool morphTrees = false);
4354 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4355 void fgDebugCheckFlags(GenTreePtr tree);
4356 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4357 void fgDebugCheckTryFinallyExits();
4360 #ifdef LEGACY_BACKEND
4361 static void fgOrderBlockOps(GenTreePtr tree,
4365 GenTreePtr* opsPtr, // OUT
4366 regMaskTP* regsPtr); // OUT
4367 #endif // LEGACY_BACKEND
4369 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4370 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4372 inline bool fgIsInlining()
4374 return fgExpandInline;
4377 void fgTraverseRPO();
4379 //--------------------- Walking the trees in the IR -----------------------
4384 fgWalkPreFn* wtprVisitorFn;
4385 fgWalkPostFn* wtpoVisitorFn;
4386 void* pCallbackData; // user-provided data
4387 bool wtprLclsOnly; // whether to only visit lclvar nodes
4388 GenTreePtr parent; // parent of current node, provided to callback
4389 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4391 bool printModified; // callback can use this
4395 template <bool computeStack>
4396 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4398 // general purpose tree-walker that is capable of doing pre- and post- order
4399 // callbacks at the same time
4400 template <bool doPreOrder, bool doPostOrder>
4401 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4403 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4404 fgWalkPreFn* visitor,
4405 void* pCallBackData = nullptr,
4406 bool lclVarsOnly = false,
4407 bool computeStack = false);
4409 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4410 fgWalkPreFn* preVisitor,
4411 fgWalkPostFn* postVisitor,
4412 void* pCallBackData = nullptr);
4414 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4418 template <bool computeStack>
4419 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4421 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4422 fgWalkPostFn* visitor,
4423 void* pCallBackData = nullptr,
4424 bool computeStack = false);
4426 // An fgWalkPreFn that looks for expressions that have inline throws in
4427 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4428 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4429 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4430 // properly propagated to parent trees). It returns WALK_CONTINUE
4432 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4433 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4434 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4436 /**************************************************************************
4438 *************************************************************************/
4441 friend class SsaBuilder;
4442 friend struct ValueNumberState;
4444 //--------------------- Detect the basic blocks ---------------------------
4446 BasicBlock** fgBBs; // Table of pointers to the BBs
4448 void fgInitBBLookup();
4449 BasicBlock* fgLookupBB(unsigned addr);
4451 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4453 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4455 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4457 void fgLinkBasicBlocks();
4459 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4461 void fgCheckBasicBlockControlFlow();
4463 void fgControlFlowPermitted(BasicBlock* blkSrc,
4464 BasicBlock* blkDest,
4465 BOOL IsLeave = false /* is the src a leave block */);
4467 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4469 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4471 void fgAdjustForAddressExposedOrWrittenThis();
4473 bool fgProfileData_ILSizeMismatch;
4474 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4475 ULONG fgProfileBufferCount;
4476 ULONG fgNumProfileRuns;
4478 unsigned fgStressBBProf()
4481 unsigned result = JitConfig.JitStressBBProf();
4484 if (compStressCompile(STRESS_BB_PROFILE, 15))
4495 bool fgHaveProfileData();
4496 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4498 bool fgIsUsingProfileWeights()
4500 return (fgHaveProfileData() || fgStressBBProf());
4502 void fgInstrumentMethod();
4504 //-------- Insert a statement at the start or end of a basic block --------
4508 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4512 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4514 public: // Used by linear scan register allocation
4515 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4518 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4519 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4521 public: // Used by linear scan register allocation
4522 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4525 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4527 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4529 // Create a new temporary variable to hold the result of *ppTree,
4530 // and transform the graph accordingly.
4531 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4532 GenTree* fgMakeMultiUse(GenTree** ppTree);
4535 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4536 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4537 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4539 //-------- Determine the order in which the trees will be evaluated -------
4541 unsigned fgTreeSeqNum;
4542 GenTree* fgTreeSeqLst;
4543 GenTree* fgTreeSeqBeg;
4545 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4546 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4547 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4548 void fgSetStmtSeq(GenTree* tree);
4549 void fgSetBlockOrder(BasicBlock* block);
4551 //------------------------- Morphing --------------------------------------
4553 unsigned fgPtrArgCntCur;
4554 unsigned fgPtrArgCntMax;
4555 hashBv* fgOutgoingArgTemps;
4556 hashBv* fgCurrentlyInUseArgTemps;
4558 bool compCanEncodePtrArgCntMax();
4560 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4563 void fgMoveOpsLeft(GenTreePtr tree);
4566 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4568 bool fgIsThrow(GenTreePtr tree);
4570 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4571 bool fgIsBlockCold(BasicBlock* block);
4573 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4575 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4577 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4579 bool fgMorphRelopToQmark(GenTreePtr tree);
4581 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4582 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4583 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4584 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4585 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4586 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4587 // small; hence the other fields of MorphAddrContext.
4588 enum MorphAddrContextKind
4593 struct MorphAddrContext
4595 MorphAddrContextKind m_kind;
4596 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4597 // top-level indirection and here have been constants.
4598 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4599 // In that case, is the sum of those constant offsets.
4601 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4606 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4607 static MorphAddrContext s_CopyBlockMAC;
4610 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4611 var_types* baseTypeOut,
4613 unsigned* simdSizeOut,
4614 bool ignoreUsedInSIMDIntrinsic = false);
4615 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4616 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4617 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4618 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4620 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4621 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4622 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4624 #endif // FEATURE_SIMD
4625 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4626 GenTreePtr fgMorphCast(GenTreePtr tree);
4627 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4628 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4630 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4633 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4634 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4636 void fgFixupStructReturn(GenTreePtr call);
4637 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4638 bool fgAddrCouldBeNull(GenTreePtr addr);
4639 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4640 bool fgCanFastTailCall(GenTreeCall* call);
4641 void fgMorphTailCall(GenTreeCall* call);
4642 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4643 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4644 fgArgTabEntryPtr argTabEntry,
4646 IL_OFFSETX callILOffset,
4647 GenTreePtr tmpAssignmentInsertionPoint,
4648 GenTreePtr paramAssignmentInsertionPoint);
4649 static int fgEstimateCallStackSize(GenTreeCall* call);
4650 GenTreePtr fgMorphCall(GenTreeCall* call);
4651 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4652 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4654 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4655 static fgWalkPreFn fgFindNonInlineCandidate;
4657 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4658 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4659 void fgAssignSetVarDef(GenTreePtr tree);
4660 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4661 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4662 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4663 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4664 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4665 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4666 void fgMorphUnsafeBlk(GenTreeObj* obj);
4667 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4668 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4669 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4670 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4671 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4672 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4673 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4675 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4676 GenTreePtr fgMorphConst(GenTreePtr tree);
4679 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4682 #if LOCAL_ASSERTION_PROP
4683 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4684 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4686 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4688 GenTreeStmt* fgMorphStmt;
4690 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4691 // used when morphing big offset.
4693 //----------------------- Liveness analysis -------------------------------
4695 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4696 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4698 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4699 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4700 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4702 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4704 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4706 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4707 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4709 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4710 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4712 void fgExtendDbgScopes();
4713 void fgExtendDbgLifetimes();
4716 void fgDispDebugScopes();
4719 //-------------------------------------------------------------------------
4721 // The following keeps track of any code we've added for things like array
4722 // range checking or explicit calls to enable GC, and so on.
4727 AddCodeDsc* acdNext;
4728 BasicBlock* acdDstBlk; // block to which we jump
4730 SpecialCodeKind acdKind; // what kind of a special block is this?
4731 unsigned short acdStkLvl;
4735 static unsigned acdHelper(SpecialCodeKind codeKind);
4737 AddCodeDsc* fgAddCodeList;
4739 bool fgRngChkThrowAdded;
4740 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4742 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4744 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4747 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4750 bool fgIsCodeAdded();
4752 bool fgIsThrowHlpBlk(BasicBlock* block);
4753 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4755 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4757 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4758 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4759 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4760 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4761 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4763 #if FEATURE_MULTIREG_RET
4764 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4765 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4766 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4767 #endif // FEATURE_MULTIREG_RET
4769 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4772 static fgWalkPreFn fgDebugCheckInlineCandidates;
4774 void CheckNoFatPointerCandidatesLeft();
4775 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4778 void fgPromoteStructs();
4779 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4780 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4781 void fgMarkImplicitByRefArgs();
4782 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4783 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4784 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4785 void fgMarkAddressExposedLocals();
4786 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4788 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4790 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4792 // The given local variable, required to be a struct variable, is being assigned via
4793 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4794 // the variable is not enregistered, and is therefore not promoted independently.
4795 void fgLclFldAssign(unsigned lclNum);
4797 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4798 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4799 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4800 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4803 bool fgPrintInlinedMethods;
4806 bool fgIsBigOffset(size_t offset);
4808 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4809 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4810 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4811 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4812 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4815 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4816 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4820 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4821 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4828 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4831 void optRemoveRangeCheck(
4832 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4833 bool optIsRangeCheckRemovable(GenTreePtr tree);
4836 static fgWalkPreFn optValidRangeCheckIndex;
4837 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4840 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4842 /**************************************************************************
4844 *************************************************************************/
4847 // Do hoisting for all loops.
4848 void optHoistLoopCode();
4850 // To represent sets of VN's that have already been hoisted in outer loops.
4851 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4852 typedef VNToBoolMap VNSet;
4854 struct LoopHoistContext
4857 // The set of variables hoisted in the current loop (or nullptr if there are none).
4858 VNSet* m_pHoistedInCurLoop;
4861 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4862 VNSet m_hoistedInParentLoops;
4863 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4864 // Previous decisions on loop-invariance of value numbers in the current loop.
4865 VNToBoolMap m_curLoopVnInvariantCache;
4867 VNSet* GetHoistedInCurLoop(Compiler* comp)
4869 if (m_pHoistedInCurLoop == nullptr)
4871 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4873 return m_pHoistedInCurLoop;
4876 VNSet* ExtractHoistedInCurLoop()
4878 VNSet* res = m_pHoistedInCurLoop;
4879 m_pHoistedInCurLoop = nullptr;
4883 LoopHoistContext(Compiler* comp)
4884 : m_pHoistedInCurLoop(nullptr)
4885 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4886 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4891 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4892 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4893 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4894 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4896 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4897 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4898 // "m_hoistedInParentLoops".
4900 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4902 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4903 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4904 // expressions to "hoistInLoop".
4905 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4907 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4908 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4910 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4911 // that are invariant in loop "lnum" (an index into the optLoopTable)
4912 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4913 // expressions to "hoistInLoop".
4914 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4915 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4916 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4917 bool optHoistLoopExprsForTree(GenTreePtr tree,
4919 LoopHoistContext* hoistCtxt,
4920 bool* firstBlockAndBeforeSideEffect,
4923 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4924 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4926 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4927 // Constants and init values are always loop invariant.
4928 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4929 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4931 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4932 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4933 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4934 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4935 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4937 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4938 // in the loop table.
4939 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4941 // Records the set of "side effects" of all loops: fields (object instance and static)
4942 // written to, and SZ-array element type equivalence classes updated.
4943 void optComputeLoopSideEffects();
4946 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4947 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4948 // static) written to, and SZ-array element type equivalence classes updated.
4949 void optComputeLoopNestSideEffects(unsigned lnum);
4951 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4952 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4954 // Hoist the expression "expr" out of loop "lnum".
4955 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4958 void optOptimizeBools();
4961 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4963 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4966 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4968 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4969 // the loop into a "do-while" loop
4970 // Also finds all natural loops and records them in the loop table
4972 // Optionally clone loops in the loop table.
4973 void optCloneLoops();
4975 // Clone loop "loopInd" in the loop table.
4976 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4978 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4979 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4980 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4982 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4984 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4987 // This enumeration describes what is killed by a call.
4991 CALLINT_NONE, // no interference (most helpers)
4992 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4993 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4994 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4995 CALLINT_ALL, // kills everything (normal method call)
4999 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5000 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5001 // in bbNext order; we use comparisons on the bbNum to decide order.)
5002 // The blocks that define the body are
5003 // first <= top <= entry <= bottom .
5004 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5005 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5006 // Compiler::optFindNaturalLoops().
5009 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5010 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5011 // loop, but not the outer loop.)
5012 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5014 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5015 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5016 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5018 callInterf lpAsgCall; // "callInterf" for calls in the loop
5019 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5020 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5022 unsigned short lpFlags; // Mask of the LPFLG_* constants
5024 unsigned char lpExitCnt; // number of exits from the loop
5026 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5027 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5028 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5029 // (Actually, an "immediately" nested loop --
5030 // no other child of this loop is a parent of lpChild.)
5031 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5032 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5033 // by following "lpChild" then "lpSibling" links.
5035 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5036 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5038 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5039 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5040 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5042 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5043 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5045 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5046 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5047 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5048 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5050 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5051 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5052 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5054 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5055 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5056 // type are assigned to.
5058 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5059 // memory side effects. If this is set, the fields below
5060 // may not be accurate (since they become irrelevant.)
5061 bool lpContainsCall; // True if executing the loop body *may* execute a call
5063 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5064 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5066 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5068 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5069 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5071 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5073 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5074 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5076 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5077 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5079 JitSimplerHashBehavior>
5081 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5082 // instance fields modified
5085 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5086 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5088 JitSimplerHashBehavior>
5090 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5091 // arrays of that type are modified
5094 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5095 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5097 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5098 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5099 // (shifted left, with a low-order bit set to distinguish.)
5100 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5101 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5103 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5105 GenTreePtr lpIterTree; // The "i <op>= const" tree
5106 unsigned lpIterVar(); // iterator variable #
5107 int lpIterConst(); // the constant with which the iterator is incremented
5108 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5109 void VERIFY_lpIterTree();
5111 var_types lpIterOperType(); // For overflow instructions
5114 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5115 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5119 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5121 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5122 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5123 void VERIFY_lpTestTree();
5125 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5126 GenTreePtr lpIterator(); // the iterator node in the loop test
5127 GenTreePtr lpLimit(); // the limit node in the loop test
5129 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5130 // LPFLG_CONST_LIMIT
5131 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5133 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5134 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5135 // LPFLG_ARRLEN_LIMIT
5137 // Returns "true" iff "*this" contains the blk.
5138 bool lpContains(BasicBlock* blk)
5140 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5142 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5143 // to be equal, but requiring bottoms to be different.)
5144 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5146 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5149 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5150 // bottoms to be different.)
5151 bool lpContains(const LoopDsc& lp2)
5153 return lpContains(lp2.lpFirst, lp2.lpBottom);
5156 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5157 // (allowing firsts to be equal, but requiring bottoms to be different.)
5158 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5160 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5163 // Returns "true" iff "*this" is (properly) contained by "lp2"
5164 // (allowing firsts to be equal, but requiring bottoms to be different.)
5165 bool lpContainedBy(const LoopDsc& lp2)
5167 return lpContains(lp2.lpFirst, lp2.lpBottom);
5170 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5171 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5173 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5175 // Returns "true" iff "*this" is disjoint from "lp2".
5176 bool lpDisjoint(const LoopDsc& lp2)
5178 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5180 // Returns "true" iff the loop is well-formed (see code for defn).
5183 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5184 lpEntry->bbNum <= lpBottom->bbNum &&
5185 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5190 bool fgMightHaveLoop(); // returns true if there are any backedges
5191 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5194 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5195 unsigned char optLoopCount; // number of tracked loops
5198 unsigned optCallCount; // number of calls made in the method
5199 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5200 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5201 unsigned optLoopsCloned; // number of loops cloned in the current method.
5204 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5205 void optPrintLoopInfo(unsigned loopNum,
5207 BasicBlock* lpFirst,
5209 BasicBlock* lpEntry,
5210 BasicBlock* lpBottom,
5211 unsigned char lpExitCnt,
5213 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5214 void optPrintLoopInfo(unsigned lnum);
5215 void optPrintLoopRecording(unsigned lnum);
5217 void optCheckPreds();
5220 void optSetBlockWeights();
5222 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5224 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5226 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5228 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5229 unsigned optIsLoopIncrTree(GenTreePtr incr);
5230 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5231 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5232 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5233 bool optExtractInitTestIncr(BasicBlock* head,
5238 GenTreePtr* ppIncr);
5240 void optRecordLoop(BasicBlock* head,
5246 unsigned char exitCnt);
5248 void optFindNaturalLoops();
5250 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5251 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5252 bool optCanonicalizeLoopNest(unsigned char loopInd);
5254 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5255 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5256 bool optCanonicalizeLoop(unsigned char loopInd);
5258 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5259 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5260 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5261 bool optLoopContains(unsigned l1, unsigned l2);
5263 // Requires "loopInd" to be a valid index into the loop table.
5264 // Updates the loop table by changing loop "loopInd", whose head is required
5265 // to be "from", to be "to". Also performs this transformation for any
5266 // loop nested in "loopInd" that shares the same head as "loopInd".
5267 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5269 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5270 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5271 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5273 // Marks the containsCall information to "lnum" and any parent loops.
5274 void AddContainsCallAllContainingLoops(unsigned lnum);
5275 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5276 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5277 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5278 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5279 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5280 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5282 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5283 // of "from".) Copies the jump destination from "from" to "to".
5284 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5286 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5287 unsigned optLoopDepth(unsigned lnum)
5289 unsigned par = optLoopTable[lnum].lpParent;
5290 if (par == BasicBlock::NOT_IN_LOOP)
5296 return 1 + optLoopDepth(par);
5300 void fgOptWhileLoop(BasicBlock* block);
5302 bool optComputeLoopRep(int constInit,
5305 genTreeOps iterOper,
5307 genTreeOps testOper,
5310 unsigned* iterCount);
5311 #if FEATURE_STACK_FP_X87
5314 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5315 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5316 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5317 #endif // FEATURE_STACK_FP_X87
5320 static fgWalkPreFn optIsVarAssgCB;
5323 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5325 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5327 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5329 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5331 /**************************************************************************
5332 * Optimization conditions
5333 *************************************************************************/
5335 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5336 bool optPentium4(void);
5337 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5338 bool optAvoidIntMult(void);
5343 // The following is the upper limit on how many expressions we'll keep track
5344 // of for the CSE analysis.
5346 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5348 static const int MIN_CSE_COST = 2;
5350 // Keeps tracked cse indices
5351 BitVecTraits* cseTraits;
5355 /* Generic list of nodes - used by the CSE logic */
5363 typedef struct treeLst* treeLstPtr;
5367 treeStmtLst* tslNext;
5368 GenTreePtr tslTree; // tree node
5369 GenTreePtr tslStmt; // statement containing the tree
5370 BasicBlock* tslBlock; // block containing the statement
5373 typedef struct treeStmtLst* treeStmtLstPtr;
5375 // The following logic keeps track of expressions via a simple hash table.
5379 CSEdsc* csdNextInBucket; // used by the hash table
5381 unsigned csdHashValue; // the orginal hashkey
5383 unsigned csdIndex; // 1..optCSECandidateCount
5384 char csdLiveAcrossCall; // 0 or 1
5386 unsigned short csdDefCount; // definition count
5387 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5389 unsigned csdDefWtCnt; // weighted def count
5390 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5392 GenTreePtr csdTree; // treenode containing the 1st occurance
5393 GenTreePtr csdStmt; // stmt containing the 1st occurance
5394 BasicBlock* csdBlock; // block containing the 1st occurance
5396 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5397 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5399 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5400 // number, this will reflect it; otherwise, NoVN.
5403 static const size_t s_optCSEhashSize;
5404 CSEdsc** optCSEhash;
5409 CSEdsc* optCSEfindDsc(unsigned index);
5410 void optUnmarkCSE(GenTreePtr tree);
5412 // user defined callback data for the tree walk function optCSE_MaskHelper()
5413 struct optCSE_MaskData
5415 EXPSET_TP CSE_defMask;
5416 EXPSET_TP CSE_useMask;
5419 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5420 static fgWalkPreFn optCSE_MaskHelper;
5422 // This function walks all the node for an given tree
5423 // and return the mask of CSE definitions and uses for the tree
5425 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5427 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5428 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5429 bool optCSE_canSwap(GenTree* tree);
5431 static fgWalkPostFn optPropagateNonCSE;
5432 static fgWalkPreFn optHasNonCSEChild;
5434 static fgWalkPreFn optUnmarkCSEs;
5436 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5437 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5439 void optCleanupCSEs();
5442 void optEnsureClearCSEInfo();
5445 #endif // FEATURE_ANYCSE
5447 #if FEATURE_VALNUM_CSE
5448 /**************************************************************************
5449 * Value Number based CSEs
5450 *************************************************************************/
5453 void optOptimizeValnumCSEs();
5456 void optValnumCSE_Init();
5457 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5458 unsigned optValnumCSE_Locate();
5459 void optValnumCSE_InitDataFlow();
5460 void optValnumCSE_DataFlow();
5461 void optValnumCSE_Availablity();
5462 void optValnumCSE_Heuristic();
5463 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5465 #endif // FEATURE_VALNUM_CSE
5468 bool optDoCSE; // True when we have found a duplicate CSE tree
5469 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5470 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5471 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5472 unsigned optCSEstart; // The first local variable number that is a CSE
5473 unsigned optCSEcount; // The total count of CSE's introduced.
5474 unsigned optCSEweight; // The weight of the current block when we are
5475 // scanning for CSE expressions
5477 bool optIsCSEcandidate(GenTreePtr tree);
5479 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5481 bool lclNumIsTrueCSE(unsigned lclNum) const
5483 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5486 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5488 bool lclNumIsCSE(unsigned lclNum) const
5490 return lvaTable[lclNum].lvIsCSE;
5494 bool optConfigDisableCSE();
5495 bool optConfigDisableCSE2();
5497 void optOptimizeCSEs();
5499 #endif // FEATURE_ANYCSE
5507 unsigned ivaVar; // Variable we are interested in, or -1
5508 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5509 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5510 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5511 callInterf ivaMaskCall; // What kind of calls are there?
5514 static callInterf optCallInterf(GenTreePtr call);
5517 // VN based copy propagation.
5518 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5519 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5520 LclNumToGenTreePtrStack;
5522 // Kill set to track variables with intervening definitions.
5523 VARSET_TP optCopyPropKillSet;
5525 // Copy propagation functions.
5526 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5527 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5528 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5529 bool optIsSsaLocal(GenTreePtr tree);
5530 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5531 void optVnCopyProp();
5533 /**************************************************************************
5534 * Early value propagation
5535 *************************************************************************/
5541 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5545 static unsigned GetHashCode(SSAName ssaNm)
5547 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5550 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5552 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5556 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5557 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5558 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5559 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5560 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5561 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5563 bool doesMethodHaveFatPointer()
5565 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5568 void setMethodHasFatPointer()
5570 optMethodFlags |= OMF_HAS_FATPOINTER;
5573 void clearMethodHasFatPointer()
5575 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5578 void addFatPointerCandidate(GenTreeCall* call)
5580 setMethodHasFatPointer();
5581 call->SetFatPointerCandidate();
5584 unsigned optMethodFlags;
5586 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5587 // No throughput diff was found with backward walk bound between 3-8.
5588 static const int optEarlyPropRecurBound = 5;
5590 enum class optPropKind
5598 bool gtIsVtableRef(GenTreePtr tree);
5599 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5600 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5601 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5602 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5603 bool optEarlyPropRewriteTree(GenTreePtr tree);
5604 bool optDoEarlyPropForBlock(BasicBlock* block);
5605 bool optDoEarlyPropForFunc();
5606 void optEarlyProp();
5607 void optFoldNullCheck(GenTreePtr tree);
5608 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5611 /**************************************************************************
5612 * Value/Assertion propagation
5613 *************************************************************************/
5615 // Data structures for assertion prop
5616 BitVecTraits* apTraits;
5620 enum optAssertionKind
5635 O1K_ARRLEN_OPER_BND,
5636 O1K_ARRLEN_LOOP_BND,
5637 O1K_CONSTANT_LOOP_BND,
5658 optAssertionKind assertionKind;
5661 unsigned lclNum; // assigned to or property of this local var number
5669 struct AssertionDscOp1
5671 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5678 struct AssertionDscOp2
5680 optOp2Kind kind; // a const or copy assignment
5684 ssize_t iconVal; // integer
5685 unsigned iconFlags; // gtFlags
5687 struct Range // integer subrange
5701 bool IsArrLenArithBound()
5703 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5705 bool IsArrLenBound()
5707 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5709 bool IsConstantBound()
5711 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5712 op1.kind == O1K_CONSTANT_LOOP_BND);
5714 bool IsBoundsCheckNoThrow()
5716 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5719 bool IsCopyAssertion()
5721 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5724 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5726 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5727 a1->op2.kind == a2->op2.kind;
5730 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5732 if (kind == OAK_EQUAL)
5734 return kind2 == OAK_NOT_EQUAL;
5736 else if (kind == OAK_NOT_EQUAL)
5738 return kind2 == OAK_EQUAL;
5743 static ssize_t GetLowerBoundForIntegralType(var_types type)
5763 static ssize_t GetUpperBoundForIntegralType(var_types type)
5787 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5789 return (op1.kind == that->op1.kind) &&
5790 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5793 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5795 if (op2.kind != that->op2.kind)
5801 case O2K_IND_CNS_INT:
5803 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5805 case O2K_CONST_LONG:
5806 return (op2.lconVal == that->op2.lconVal);
5808 case O2K_CONST_DOUBLE:
5809 // exact match because of positive and negative zero.
5810 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5812 case O2K_LCLVAR_COPY:
5814 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5815 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5818 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5821 // we will return false
5825 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5831 bool Complementary(AssertionDsc* that, bool vnBased)
5833 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5834 HasSameOp2(that, vnBased);
5837 bool Equals(AssertionDsc* that, bool vnBased)
5839 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5843 typedef unsigned short AssertionIndex;
5846 static fgWalkPreFn optAddCopiesCallback;
5847 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5848 unsigned optAddCopyLclNum;
5849 GenTreePtr optAddCopyAsgnNode;
5851 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5852 bool optAssertionPropagated; // set to true if we modified the trees
5853 bool optAssertionPropagatedCurrentStmt;
5855 GenTreePtr optAssertionPropCurrentTree;
5857 AssertionIndex* optComplementaryAssertionMap;
5858 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5859 // using the value of a local var) for each local var
5860 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5861 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5862 AssertionIndex optMaxAssertionCount;
5865 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5866 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5867 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5868 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5869 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5870 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5872 AssertionIndex GetAssertionCount()
5874 return optAssertionCount;
5876 ASSERT_TP* bbJtrueAssertionOut;
5877 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5878 ValueNumToAssertsMap;
5879 ValueNumToAssertsMap* optValueNumToAsserts;
5881 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5883 // Assertion prop helpers.
5884 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5885 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5886 void optAssertionInit(bool isLocalProp);
5887 void optAssertionTraitsInit(AssertionIndex assertionCount);
5888 #if LOCAL_ASSERTION_PROP
5889 void optAssertionReset(AssertionIndex limit);
5890 void optAssertionRemove(AssertionIndex index);
5893 // Assertion prop data flow functions.
5894 void optAssertionPropMain();
5895 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5896 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5897 ASSERT_TP* optInitAssertionDataflowFlags();
5898 ASSERT_TP* optComputeAssertionGen();
5900 // Assertion Gen functions.
5901 void optAssertionGen(GenTreePtr tree);
5902 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5903 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5904 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5905 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5906 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5907 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5909 // Assertion creation functions.
5910 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5911 AssertionIndex optCreateAssertion(GenTreePtr op1,
5913 optAssertionKind assertionKind,
5914 AssertionDsc* assertion);
5915 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5917 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5918 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5919 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5921 void optPrintVnAssertionMapping();
5923 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5925 // Used for respective assertion propagations.
5926 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5927 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5928 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5929 bool optAssertionIsNonNull(GenTreePtr op,
5930 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5932 // Used for Relop propagation.
5933 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5934 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5935 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5937 // Assertion prop for lcl var functions.
5938 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5939 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5941 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5942 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5943 const GenTreePtr tree,
5944 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5945 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5947 // Assertion propagation functions.
5948 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5949 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5950 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5951 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5952 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5953 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5954 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5955 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5956 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5957 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5958 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5959 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5961 // Implied assertion functions.
5962 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5963 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5964 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5965 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5967 ASSERT_VALRET_TP optNewFullAssertSet();
5968 ASSERT_VALRET_TP optNewEmptyAssertSet();
5971 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5972 void optDebugCheckAssertion(AssertionDsc* assertion);
5973 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5975 void optAddCopies();
5976 #endif // ASSERTION_PROP
5978 /**************************************************************************
5980 *************************************************************************/
5983 struct LoopCloneVisitorInfo
5985 LoopCloneContext* context;
5988 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5989 : context(context), loopNum(loopNum), stmt(nullptr)
5994 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5995 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5996 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5997 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5998 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5999 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
6000 void optObtainLoopCloningOpts(LoopCloneContext* context);
6001 bool optIsLoopClonable(unsigned loopInd);
6003 bool optCanCloneLoops();
6006 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
6008 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6009 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6010 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6011 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6015 void optInsertLoopCloningStress(BasicBlock* head);
6017 #if COUNT_RANGECHECKS
6018 static unsigned optRangeChkRmv;
6019 static unsigned optRangeChkAll;
6028 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6033 RngChkDsc* rcdNextInBucket; // used by the hash table
6035 unsigned short rcdHashValue; // to make matching faster
6036 unsigned short rcdIndex; // 0..optRngChkCount-1
6038 GenTreePtr rcdTree; // the array index tree
6041 unsigned optRngChkCount;
6042 static const size_t optRngChkHashSize;
6044 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6045 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6047 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6050 bool optLoopsMarked;
6053 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6054 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6058 XX Does the register allocation and puts the remaining lclVars on the stack XX
6060 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6061 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6065 #ifndef LEGACY_BACKEND
6070 #else // LEGACY_BACKEND
6075 #endif // LEGACY_BACKEND
6077 #ifdef LEGACY_BACKEND
6079 void raAssignVars(); // register allocation
6080 #endif // LEGACY_BACKEND
6082 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6084 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6086 void raMarkStkVars();
6089 // Some things are used by both LSRA and regpredict allocators.
6091 FrameType rpFrameType;
6092 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6094 #ifdef LEGACY_BACKEND
6095 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6097 #endif // LEGACY_BACKEND
6099 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6101 #if FEATURE_FP_REGALLOC
6102 enum enumConfigRegisterFP
6104 CONFIG_REGISTER_FP_NONE = 0x0,
6105 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6106 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6107 CONFIG_REGISTER_FP_FULL = 0x3,
6109 enumConfigRegisterFP raConfigRegisterFP();
6110 #endif // FEATURE_FP_REGALLOC
6113 regMaskTP raConfigRestrictMaskFP();
6116 #ifndef LEGACY_BACKEND
6117 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6118 #else // LEGACY_BACKEND
6119 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6120 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6121 bool raNewBlocks; // True is we added killing blocks for FPU registers
6122 unsigned rpPasses; // Number of passes made by the register predicter
6123 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6124 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6125 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6126 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6127 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6128 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6129 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6130 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6131 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6132 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6133 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6134 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6136 bool rpRegAllocDone; // Set to true after we have completed register allocation
6138 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6140 void raSetupArgMasks(RegState* r);
6142 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6144 void raDumpVarIntf(); // Dump the variable to variable interference graph
6145 void raDumpRegIntf(); // Dump the variable to register interference graph
6147 void raAdjustVarIntf();
6149 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6151 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6153 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6154 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6156 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6158 static fgWalkPreFn rpMarkRegIntf;
6160 regMaskTP rpPredictAddressMode(
6161 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6163 void rpPredictRefAssign(unsigned lclNum);
6165 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6167 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6169 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6171 void rpPredictRegUse(); // Entry point
6173 unsigned raPredictTreeRegUse(GenTreePtr tree);
6174 unsigned raPredictListRegUse(GenTreePtr list);
6176 void raSetRegVarOrder(var_types regType,
6177 regNumber* customVarOrder,
6178 unsigned* customVarOrderSize,
6180 regMaskTP avoidReg);
6182 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6183 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6184 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6185 void raAddToStkPredict(unsigned val)
6187 unsigned newStkPredict = rpStkPredict + val;
6188 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6189 rpStkPredict = UINT_MAX - 1;
6191 rpStkPredict = newStkPredict;
6195 #if !FEATURE_FP_REGALLOC
6196 void raDispFPlifeInfo();
6200 regMaskTP genReturnRegForTree(GenTreePtr tree);
6201 #endif // LEGACY_BACKEND
6203 /* raIsVarargsStackArg is called by raMaskStkVars and by
6204 lvaSortByRefCount. It identifies the special case
6205 where a varargs function has a parameter passed on the
6206 stack, other than the special varargs handle. Such parameters
6207 require special treatment, because they cannot be tracked
6208 by the GC (their offsets in the stack are not known
6212 bool raIsVarargsStackArg(unsigned lclNum)
6216 LclVarDsc* varDsc = &lvaTable[lclNum];
6218 assert(varDsc->lvIsParam);
6220 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6222 #else // _TARGET_X86_
6226 #endif // _TARGET_X86_
6229 #ifdef LEGACY_BACKEND
6230 // Records the current prediction, if it's better than any previous recorded prediction.
6231 void rpRecordPrediction();
6232 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6233 void rpUseRecordedPredictionIfBetter();
6235 // Data members used in the methods above.
6236 unsigned rpBestRecordedStkPredict;
6237 struct VarRegPrediction
6239 bool m_isEnregistered;
6240 regNumberSmall m_regNum;
6241 regNumberSmall m_otherReg;
6243 VarRegPrediction* rpBestRecordedPrediction;
6244 #endif // LEGACY_BACKEND
6247 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6248 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6252 XX Get to the class and method info from the Execution Engine given XX
6253 XX tokens for the class and method XX
6255 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6256 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6260 /* These are the different addressing modes used to access a local var.
6261 * The JIT has to report the location of the locals back to the EE
6262 * for debugging purposes.
6268 VLT_REG_BYREF, // this type is currently only used for value types on X64
6271 VLT_STK_BYREF, // this type is currently only used for value types on X64
6285 siVarLocType vlType;
6288 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6290 // VLT_REG_BYREF -- the specified register contains the address of the variable
6298 // VLT_STK -- Any 32 bit value which is on the stack
6299 // eg. [ESP+0x20], or [EBP-0x28]
6300 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6301 // eg. mov EAX, [ESP+0x20]; [EAX]
6305 regNumber vlsBaseReg;
6306 NATIVE_OFFSET vlsOffset;
6309 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6318 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6319 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6327 regNumber vlrssBaseReg;
6328 NATIVE_OFFSET vlrssOffset;
6332 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6333 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6339 regNumber vlsrsBaseReg;
6340 NATIVE_OFFSET vlsrsOffset;
6346 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6347 // eg 2 DWords at [ESP+0x10]
6351 regNumber vls2BaseReg;
6352 NATIVE_OFFSET vls2Offset;
6355 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6356 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6363 // VLT_FIXED_VA -- fixed argument of a varargs function.
6364 // The argument location depends on the size of the variable
6365 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6366 // location of the first arg. This argument can then be accessed
6367 // relative to the position of the first arg
6371 unsigned vlfvOffset;
6378 void* rpValue; // pointer to the in-process
6379 // location of the value.
6385 bool vlIsInReg(regNumber reg);
6386 bool vlIsOnStk(regNumber reg, signed offset);
6389 /*************************************************************************/
6394 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6395 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6396 CORINFO_CALLINFO_FLAGS flags,
6397 CORINFO_CALL_INFO* pResult);
6398 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6400 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6401 CORINFO_ACCESS_FLAGS flags,
6402 CORINFO_FIELD_INFO* pResult);
6406 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6408 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6410 bool IsSuperPMIException(unsigned code)
6412 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6414 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6415 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6416 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6417 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6418 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6419 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6420 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6421 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6425 case EXCEPTIONCODE_DebugBreakorAV:
6426 case EXCEPTIONCODE_MC:
6427 case EXCEPTIONCODE_LWM:
6428 case EXCEPTIONCODE_SASM:
6429 case EXCEPTIONCODE_SSYM:
6430 case EXCEPTIONCODE_CALLUTILS:
6431 case EXCEPTIONCODE_TYPEUTILS:
6432 case EXCEPTIONCODE_ASSERT:
6439 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6440 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6442 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6443 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6446 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6447 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6448 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6450 // VOM info, method sigs
6452 void eeGetSig(unsigned sigTok,
6453 CORINFO_MODULE_HANDLE scope,
6454 CORINFO_CONTEXT_HANDLE context,
6455 CORINFO_SIG_INFO* retSig);
6457 void eeGetCallSiteSig(unsigned sigTok,
6458 CORINFO_MODULE_HANDLE scope,
6459 CORINFO_CONTEXT_HANDLE context,
6460 CORINFO_SIG_INFO* retSig);
6462 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6464 // Method entry-points, instrs
6466 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6468 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6470 CORINFO_EE_INFO eeInfo;
6471 bool eeInfoInitialized;
6473 CORINFO_EE_INFO* eeGetEEInfo();
6475 // Gets the offset of a SDArray's first element
6476 unsigned eeGetArrayDataOffset(var_types type);
6477 // Gets the offset of a MDArray's first element
6478 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6480 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6482 // Returns the page size for the target machine as reported by the EE.
6483 inline size_t eeGetPageSize()
6485 #if COR_JIT_EE_VERSION > 460
6486 return eeGetEEInfo()->osPageSize;
6487 #else // COR_JIT_EE_VERSION <= 460
6488 return CORINFO_PAGE_SIZE;
6489 #endif // COR_JIT_EE_VERSION > 460
6492 // Returns the frame size at which we will generate a loop to probe the stack.
6493 inline size_t getVeryLargeFrameSize()
6496 // The looping probe code is 40 bytes, whereas the straight-line probing for
6497 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6498 // or greater, to generate smaller code.
6499 return 2 * eeGetPageSize();
6501 return 3 * eeGetPageSize();
6505 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6507 #if COR_JIT_EE_VERSION > 460
6508 return eeGetEEInfo()->targetAbi == abi;
6510 return CORINFO_DESKTOP_ABI == abi;
6514 inline bool generateCFIUnwindCodes()
6516 #ifdef UNIX_AMD64_ABI
6517 return IsTargetAbi(CORINFO_CORERT_ABI);
6525 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6527 // Debugging support - Line number info
6529 void eeGetStmtOffsets();
6531 unsigned eeBoundariesCount;
6533 struct boundariesDsc
6535 UNATIVE_OFFSET nativeIP;
6537 unsigned sourceReason;
6538 } * eeBoundaries; // Boundaries to report to EE
6539 void eeSetLIcount(unsigned count);
6540 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6544 static void eeDispILOffs(IL_OFFSET offs);
6545 static void eeDispLineInfo(const boundariesDsc* line);
6546 void eeDispLineInfos();
6549 // Debugging support - Local var info
6553 unsigned eeVarsCount;
6555 struct VarResultInfo
6557 UNATIVE_OFFSET startOffset;
6558 UNATIVE_OFFSET endOffset;
6562 void eeSetLVcount(unsigned count);
6563 void eeSetLVinfo(unsigned which,
6564 UNATIVE_OFFSET startOffs,
6565 UNATIVE_OFFSET length,
6570 const siVarLoc& loc);
6574 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6575 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6578 // ICorJitInfo wrappers
6580 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6582 void eeAllocUnwindInfo(BYTE* pHotCode,
6588 CorJitFuncKind funcKind);
6590 void eeSetEHcount(unsigned cEH);
6592 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6594 WORD eeGetRelocTypeHint(void* target);
6596 // ICorStaticInfo wrapper functions
6598 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6600 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6602 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6605 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6606 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6607 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6608 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6610 template <typename ParamType>
6611 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6613 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6616 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6618 // Utility functions
6620 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6623 const wchar_t* eeGetCPString(size_t stringHandle);
6626 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6628 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6629 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6631 static fgWalkPreFn CountSharedStaticHelper;
6632 static bool IsSharedStaticHelper(GenTreePtr tree);
6633 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6635 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6636 // returns true/false if 'field' is a Jit Data offset
6637 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6638 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6639 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6641 /*****************************************************************************/
6646 enum TEMP_USAGE_TYPE
6652 static var_types tmpNormalizeType(var_types type);
6653 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6654 void tmpRlsTemp(TempDsc* temp);
6655 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6658 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6659 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6663 bool tmpAllFree() const;
6666 #ifndef LEGACY_BACKEND
6667 void tmpPreAllocateTemps(var_types type, unsigned count);
6668 #endif // !LEGACY_BACKEND
6671 #ifdef LEGACY_BACKEND
6672 unsigned tmpIntSpillMax; // number of int-sized spill temps
6673 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6674 #endif // LEGACY_BACKEND
6676 unsigned tmpCount; // Number of temps
6677 unsigned tmpSize; // Size of all the temps
6680 // Used by RegSet::rsSpillChk()
6681 unsigned tmpGetCount; // Temps which haven't been released yet
6684 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6686 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6687 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6690 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6691 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6695 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6696 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6700 CodeGenInterface* codeGen;
6702 // The following holds information about instr offsets in terms of generated code.
6706 IPmappingDsc* ipmdNext; // next line# record
6707 IL_OFFSETX ipmdILoffsx; // the instr offset
6708 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6709 bool ipmdIsLabel; // Can this code be a branch label?
6712 // Record the instr offset mapping to the generated code
6714 IPmappingDsc* genIPmappingList;
6715 IPmappingDsc* genIPmappingLast;
6717 // Managed RetVal - A side hash table meant to record the mapping from a
6718 // GT_CALL node to its IL offset. This info is used to emit sequence points
6719 // that can be used by debugger to determine the native offset at which the
6720 // managed RetVal will be available.
6722 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6723 // favor of a side table for two reasons: 1) We need IL offset for only those
6724 // GT_CALL nodes (created during importation) that correspond to an IL call and
6725 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6726 // structure and IL offset is needed only when generating debuggable code. Therefore
6727 // it is desirable to avoid memory size penalty in retail scenarios.
6728 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6729 CallSiteILOffsetTable;
6730 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6732 unsigned genReturnLocal; // Local number for the return value when applicable.
6733 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6735 // The following properties are part of CodeGenContext. Getters are provided here for
6736 // convenience and backward compatibility, but the properties can only be set by invoking
6737 // the setter on CodeGenContext directly.
6739 __declspec(property(get = getEmitter)) emitter* genEmitter;
6740 emitter* getEmitter()
6742 return codeGen->getEmitter();
6745 const bool isFramePointerUsed()
6747 return codeGen->isFramePointerUsed();
6750 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6751 bool getInterruptible()
6753 return codeGen->genInterruptible;
6755 void setInterruptible(bool value)
6757 codeGen->setInterruptible(value);
6761 const bool genDoubleAlign()
6763 return codeGen->doDoubleAlign();
6765 DWORD getCanDoubleAlign();
6766 bool shouldDoubleAlign(unsigned refCntStk,
6768 unsigned refCntWtdReg,
6769 unsigned refCntStkParam,
6770 unsigned refCntWtdStkDbl);
6771 #endif // DOUBLE_ALIGN
6773 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6774 bool getFullPtrRegMap()
6776 return codeGen->genFullPtrRegMap;
6778 void setFullPtrRegMap(bool value)
6780 codeGen->setFullPtrRegMap(value);
6783 // Things that MAY belong either in CodeGen or CodeGenContext
6785 #if FEATURE_EH_FUNCLETS
6786 FuncInfoDsc* compFuncInfos;
6787 unsigned short compCurrFuncIdx;
6788 unsigned short compFuncInfoCount;
6790 unsigned short compFuncCount()
6792 assert(fgFuncletsCreated);
6793 return compFuncInfoCount;
6796 #else // !FEATURE_EH_FUNCLETS
6798 // This is a no-op when there are no funclets!
6799 void genUpdateCurrentFunclet(BasicBlock* block)
6804 FuncInfoDsc compFuncInfoRoot;
6806 static const unsigned compCurrFuncIdx = 0;
6808 unsigned short compFuncCount()
6813 #endif // !FEATURE_EH_FUNCLETS
6815 FuncInfoDsc* funCurrentFunc();
6816 void funSetCurrentFunc(unsigned funcIdx);
6817 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6818 unsigned int funGetFuncIdx(BasicBlock* block);
6822 VARSET_TP compCurLife; // current live variables
6823 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6825 template <bool ForCodeGen>
6826 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6828 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6830 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6833 template <bool ForCodeGen>
6834 void compUpdateLife(GenTreePtr tree);
6836 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6837 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6838 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6839 template <bool ForCodeGen>
6840 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6842 template <bool ForCodeGen>
6843 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6845 // Gets a register mask that represent the kill set for a helper call since
6846 // not all JIT Helper calls follow the standard ABI on the target architecture.
6847 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6849 // Gets a register mask that represent the kill set for a NoGC helper call.
6850 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6853 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6854 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6855 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6856 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6857 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6858 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6859 #endif // _TARGET_ARM_
6861 // 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
6863 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6865 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6866 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6867 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6868 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6869 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6870 // for the tracked var indices of the field vars, as in a live var set).
6871 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6873 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6875 if (m_promotedStructDeathVars == nullptr)
6877 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6879 return m_promotedStructDeathVars;
6883 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6884 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6888 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6889 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6892 #if !defined(__GNUC__)
6893 #pragma region Unwind information
6898 // Infrastructure functions: start/stop/reserve/emit.
6901 void unwindBegProlog();
6902 void unwindEndProlog();
6903 void unwindBegEpilog();
6904 void unwindEndEpilog();
6905 void unwindReserve();
6906 void unwindEmit(void* pHotCode, void* pColdCode);
6909 // Specific unwind information functions: called by code generation to indicate a particular
6910 // prolog or epilog unwindable instruction has been generated.
6913 void unwindPush(regNumber reg);
6914 void unwindAllocStack(unsigned size);
6915 void unwindSetFrameReg(regNumber reg, unsigned offset);
6916 void unwindSaveReg(regNumber reg, unsigned offset);
6918 #if defined(_TARGET_ARM_)
6919 void unwindPushMaskInt(regMaskTP mask);
6920 void unwindPushMaskFloat(regMaskTP mask);
6921 void unwindPopMaskInt(regMaskTP mask);
6922 void unwindPopMaskFloat(regMaskTP mask);
6923 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6924 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6925 // called via unwindPadding().
6926 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6927 // instruction and the current location.
6928 #endif // _TARGET_ARM_
6930 #if defined(_TARGET_ARM64_)
6932 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6933 // instruction and the current location.
6934 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6935 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6936 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6937 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6938 void unwindSaveNext(); // unwind code: save_next
6939 void unwindReturn(regNumber reg); // ret lr
6940 #endif // defined(_TARGET_ARM64_)
6943 // Private "helper" functions for the unwind implementation.
6947 #if FEATURE_EH_FUNCLETS
6948 void unwindGetFuncLocations(FuncInfoDsc* func,
6949 bool getHotSectionData,
6950 /* OUT */ emitLocation** ppStartLoc,
6951 /* OUT */ emitLocation** ppEndLoc);
6952 #endif // FEATURE_EH_FUNCLETS
6954 void unwindReserveFunc(FuncInfoDsc* func);
6955 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6957 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6959 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6960 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6962 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6964 #if defined(_TARGET_AMD64_)
6966 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6968 void unwindBegPrologWindows();
6969 void unwindPushWindows(regNumber reg);
6970 void unwindAllocStackWindows(unsigned size);
6971 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6972 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6974 #ifdef UNIX_AMD64_ABI
6975 void unwindBegPrologCFI();
6976 void unwindPushCFI(regNumber reg);
6977 void unwindAllocStackCFI(unsigned size);
6978 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6979 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6980 int mapRegNumToDwarfReg(regNumber reg);
6981 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6982 #endif // UNIX_AMD64_ABI
6983 #elif defined(_TARGET_ARM_)
6985 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6986 void unwindPushPopMaskFloat(regMaskTP mask);
6987 void unwindSplit(FuncInfoDsc* func);
6989 #endif // _TARGET_ARM_
6991 #if !defined(__GNUC__)
6992 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6996 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6997 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7001 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7002 XX that contains the distinguished, well-known SIMD type definitions). XX
7004 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7005 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7008 // Get highest available instruction set for floating point codegen
7009 InstructionSet getFloatingPointInstructionSet()
7011 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7014 return InstructionSet_AVX;
7019 return InstructionSet_SSE3_4;
7023 assert(canUseSSE2());
7024 return InstructionSet_SSE2;
7026 assert(!"getFPInstructionSet() is not implemented for target arch");
7028 return InstructionSet_NONE;
7032 // Get highest available instruction set for SIMD codegen
7033 InstructionSet getSIMDInstructionSet()
7035 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7036 return getFloatingPointInstructionSet();
7038 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7040 return InstructionSet_NONE;
7046 // Should we support SIMD intrinsics?
7049 // Have we identified any SIMD types?
7050 // This is currently used by struct promotion to avoid getting type information for a struct
7051 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7053 bool _usesSIMDTypes;
7054 bool usesSIMDTypes()
7056 return _usesSIMDTypes;
7058 void setUsesSIMDTypes(bool value)
7060 _usesSIMDTypes = value;
7063 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7064 // that require indexed access to the individual fields of the vector, which is not well supported
7065 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7066 unsigned lvaSIMDInitTempVarNum;
7069 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7070 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7071 CORINFO_CLASS_HANDLE SIMDIntHandle;
7072 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7073 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7074 CORINFO_CLASS_HANDLE SIMDShortHandle;
7075 CORINFO_CLASS_HANDLE SIMDByteHandle;
7076 CORINFO_CLASS_HANDLE SIMDLongHandle;
7077 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7078 CORINFO_CLASS_HANDLE SIMDULongHandle;
7079 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7080 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7081 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7082 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7084 // Get the handle for a SIMD type.
7085 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7087 if (simdBaseType == TYP_FLOAT)
7092 return SIMDVector2Handle;
7094 return SIMDVector3Handle;
7096 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7098 return SIMDVector4Handle;
7107 assert(simdType == getSIMDVectorType());
7108 switch (simdBaseType)
7111 return SIMDFloatHandle;
7113 return SIMDDoubleHandle;
7115 return SIMDIntHandle;
7117 return SIMDUShortHandle;
7119 return SIMDUShortHandle;
7121 return SIMDUByteHandle;
7123 return SIMDShortHandle;
7125 return SIMDByteHandle;
7127 return SIMDLongHandle;
7129 return SIMDUIntHandle;
7131 return SIMDULongHandle;
7133 assert(!"Didn't find a class handle for simdType");
7135 return NO_CLASS_HANDLE;
7139 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7140 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7141 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7143 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7144 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7145 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7146 bool isSIMDTypeLocal(GenTree* tree)
7148 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7151 // Returns true if the type of the tree is a byref of TYP_SIMD
7152 bool isAddrOfSIMDType(GenTree* tree)
7154 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7156 switch (tree->OperGet())
7159 return varTypeIsSIMD(tree->gtGetOp1());
7161 case GT_LCL_VAR_ADDR:
7162 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7165 return isSIMDTypeLocal(tree);
7172 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7174 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7175 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7176 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7179 // Returns base type of a TYP_SIMD local.
7180 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7181 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7183 if (isSIMDTypeLocal(tree))
7185 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7191 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7193 return info.compCompHnd->isInSIMDModule(clsHnd);
7196 bool isSIMDClass(typeInfo* pTypeInfo)
7198 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7201 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7202 // if it is not a SIMD type or is an unsupported base type.
7203 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7205 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7207 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7210 // Get SIMD Intrinsic info given the method handle.
7211 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7212 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7213 CORINFO_METHOD_HANDLE methodHnd,
7214 CORINFO_SIG_INFO* sig,
7217 var_types* baseType,
7218 unsigned* sizeBytes);
7220 // Pops and returns GenTree node from importers type stack.
7221 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7222 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7224 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7225 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7227 // Creates a GT_SIMD tree for Select operation
7228 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7230 unsigned simdVectorSize,
7235 // Creates a GT_SIMD tree for Min/Max operation
7236 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7237 CORINFO_CLASS_HANDLE typeHnd,
7239 unsigned simdVectorSize,
7243 // Transforms operands and returns the SIMD intrinsic to be applied on
7244 // transformed operands to obtain given relop result.
7245 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7246 CORINFO_CLASS_HANDLE typeHnd,
7247 unsigned simdVectorSize,
7248 var_types* baseType,
7252 // Creates a GT_SIMD tree for Abs intrinsic.
7253 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7255 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7256 // Transforms operands and returns the SIMD intrinsic to be applied on
7257 // transformed operands to obtain == comparison result.
7258 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7259 unsigned simdVectorSize,
7263 // Transforms operands and returns the SIMD intrinsic to be applied on
7264 // transformed operands to obtain > comparison result.
7265 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7266 unsigned simdVectorSize,
7270 // Transforms operands and returns the SIMD intrinsic to be applied on
7271 // transformed operands to obtain >= comparison result.
7272 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7273 unsigned simdVectorSize,
7277 // Transforms operands and returns the SIMD intrinsic to be applied on
7278 // transformed operands to obtain >= comparison result in case of int32
7279 // and small int base type vectors.
7280 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7281 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7282 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7284 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7285 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7286 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7287 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7288 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7290 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7291 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7292 GenTreePtr newobjThis,
7293 CORINFO_CLASS_HANDLE clsHnd,
7294 CORINFO_METHOD_HANDLE method,
7295 CORINFO_SIG_INFO* sig,
7298 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7300 // Whether SIMD vector occupies part of SIMD register.
7301 // SSE2: vector2f/3f are considered sub register SIMD types.
7302 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7303 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7305 unsigned sizeBytes = 0;
7306 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7307 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7310 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7312 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7315 // Get the type for the hardware SIMD vector.
7316 // This is the maximum SIMD type supported for this target.
7317 var_types getSIMDVectorType()
7319 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7326 assert(canUseSSE2());
7330 assert(!"getSIMDVectorType() unimplemented on target arch");
7335 // Get the size of the SIMD type in bytes
7336 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7338 unsigned sizeBytes = 0;
7339 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7343 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7344 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7346 // Get the the number of elements of basetype of SIMD vector given by its type handle
7347 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7349 // Get preferred alignment of SIMD type.
7350 int getSIMDTypeAlignment(var_types simdType);
7352 // Get the number of bytes in a SIMD Vector for the current compilation.
7353 unsigned getSIMDVectorRegisterByteLength()
7355 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7358 return YMM_REGSIZE_BYTES;
7362 assert(canUseSSE2());
7363 return XMM_REGSIZE_BYTES;
7366 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7371 // The minimum and maximum possible number of bytes in a SIMD vector.
7372 unsigned int maxSIMDStructBytes()
7374 return getSIMDVectorRegisterByteLength();
7376 unsigned int minSIMDStructBytes()
7378 return emitTypeSize(TYP_SIMD8);
7381 #ifdef FEATURE_AVX_SUPPORT
7382 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7383 static const unsigned maxPossibleSIMDStructBytes = 32;
7384 #else // !FEATURE_AVX_SUPPORT
7385 static const unsigned maxPossibleSIMDStructBytes = 16;
7386 #endif // !FEATURE_AVX_SUPPORT
7388 // Returns the codegen type for a given SIMD size.
7389 var_types getSIMDTypeForSize(unsigned size)
7391 var_types simdType = TYP_UNDEF;
7394 simdType = TYP_SIMD8;
7396 else if (size == 12)
7398 simdType = TYP_SIMD12;
7400 else if (size == 16)
7402 simdType = TYP_SIMD16;
7404 #ifdef FEATURE_AVX_SUPPORT
7405 else if (size == 32)
7407 simdType = TYP_SIMD32;
7409 #endif // FEATURE_AVX_SUPPORT
7412 noway_assert(!"Unexpected size for SIMD type");
7417 unsigned getSIMDInitTempVarNum()
7419 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7421 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7422 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7424 return lvaSIMDInitTempVarNum;
7427 #endif // FEATURE_SIMD
7430 //------------------------------------------------------------------------
7431 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7433 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7434 // candidate for enregistration.
7436 unsigned largestEnregisterableStructSize()
7439 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7440 if (vectorRegSize > TARGET_POINTER_SIZE)
7442 return vectorRegSize;
7445 #endif // FEATURE_SIMD
7447 return TARGET_POINTER_SIZE;
7452 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7453 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7454 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7456 // Is this var is of type simd struct?
7457 bool lclVarIsSIMDType(unsigned varNum)
7459 LclVarDsc* varDsc = lvaTable + varNum;
7460 return varDsc->lvIsSIMDType();
7463 // Is this Local node a SIMD local?
7464 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7466 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7469 // Returns true if the TYP_SIMD locals on stack are aligned at their
7470 // preferred byte boundary specified by getSIMDTypeAlignment().
7472 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7473 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7474 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7475 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7476 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7477 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7478 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7481 bool isSIMDTypeLocalAligned(unsigned varNum)
7483 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7484 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7487 int off = lvaFrameAddress(varNum, &ebpBased);
7488 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7489 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7490 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7493 #endif // FEATURE_SIMD
7498 // Whether SSE2 is available
7499 bool canUseSSE2() const
7501 #ifdef _TARGET_XARCH_
7502 return opts.compCanUseSSE2;
7508 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7509 bool CanUseSSE3_4() const
7511 #ifdef _TARGET_XARCH_
7512 return opts.compCanUseSSE3_4;
7518 bool canUseAVX() const
7520 #ifdef FEATURE_AVX_SUPPORT
7521 return opts.compCanUseAVX;
7528 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7529 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7533 XX Generic info about the compilation and the method being compiled. XX
7534 XX It is responsible for driving the other phases. XX
7535 XX It is also responsible for all the memory management. XX
7537 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7538 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7542 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7544 InlineResult* compInlineResult; // The result of importing the inlinee method.
7546 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7547 bool compJmpOpUsed; // Does the method do a JMP
7548 bool compLongUsed; // Does the method use TYP_LONG
7549 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7550 bool compTailCallUsed; // Does the method do a tailcall
7551 bool compLocallocUsed; // Does the method use localloc.
7552 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7553 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7554 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7556 // NOTE: These values are only reliable after
7557 // the importing is completely finished.
7559 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7560 // we can iterate over these efficiently.
7562 #if CPU_USES_BLOCK_MOVE
7563 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7567 // State information - which phases have completed?
7568 // These are kept together for easy discoverability
7570 bool bRangeAllowStress;
7571 bool compCodeGenDone;
7572 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7573 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7574 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7575 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7578 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7579 bool fgLocalVarLivenessChanged;
7581 bool compStackProbePrologDone;
7583 #ifndef LEGACY_BACKEND
7585 #endif // !LEGACY_BACKEND
7586 bool compRationalIRForm;
7588 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7590 bool compGeneratingProlog;
7591 bool compGeneratingEpilog;
7592 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7593 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7594 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7595 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7596 bool getNeedsGSSecurityCookie() const
7598 return compNeedsGSSecurityCookie;
7600 void setNeedsGSSecurityCookie()
7602 compNeedsGSSecurityCookie = true;
7605 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7606 // frame layout calculations, this is the level we are currently
7609 //---------------------------- JITing options -----------------------------
7622 JitFlags* jitFlags; // all flags passed from the EE
7623 unsigned compFlags; // method attributes
7625 codeOptimize compCodeOpt; // what type of code optimizations
7629 #ifdef _TARGET_XARCH_
7630 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7631 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7633 #ifdef FEATURE_AVX_SUPPORT
7634 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7635 #endif // FEATURE_AVX_SUPPORT
7636 #endif // _TARGET_XARCH_
7638 // optimize maximally and/or favor speed over size?
7640 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7641 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7642 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7643 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7644 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7646 // Maximun number of locals before turning off the inlining
7647 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7650 unsigned instrCount;
7651 unsigned lvRefCount;
7652 bool compMinOptsIsSet;
7654 bool compMinOptsIsUsed;
7656 inline bool MinOpts()
7658 assert(compMinOptsIsSet);
7659 compMinOptsIsUsed = true;
7662 inline bool IsMinOptsSet()
7664 return compMinOptsIsSet;
7667 inline bool MinOpts()
7671 inline bool IsMinOptsSet()
7673 return compMinOptsIsSet;
7676 inline void SetMinOpts(bool val)
7678 assert(!compMinOptsIsUsed);
7679 assert(!compMinOptsIsSet || (compMinOpts == val));
7681 compMinOptsIsSet = true;
7684 // true if the CLFLG_* for an optimization is set.
7685 inline bool OptEnabled(unsigned optFlag)
7687 return !!(compFlags & optFlag);
7690 #ifdef FEATURE_READYTORUN_COMPILER
7691 inline bool IsReadyToRun()
7693 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7696 inline bool IsReadyToRun()
7702 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7703 // PInvoke transitions inline (e.g. when targeting CoreRT).
7704 inline bool ShouldUsePInvokeHelpers()
7706 #if COR_JIT_EE_VERSION > 460
7707 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7713 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7715 inline bool IsReversePInvoke()
7717 #if COR_JIT_EE_VERSION > 460
7718 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7724 // true if we must generate code compatible with JIT32 quirks
7725 inline bool IsJit32Compat()
7727 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7728 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7734 // true if we must generate code compatible with Jit64 quirks
7735 inline bool IsJit64Compat()
7737 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7738 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7739 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7746 bool compScopeInfo; // Generate the LocalVar info ?
7747 bool compDbgCode; // Generate debugger-friendly code?
7748 bool compDbgInfo; // Gather debugging info?
7751 #ifdef PROFILING_SUPPORTED
7752 bool compNoPInvokeInlineCB;
7754 static const bool compNoPInvokeInlineCB;
7758 bool compGcChecks; // Check arguments and return values to ensure they are sane
7759 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7760 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7764 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7765 // to be allocated on the stack.
7766 // It will be set to true in the following cases:
7767 // 1. When the method being compiled has a declarative security
7768 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7769 // This is also the case when we inject a prolog and epilog in the method.
7771 // 2. When the method being compiled has imperative security (i.e. the method
7772 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7774 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7776 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7777 // which gets reported as a GC root to stackwalker.
7778 // (See also ICodeManager::GetAddrOfSecurityObject.)
7785 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7786 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7790 #ifdef UNIX_AMD64_ABI
7791 // This flag is indicating if there is a need to align the frame.
7792 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7793 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7794 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7795 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7796 // there are calls and making sure the frame alignment logic is executed.
7797 bool compNeedToAlignFrame;
7798 #endif // UNIX_AMD64_ABI
7800 bool compProcedureSplitting; // Separate cold code from hot code
7802 bool genFPorder; // Preserve FP order (operations are non-commutative)
7803 bool genFPopt; // Can we do frame-pointer-omission optimization?
7804 bool altJit; // True if we are an altjit and are compiling this method
7807 bool optRepeat; // Repeat optimizer phases k times
7811 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7812 bool dspCode; // Display native code generated
7813 bool dspEHTable; // Display the EH table reported to the VM
7814 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7815 bool dspEmit; // Display emitter output
7816 bool dspLines; // Display source-code lines intermixed with native code output
7817 bool dmpHex; // Display raw bytes in hex of native code output
7818 bool varNames; // Display variables names in native code output
7819 bool disAsm; // Display native code as it is generated
7820 bool disAsmSpilled; // Display native code when any register spilling occurs
7821 bool disDiffable; // Makes the Disassembly code 'diff-able'
7822 bool disAsm2; // Display native code after it is generated using external disassembler
7823 bool dspOrder; // Display names of each of the methods that we ngen/jit
7824 bool dspUnwind; // Display the unwind info output
7825 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7826 bool compLongAddress; // Force using large pseudo instructions for long address
7827 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7828 bool dspGCtbls; // Display the GC tables
7832 bool doLateDisasm; // Run the late disassembler
7833 #endif // LATE_DISASM
7835 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7836 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7837 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7838 static const bool dspGCtbls = true;
7841 // We need stack probes to guarantee that we won't trigger a stack overflow
7842 // when calling unmanaged code until they get a chance to set up a frame, because
7843 // the EE will have no idea where it is.
7845 // We will only be doing this currently for hosted environments. Unfortunately
7846 // we need to take care of stubs, so potentially, we will have to do the probes
7847 // for any call. We have a plan for not needing for stubs though
7848 bool compNeedStackProbes;
7850 #ifdef PROFILING_SUPPORTED
7851 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7852 // This option helps make the JIT behave as if it is running under a profiler.
7853 bool compJitELTHookEnabled;
7854 #endif // PROFILING_SUPPORTED
7856 #if FEATURE_TAILCALL_OPT
7857 // Whether opportunistic or implicit tail call optimization is enabled.
7858 bool compTailCallOpt;
7859 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7860 bool compTailCallLoopOpt;
7864 static const bool compUseSoftFP = true;
7865 #else // !ARM_SOFTFP
7866 static const bool compUseSoftFP = false;
7869 GCPollType compGCPollType;
7873 static bool s_pAltJitExcludeAssembliesListInitialized;
7874 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7879 template <typename T>
7882 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7885 template <typename T>
7888 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7891 static int dspTreeID(GenTree* tree)
7893 return tree->gtTreeID;
7895 static void printTreeID(GenTree* tree)
7897 if (tree == nullptr)
7903 printf("[%06d]", dspTreeID(tree));
7910 #define STRESS_MODES \
7914 /* "Variations" stress areas which we try to mix up with each other. */ \
7915 /* These should not be exhaustively used as they might */ \
7916 /* hide/trivialize other areas */ \
7919 STRESS_MODE(DBL_ALN) \
7920 STRESS_MODE(LCL_FLDS) \
7921 STRESS_MODE(UNROLL_LOOPS) \
7922 STRESS_MODE(MAKE_CSE) \
7923 STRESS_MODE(LEGACY_INLINE) \
7924 STRESS_MODE(CLONE_EXPR) \
7925 STRESS_MODE(USE_FCOMI) \
7926 STRESS_MODE(USE_CMOV) \
7928 STRESS_MODE(BB_PROFILE) \
7929 STRESS_MODE(OPT_BOOLS_GC) \
7930 STRESS_MODE(REMORPH_TREES) \
7931 STRESS_MODE(64RSLT_MUL) \
7932 STRESS_MODE(DO_WHILE_LOOPS) \
7933 STRESS_MODE(MIN_OPTS) \
7934 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7935 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7936 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7937 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7938 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7939 STRESS_MODE(NULL_OBJECT_CHECK) \
7940 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7941 STRESS_MODE(RANDOM_INLINE) \
7942 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
7943 STRESS_MODE(GENERIC_VARN) \
7945 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
7947 STRESS_MODE(COUNT_VARN) \
7949 /* "Check" stress areas that can be exhaustively used if we */ \
7950 /* dont care about performance at all */ \
7952 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7953 STRESS_MODE(CHK_FLOW_UPDATE) \
7954 STRESS_MODE(EMITTER) \
7955 STRESS_MODE(CHK_REIMPORT) \
7956 STRESS_MODE(FLATFP) \
7957 STRESS_MODE(GENERIC_CHECK) \
7962 #define STRESS_MODE(mode) STRESS_##mode,
7969 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7970 BYTE compActiveStressModes[STRESS_COUNT];
7973 #define MAX_STRESS_WEIGHT 100
7975 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7979 bool compInlineStress()
7981 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7984 bool compRandomInlineStress()
7986 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7991 bool compTailCallStress()
7994 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8000 codeOptimize compCodeOpt()
8003 // Switching between size & speed has measurable throughput impact
8004 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8005 // DEBUG, but should generate identical code between CHK & RET builds,
8006 // so that's not acceptable.
8007 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8008 // Investigate the cause of the throughput regression.
8010 return opts.compCodeOpt;
8012 return BLENDED_CODE;
8016 //--------------------- Info about the procedure --------------------------
8020 COMP_HANDLE compCompHnd;
8021 CORINFO_MODULE_HANDLE compScopeHnd;
8022 CORINFO_CLASS_HANDLE compClassHnd;
8023 CORINFO_METHOD_HANDLE compMethodHnd;
8024 CORINFO_METHOD_INFO* compMethodInfo;
8026 BOOL hasCircularClassConstraints;
8027 BOOL hasCircularMethodConstraints;
8029 #if defined(DEBUG) || defined(LATE_DISASM)
8030 const char* compMethodName;
8031 const char* compClassName;
8032 const char* compFullName;
8033 #endif // defined(DEBUG) || defined(LATE_DISASM)
8035 #if defined(DEBUG) || defined(INLINE_DATA)
8036 // Method hash is logcally const, but computed
8038 mutable unsigned compMethodHashPrivate;
8039 unsigned compMethodHash() const;
8040 #endif // defined(DEBUG) || defined(INLINE_DATA)
8042 #ifdef PSEUDORANDOM_NOP_INSERTION
8043 // things for pseudorandom nop insertion
8044 unsigned compChecksum;
8048 // The following holds the FLG_xxxx flags for the method we're compiling.
8051 // The following holds the class attributes for the method we're compiling.
8052 unsigned compClassAttr;
8054 const BYTE* compCode;
8055 IL_OFFSET compILCodeSize; // The IL code size
8056 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8057 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8058 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8059 // (2) the code is hot/cold split, and we issued less code than we expected
8060 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8062 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8063 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8064 bool compIsContextful : 1; // contextful method
8065 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8066 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8067 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8068 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8069 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8071 var_types compRetType; // Return type of the method as declared in IL
8072 var_types compRetNativeType; // Normalized return type as per target arch ABI
8073 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8074 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8075 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8076 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8077 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8078 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8079 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8080 unsigned compMaxStack;
8081 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8082 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8084 unsigned compCallUnmanaged; // count of unmanaged calls
8085 unsigned compLvFrameListRoot; // lclNum for the Frame root
8086 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8087 // You should generally use compHndBBtabCount instead: it is the
8088 // current number of EH clauses (after additions like synchronized
8089 // methods and funclets, and removals like unreachable code deletion).
8091 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8092 // and the VM expects that, or the JIT is a "self-host" compiler
8093 // (e.g., x86 hosted targeting x86) and the VM expects that.
8095 /* The following holds IL scope information about local variables.
8098 unsigned compVarScopesCount;
8099 VarScopeDsc* compVarScopes;
8101 /* The following holds information about instr offsets for
8102 * which we need to report IP-mappings
8105 IL_OFFSET* compStmtOffsets; // sorted
8106 unsigned compStmtOffsetsCount;
8107 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8109 #define CPU_X86 0x0100 // The generic X86 CPU
8110 #define CPU_X86_PENTIUM_4 0x0110
8112 #define CPU_X64 0x0200 // The generic x64 CPU
8113 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8114 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8116 #define CPU_ARM 0x0300 // The generic ARM CPU
8118 unsigned genCPU; // What CPU are we running on
8121 // Returns true if the method being compiled returns a non-void and non-struct value.
8122 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8123 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8124 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8125 // Methods returning such structs are considered to return non-struct return value and
8126 // this method returns true in that case.
8127 bool compMethodReturnsNativeScalarType()
8129 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8132 // Returns true if the method being compiled returns RetBuf addr as its return value
8133 bool compMethodReturnsRetBufAddr()
8135 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8136 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8138 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8139 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8140 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8141 // methods with hidden RetBufArg.
8143 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8144 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8145 // returning the address of RetBuf.
8147 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8148 // to be returned in RAX.
8149 CLANG_FORMAT_COMMENT_ANCHOR;
8151 #ifdef _TARGET_AMD64_
8152 return (info.compRetBuffArg != BAD_VAR_NUM);
8153 #else // !_TARGET_AMD64_
8154 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8155 #endif // !_TARGET_AMD64_
8158 // Returns true if the method returns a value in more than one return register
8159 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8160 // TODO-ARM64: Does this apply for ARM64 too?
8161 bool compMethodReturnsMultiRegRetType()
8163 #if FEATURE_MULTIREG_RET
8164 #if defined(_TARGET_X86_)
8165 // On x86 only 64-bit longs are returned in multiple registers
8166 return varTypeIsLong(info.compRetNativeType);
8167 #else // targets: X64-UNIX, ARM64 or ARM32
8168 // On all other targets that support multireg return values:
8169 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8170 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8171 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8172 #endif // TARGET_XXX
8174 #else // not FEATURE_MULTIREG_RET
8176 // For this architecture there are no multireg returns
8179 #endif // FEATURE_MULTIREG_RET
8182 #if FEATURE_MULTIREG_ARGS
8183 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8184 // return the gcPtr layout for the pointers sized fields
8185 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8186 #endif // FEATURE_MULTIREG_ARGS
8188 // Returns true if the method being compiled returns a value
8189 bool compMethodHasRetVal()
8191 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8192 compMethodReturnsMultiRegRetType();
8197 void compDispLocalVars();
8201 //-------------------------- Global Compiler Data ------------------------------------
8204 static unsigned s_compMethodsCount; // to produce unique label names
8205 unsigned compGenTreeID;
8208 BasicBlock* compCurBB; // the current basic block in process
8209 GenTreePtr compCurStmt; // the current statement in process
8211 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8214 // The following is used to create the 'method JIT info' block.
8215 size_t compInfoBlkSize;
8216 BYTE* compInfoBlkAddr;
8218 EHblkDsc* compHndBBtab; // array of EH data
8219 unsigned compHndBBtabCount; // element count of used elements in EH data array
8220 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8222 #if defined(_TARGET_X86_)
8224 //-------------------------------------------------------------------------
8225 // Tracking of region covered by the monitor in synchronized methods
8226 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8227 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8229 #endif // !_TARGET_X86_
8231 Phases previousCompletedPhase; // the most recently completed phase
8233 //-------------------------------------------------------------------------
8234 // The following keeps track of how many bytes of local frame space we've
8235 // grabbed so far in the current function, and how many argument bytes we
8236 // need to pop when we return.
8239 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8241 // Count of callee-saved regs we pushed in the prolog.
8242 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8243 // In case of Amd64 this doesn't include float regs saved on stack.
8244 unsigned compCalleeRegsPushed;
8246 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8247 // Mask of callee saved float regs on stack.
8248 regMaskTP compCalleeFPRegsSavedMask;
8250 #ifdef _TARGET_AMD64_
8251 // Quirk for VS debug-launch scenario to work:
8252 // Bytes of padding between save-reg area and locals.
8253 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8254 unsigned compVSQuirkStackPaddingNeeded;
8255 bool compQuirkForPPPflag;
8258 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8260 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8261 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8262 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8264 //-------------------------------------------------------------------------
8266 static void compStartup(); // One-time initialization
8267 static void compShutdown(); // One-time finalization
8269 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8272 static void compDisplayStaticSizes(FILE* fout);
8274 //------------ Some utility functions --------------
8276 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8277 void** ppIndirection); /* OUT */
8279 // Several JIT/EE interface functions return a CorInfoType, and also return a
8280 // class handle as an out parameter if the type is a value class. Returns the
8281 // size of the type these describe.
8282 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8285 // Components used by the compiler may write unit test suites, and
8286 // have them run within this method. They will be run only once per process, and only
8287 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8288 // These should fail by asserting.
8289 void compDoComponentUnitTestsOnce();
8292 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8293 CORINFO_MODULE_HANDLE classPtr,
8294 COMP_HANDLE compHnd,
8295 CORINFO_METHOD_INFO* methodInfo,
8296 void** methodCodePtr,
8297 ULONG* methodCodeSize,
8298 JitFlags* compileFlags);
8299 void compCompileFinish();
8300 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8301 COMP_HANDLE compHnd,
8302 CORINFO_METHOD_INFO* methodInfo,
8303 void** methodCodePtr,
8304 ULONG* methodCodeSize,
8305 JitFlags* compileFlags,
8306 CorInfoInstantiationVerification instVerInfo);
8308 ArenaAllocator* compGetAllocator();
8310 #if MEASURE_MEM_ALLOC
8312 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8316 unsigned allocCnt; // # of allocs
8317 UINT64 allocSz; // total size of those alloc.
8318 UINT64 allocSzMax; // Maximum single allocation.
8319 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8320 UINT64 nraTotalSizeAlloc;
8321 UINT64 nraTotalSizeUsed;
8323 static const char* s_CompMemKindNames[]; // Names of the kinds.
8325 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8327 for (int i = 0; i < CMK_Count; i++)
8329 allocSzByKind[i] = 0;
8332 MemStats(const MemStats& ms)
8333 : allocCnt(ms.allocCnt)
8334 , allocSz(ms.allocSz)
8335 , allocSzMax(ms.allocSzMax)
8336 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8337 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8339 for (int i = 0; i < CMK_Count; i++)
8341 allocSzByKind[i] = ms.allocSzByKind[i];
8345 // Until we have ubiquitous constructors.
8348 this->MemStats::MemStats();
8351 void AddAlloc(size_t sz, CompMemKind cmk)
8355 if (sz > allocSzMax)
8359 allocSzByKind[cmk] += sz;
8362 void Print(FILE* f); // Print these stats to f.
8363 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8365 MemStats genMemStats;
8367 struct AggregateMemStats : public MemStats
8371 AggregateMemStats() : MemStats(), nMethods(0)
8375 void Add(const MemStats& ms)
8378 allocCnt += ms.allocCnt;
8379 allocSz += ms.allocSz;
8380 allocSzMax = max(allocSzMax, ms.allocSzMax);
8381 for (int i = 0; i < CMK_Count; i++)
8383 allocSzByKind[i] += ms.allocSzByKind[i];
8385 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8386 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8389 void Print(FILE* f); // Print these stats to jitstdout.
8392 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8393 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8394 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8396 #endif // MEASURE_MEM_ALLOC
8398 #if LOOP_HOIST_STATS
8399 unsigned m_loopsConsidered;
8400 bool m_curLoopHasHoistedExpression;
8401 unsigned m_loopsWithHoistedExpressions;
8402 unsigned m_totalHoistedExpressions;
8404 void AddLoopHoistStats();
8405 void PrintPerMethodLoopHoistStats();
8407 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8408 static unsigned s_loopsConsidered;
8409 static unsigned s_loopsWithHoistedExpressions;
8410 static unsigned s_totalHoistedExpressions;
8412 static void PrintAggregateLoopHoistStats(FILE* f);
8413 #endif // LOOP_HOIST_STATS
8415 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8416 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8417 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8418 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8419 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8420 void compFreeMem(void*);
8422 bool compIsForImportOnly();
8423 bool compIsForInlining();
8424 bool compDonotInline();
8427 const char* compLocalVarName(unsigned varNum, unsigned offs);
8428 VarName compVarName(regNumber reg, bool isFloatReg = false);
8429 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8430 const char* compRegPairName(regPairNo regPair);
8431 const char* compRegNameForSize(regNumber reg, size_t size);
8432 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8433 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8434 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8437 //-------------------------------------------------------------------------
8439 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8441 struct VarScopeMapInfo
8443 VarScopeListNode* head;
8444 VarScopeListNode* tail;
8445 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8447 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8454 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8455 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8457 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8458 VarNumToScopeDscMap;
8460 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8461 VarNumToScopeDscMap* compVarScopeMap;
8463 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8465 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8467 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8469 void compInitVarScopeMap();
8471 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8472 // enter scope, sorted by instr offset
8473 unsigned compNextEnterScope;
8475 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8476 // go out of scope, sorted by instr offset
8477 unsigned compNextExitScope;
8479 void compInitScopeLists();
8481 void compResetScopeLists();
8483 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8485 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8487 void compProcessScopesUntil(unsigned offset,
8489 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8490 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8493 void compDispScopeLists();
8496 bool compIsProfilerHookNeeded();
8498 //-------------------------------------------------------------------------
8499 /* Statistical Data Gathering */
8501 void compJitStats(); // call this function and enable
8502 // various ifdef's below for statistical data
8505 void compCallArgStats();
8506 static void compDispCallArgStats(FILE* fout);
8509 //-------------------------------------------------------------------------
8516 ArenaAllocator* compAllocator;
8519 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8520 // suitable for use by utilcode collection types.
8521 IAllocator* compAsIAllocator;
8523 #if MEASURE_MEM_ALLOC
8524 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8525 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8526 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8528 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8530 #endif // MEASURE_MEM_ALLOC
8532 void compFunctionTraceStart();
8533 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8536 size_t compMaxUncheckedOffsetForNullObject;
8538 void compInitOptions(JitFlags* compileFlags);
8540 void compSetProcessor();
8541 void compInitDebuggingInfo();
8542 void compSetOptimizationLevel();
8543 #ifdef _TARGET_ARMARCH_
8544 bool compRsvdRegCheck(FrameLayoutState curState);
8546 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8548 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8549 void ResetOptAnnotations();
8551 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8552 void RecomputeLoopInfo();
8554 #ifdef PROFILING_SUPPORTED
8555 // Data required for generating profiler Enter/Leave/TailCall hooks
8557 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8558 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8559 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8562 #ifdef _TARGET_AMD64_
8563 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8566 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8567 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8569 IAllocator* getAllocator()
8571 return compAsIAllocator;
8574 #if MEASURE_MEM_ALLOC
8575 IAllocator* getAllocatorBitset()
8577 return compAsIAllocatorBitset;
8579 IAllocator* getAllocatorGC()
8581 return compAsIAllocatorGC;
8583 IAllocator* getAllocatorLoopHoist()
8585 return compAsIAllocatorLoopHoist;
8587 #else // !MEASURE_MEM_ALLOC
8588 IAllocator* getAllocatorBitset()
8590 return compAsIAllocator;
8592 IAllocator* getAllocatorGC()
8594 return compAsIAllocator;
8596 IAllocator* getAllocatorLoopHoist()
8598 return compAsIAllocator;
8600 #endif // !MEASURE_MEM_ALLOC
8603 IAllocator* getAllocatorDebugOnly()
8605 #if MEASURE_MEM_ALLOC
8606 return compAsIAllocatorDebugOnly;
8607 #else // !MEASURE_MEM_ALLOC
8608 return compAsIAllocator;
8609 #endif // !MEASURE_MEM_ALLOC
8614 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8615 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8619 XX Checks for type compatibility and merges types XX
8621 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8622 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8626 // Set to TRUE if verification cannot be skipped for this method
8627 // If we detect unverifiable code, we will lazily check
8628 // canSkipMethodVerification() to see if verification is REALLY needed.
8629 BOOL tiVerificationNeeded;
8631 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8632 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8633 BOOL tiIsVerifiableCode;
8635 // Set to TRUE if runtime callout is needed for this method
8636 BOOL tiRuntimeCalloutNeeded;
8638 // Set to TRUE if security prolog/epilog callout is needed for this method
8639 // Note: This flag is different than compNeedSecurityCheck.
8640 // compNeedSecurityCheck means whether or not a security object needs
8641 // to be allocated on the stack, which is currently true for EnC as well.
8642 // tiSecurityCalloutNeeded means whether or not security callouts need
8643 // to be inserted in the jitted code.
8644 BOOL tiSecurityCalloutNeeded;
8646 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8647 // This support is necessary to suport attributes that are not described in
8648 // for example, signatures. For example, the permanent home byref (byref that
8649 // points to the gc heap), isn't a property of method signatures, therefore,
8650 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8651 // but when deciding if we need to reimport a block, we need to take these
8653 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8655 // Returns TRUE if child is equal to or a subtype of parent.
8656 // normalisedForStack indicates that both types are normalised for the stack
8657 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8659 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8660 // *pDest is modified to represent the merged type. Sets "*changed" to true
8661 // if this changes "*pDest".
8662 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8664 // Set pDest from the primitive value type.
8665 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8667 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8670 // <BUGNUM> VSW 471305
8671 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8672 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8673 // We use a "short" as we need to push/pop this scope.
8675 short compRegSetCheckLevel;
8679 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8680 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8682 XX IL verification stuff XX
8685 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8686 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8690 // The following is used to track liveness of local variables, initialization
8691 // of valueclass constructors, and type safe use of IL instructions.
8693 // dynamic state info needed for verification
8694 EntryState verCurrentState;
8696 // this ptr of object type .ctors are considered intited only after
8697 // the base class ctor is called, or an alternate ctor is called.
8698 // An uninited this ptr can be used to access fields, but cannot
8699 // be used to call a member function.
8700 BOOL verTrackObjCtorInitState;
8702 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8704 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8705 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8706 void verInitCurrentState();
8707 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8709 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8710 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8711 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8713 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8714 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8715 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8716 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8717 typeInfo verMakeTypeInfo(CorInfoType ciType,
8718 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8719 BOOL verIsSDArray(typeInfo ti);
8720 typeInfo verGetArrayElemType(typeInfo ti);
8722 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8723 BOOL verNeedsVerification();
8724 BOOL verIsByRefLike(const typeInfo& ti);
8725 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8727 // generic type variables range over types that satisfy IsBoxable
8728 BOOL verIsBoxable(const typeInfo& ti);
8730 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8731 DEBUGARG(unsigned line));
8732 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8733 DEBUGARG(unsigned line));
8734 bool verCheckTailCallConstraint(OPCODE opcode,
8735 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8736 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8737 // on a type parameter?
8738 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8739 // return false to the caller.
8740 // If false, it will throw.
8742 bool verIsBoxedValueType(typeInfo ti);
8744 void verVerifyCall(OPCODE opcode,
8745 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8746 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8748 bool readonlyCall, // is this a "readonly." call?
8749 const BYTE* delegateCreateStart,
8750 const BYTE* codeAddr,
8751 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8753 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8755 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8756 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8757 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8758 const CORINFO_FIELD_INFO& fieldInfo,
8759 const typeInfo* tiThis,
8761 BOOL allowPlainStructAsThis = FALSE);
8762 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8763 void verVerifyThisPtrInitialised();
8764 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8766 // Register allocator
8767 void raInitStackFP();
8768 void raEnregisterVarsPrePassStackFP();
8769 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8770 void raEnregisterVarsPostPassStackFP();
8771 void raGenerateFPRefCounts();
8772 void raEnregisterVarsStackFP();
8773 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8775 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8776 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8778 // returns true if enregistering v1 would save more mem accesses than v2
8779 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8782 void raDumpHeightsStackFP();
8783 void raDumpVariableRegIntfFloat();
8786 #if FEATURE_STACK_FP_X87
8788 // Currently, we use FP transition blocks in only 2 situations:
8790 // -conditional jump on longs where FP stack differs with target: it's not strictly
8791 // necessary, but its low frequency and the code would get complicated if we try to
8792 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8793 // minimize the way we generate the jump code.
8794 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8795 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8797 // However, transition blocks have 2 problems
8799 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8800 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8801 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8802 // in the right place without preordering them), this causes us to have to generate the transition
8803 // blocks in the cold area if we want procedure splitting.
8806 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8807 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8808 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8809 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8810 // a big change in the exception.
8812 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8813 // optimizations. For these 2 cases:
8815 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8816 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8817 // a switch statement.
8819 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8820 // current procedure splitting and exception code have.
8821 bool compMayHaveTransitionBlocks;
8823 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8825 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8827 unsigned raCntStkStackFP;
8828 unsigned raCntWtdStkDblStackFP;
8829 unsigned raCntStkParamDblStackFP;
8831 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8832 // TODO: Do we want to put this in LclVarDsc?
8833 unsigned raPayloadStackFP[lclMAX_TRACKED];
8834 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8836 // Useful for debugging
8837 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8839 #endif // FEATURE_STACK_FP_X87
8842 // One line log function. Default level is 0. Increasing it gives you
8843 // more log information
8845 // levels are currently unused: #define JITDUMP(level,...) ();
8846 void JitLogEE(unsigned level, const char* fmt, ...);
8848 bool compDebugBreak;
8850 bool compJitHaltMethod();
8855 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8856 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8858 XX GS Security checks for unsafe buffers XX
8860 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8861 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8864 struct ShadowParamVarInfo
8866 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8867 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8869 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8871 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8872 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8873 // slots and update all trees to refer to shadow slots is done immediately after
8874 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8875 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8876 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8877 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8878 // creating a shadow slot even though this routine returns true.
8880 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8881 // required. There are two cases under which a reg arg could potentially be used from its
8883 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8884 // b) LSRA spills it
8886 // Possible solution to address case (a)
8887 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8888 // in this routine. Note that live out of exception handler is something we may not be
8889 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8890 // Therefore, for methods with exception handling and need GS cookie check we might have
8891 // to take conservative approach.
8893 // Possible solution to address case (b)
8894 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8895 // create a new spill temp if the method needs GS cookie check.
8896 return varDsc->lvIsParam;
8897 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8898 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8905 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8910 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8911 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8912 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8914 void gsGSChecksInitCookie(); // Grabs cookie variable
8915 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8916 bool gsFindVulnerableParams(); // Shadow param analysis code
8917 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8919 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8920 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8922 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8923 // This can be overwritten by setting complus_JITInlineSize env variable.
8925 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8928 #ifdef FEATURE_JIT_METHOD_PERF
8929 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8930 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8932 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8933 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8935 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8937 #if MEASURE_CLRAPI_CALLS
8938 // Thin wrappers that call into JitTimer (if present).
8939 inline void CLRApiCallEnter(unsigned apix);
8940 inline void CLRApiCallLeave(unsigned apix);
8943 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8944 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8949 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8950 // These variables are associated with maintaining SQM data about compile time.
8951 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8952 // in the current compilation.
8953 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8954 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8955 // the inlining phase in the current compilation.
8956 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8958 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8959 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8960 // type-loading and class initialization).
8961 void RecordStateAtEndOfInlining();
8962 // Assumes being called at the end of compilation. Update the SQM state.
8963 void RecordStateAtEndOfCompilation();
8965 #ifdef FEATURE_CLRSQM
8966 // Does anything SQM related necessary at process shutdown time.
8967 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8968 #endif // FEATURE_CLRSQM
8971 #if FUNC_INFO_LOGGING
8972 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8973 // filename to write it to.
8974 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8975 #endif // FUNC_INFO_LOGGING
8977 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8979 // Is the compilation in a full trust context?
8980 bool compIsFullTrust();
8982 #ifndef FEATURE_TRACELOGGING
8983 // Should we actually fire the noway assert body and the exception handler?
8984 bool compShouldThrowOnNoway();
8985 #else // FEATURE_TRACELOGGING
8986 // Should we actually fire the noway assert body and the exception handler?
8987 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8989 // Telemetry instance to use per method compilation.
8990 JitTelemetry compJitTelemetry;
8992 // Get common parameters that have to be logged with most telemetry data.
8993 void compGetTelemetryDefaults(const char** assemblyName,
8994 const char** scopeName,
8995 const char** methodName,
8996 unsigned* methodHash);
8997 #endif // !FEATURE_TRACELOGGING
9001 NodeToTestDataMap* m_nodeTestData;
9003 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9004 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9005 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9006 // Current kept in this.
9008 NodeToTestDataMap* GetNodeTestData()
9010 Compiler* compRoot = impInlineRoot();
9011 if (compRoot->m_nodeTestData == nullptr)
9013 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9015 return compRoot->m_nodeTestData;
9018 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9020 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9021 // currently occur in the AST graph.
9022 NodeToIntMap* FindReachableNodesInNodeTestData();
9024 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9025 // test data, associate that data with "to".
9026 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9028 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9029 // have annotations, attach similar annotations to the corresponding nodes in "to".
9030 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9032 // These are the methods that test that the various conditions implied by the
9033 // test attributes are satisfied.
9034 void JitTestCheckSSA(); // SSA builder tests.
9035 void JitTestCheckVN(); // Value numbering tests.
9038 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9040 FieldSeqStore* m_fieldSeqStore;
9042 FieldSeqStore* GetFieldSeqStore()
9044 Compiler* compRoot = impInlineRoot();
9045 if (compRoot->m_fieldSeqStore == nullptr)
9047 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9048 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9049 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9051 return compRoot->m_fieldSeqStore;
9054 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9056 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9057 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9058 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9059 // attach the field sequence directly to the address node.
9060 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9062 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9064 // Don't need to worry about inlining here
9065 if (m_zeroOffsetFieldMap == nullptr)
9067 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9069 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9070 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9072 return m_zeroOffsetFieldMap;
9075 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9076 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9077 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9078 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9079 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9080 // record the the field sequence using the ZeroOffsetFieldMap described above.
9082 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9083 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9084 // CoreRT. Such case is handled same as the default case.
9085 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9087 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9089 NodeToArrayInfoMap* m_arrayInfoMap;
9091 NodeToArrayInfoMap* GetArrayInfoMap()
9093 Compiler* compRoot = impInlineRoot();
9094 if (compRoot->m_arrayInfoMap == nullptr)
9096 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9097 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9098 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9100 return compRoot->m_arrayInfoMap;
9103 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9105 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9106 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9107 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9108 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9110 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9112 // Use the same map for GCHeap and ByrefExposed when their states match.
9113 memoryKind = ByrefExposed;
9116 assert(memoryKind < MemoryKindCount);
9117 Compiler* compRoot = impInlineRoot();
9118 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9120 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9121 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9122 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9124 return compRoot->m_memorySsaMap[memoryKind];
9127 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9128 CORINFO_CLASS_HANDLE m_refAnyClass;
9129 CORINFO_FIELD_HANDLE GetRefanyDataField()
9131 if (m_refAnyClass == nullptr)
9133 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9135 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9137 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9139 if (m_refAnyClass == nullptr)
9141 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9143 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9147 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9149 #if ALLVARSET_COUNTOPS
9150 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9153 static HelperCallProperties s_helperCallProperties;
9155 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9156 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9157 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9159 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9162 unsigned __int8* offset0,
9163 unsigned __int8* offset1);
9164 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9165 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9167 void fgMorphMultiregStructArgs(GenTreeCall* call);
9168 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9170 }; // end of class Compiler
9172 // Inline methods of CompAllocator.
9173 void* CompAllocator::Alloc(size_t sz)
9175 #if MEASURE_MEM_ALLOC
9176 return m_comp->compGetMem(sz, m_cmk);
9178 return m_comp->compGetMem(sz);
9182 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9184 #if MEASURE_MEM_ALLOC
9185 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9187 return m_comp->compGetMemArray(elems, elemSize);
9191 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9192 inline LclVarDsc::LclVarDsc(Compiler* comp)
9193 : // Initialize the ArgRegs to REG_STK.
9194 // The morph will do the right thing to change
9195 // to the right register if passed in register.
9198 #if FEATURE_MULTIREG_ARGS
9199 _lvOtherArgReg(REG_STK)
9201 #endif // FEATURE_MULTIREG_ARGS
9203 lvRefBlks(BlockSetOps::UninitVal())
9205 #endif // ASSERTION_PROP
9206 lvPerSsaData(comp->getAllocator())
9211 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9212 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9214 XX Miscellaneous Compiler stuff XX
9216 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9217 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9220 // Values used to mark the types a stack slot is used for
9222 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9223 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9224 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9225 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9226 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9227 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9228 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9229 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9231 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9233 /*****************************************************************************
9235 * Variables to keep track of total code amounts.
9240 extern size_t grossVMsize;
9241 extern size_t grossNCsize;
9242 extern size_t totalNCsize;
9244 extern unsigned genMethodICnt;
9245 extern unsigned genMethodNCnt;
9246 extern size_t gcHeaderISize;
9247 extern size_t gcPtrMapISize;
9248 extern size_t gcHeaderNSize;
9249 extern size_t gcPtrMapNSize;
9251 #endif // DISPLAY_SIZES
9253 /*****************************************************************************
9255 * Variables to keep track of basic block counts (more data on 1 BB methods)
9258 #if COUNT_BASIC_BLOCKS
9259 extern Histogram bbCntTable;
9260 extern Histogram bbOneBBSizeTable;
9263 /*****************************************************************************
9265 * Used by optFindNaturalLoops to gather statistical information such as
9266 * - total number of natural loops
9267 * - number of loops with 1, 2, ... exit conditions
9268 * - number of loops that have an iterator (for like)
9269 * - number of loops that have a constant iterator
9274 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9275 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9276 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9277 extern unsigned totalLoopCount; // counts the total number of natural loops
9278 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9279 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9280 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9281 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9283 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9284 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9285 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9286 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9287 extern Histogram loopCountTable; // Histogram of loop counts
9288 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9290 #endif // COUNT_LOOPS
9292 /*****************************************************************************
9293 * variables to keep track of how many iterations we go in a dataflow pass
9298 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9299 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9301 #endif // DATAFLOW_ITER
9303 #if MEASURE_BLOCK_SIZE
9304 extern size_t genFlowNodeSize;
9305 extern size_t genFlowNodeCnt;
9306 #endif // MEASURE_BLOCK_SIZE
9308 #if MEASURE_NODE_SIZE
9309 struct NodeSizeStats
9314 genTreeNodeSize = 0;
9315 genTreeNodeActualSize = 0;
9318 size_t genTreeNodeCnt;
9319 size_t genTreeNodeSize; // The size we allocate
9320 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9321 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9322 // a smaller node to a larger one. TODO-Cleanup: add stats on
9323 // SetOper()/ChangeOper() usage to quanitfy this.
9325 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9326 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9327 extern Histogram genTreeNcntHist;
9328 extern Histogram genTreeNsizHist;
9329 #endif // MEASURE_NODE_SIZE
9331 /*****************************************************************************
9332 * Count fatal errors (including noway_asserts).
9336 extern unsigned fatal_badCode;
9337 extern unsigned fatal_noWay;
9338 extern unsigned fatal_NOMEM;
9339 extern unsigned fatal_noWayAssertBody;
9341 extern unsigned fatal_noWayAssertBodyArgs;
9343 extern unsigned fatal_NYI;
9344 #endif // MEASURE_FATAL
9346 /*****************************************************************************
9350 #ifdef _TARGET_XARCH_
9352 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9353 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9354 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9356 const instruction INS_AND = INS_and;
9357 const instruction INS_OR = INS_or;
9358 const instruction INS_XOR = INS_xor;
9359 const instruction INS_NEG = INS_neg;
9360 const instruction INS_TEST = INS_test;
9361 const instruction INS_MUL = INS_imul;
9362 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9363 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9364 const instruction INS_BREAKPOINT = INS_int3;
9365 const instruction INS_ADDC = INS_adc;
9366 const instruction INS_SUBC = INS_sbb;
9367 const instruction INS_NOT = INS_not;
9373 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9374 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9375 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9377 const instruction INS_AND = INS_and;
9378 const instruction INS_OR = INS_orr;
9379 const instruction INS_XOR = INS_eor;
9380 const instruction INS_NEG = INS_rsb;
9381 const instruction INS_TEST = INS_tst;
9382 const instruction INS_MUL = INS_mul;
9383 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9384 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9385 const instruction INS_BREAKPOINT = INS_bkpt;
9386 const instruction INS_ADDC = INS_adc;
9387 const instruction INS_SUBC = INS_sbc;
9388 const instruction INS_NOT = INS_mvn;
9392 #ifdef _TARGET_ARM64_
9394 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9395 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9396 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9398 const instruction INS_AND = INS_and;
9399 const instruction INS_OR = INS_orr;
9400 const instruction INS_XOR = INS_eor;
9401 const instruction INS_NEG = INS_neg;
9402 const instruction INS_TEST = INS_tst;
9403 const instruction INS_MUL = INS_mul;
9404 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9405 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9406 const instruction INS_BREAKPOINT = INS_bkpt;
9407 const instruction INS_ADDC = INS_adc;
9408 const instruction INS_SUBC = INS_sbc;
9409 const instruction INS_NOT = INS_mvn;
9413 /*****************************************************************************/
9415 extern const BYTE genTypeSizes[];
9416 extern const BYTE genTypeAlignments[];
9417 extern const BYTE genTypeStSzs[];
9418 extern const BYTE genActualTypes[];
9420 /*****************************************************************************/
9422 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9423 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9426 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9427 #elif defined(_TARGET_ARM64_)
9428 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9431 /*****************************************************************************/
9433 #define REG_CORRUPT regNumber(REG_NA + 1)
9434 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9435 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9437 /*****************************************************************************/
9439 extern BasicBlock dummyBB;
9441 /*****************************************************************************/
9442 /*****************************************************************************/
9444 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9445 // nodes of a statement in execution order.
9446 // __stmt: a GT_STMT type GenTree*
9447 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9449 #define foreach_treenode_execution_order(__node, __stmt) \
9450 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9452 // foreach_block: An iterator over all blocks in the function.
9453 // __compiler: the Compiler* object
9454 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9456 #define foreach_block(__compiler, __block) \
9457 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9459 /*****************************************************************************/
9460 /*****************************************************************************/
9464 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9466 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9467 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9469 XX Debugging helpers XX
9471 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9472 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9475 /*****************************************************************************/
9476 /* The following functions are intended to be called from the debugger, to dump
9477 * various data structures. The can be used in the debugger Watch or Quick Watch
9478 * windows. They are designed to be short to type and take as few arguments as
9479 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9480 * See the function definition comment for more details.
9483 void cBlock(Compiler* comp, BasicBlock* block);
9484 void cBlocks(Compiler* comp);
9485 void cBlocksV(Compiler* comp);
9486 void cTree(Compiler* comp, GenTree* tree);
9487 void cTrees(Compiler* comp);
9488 void cEH(Compiler* comp);
9489 void cVar(Compiler* comp, unsigned lclNum);
9490 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9491 void cVars(Compiler* comp);
9492 void cVarsFinal(Compiler* comp);
9493 void cBlockPreds(Compiler* comp, BasicBlock* block);
9494 void cReach(Compiler* comp);
9495 void cDoms(Compiler* comp);
9496 void cLiveness(Compiler* comp);
9497 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9499 void cFuncIR(Compiler* comp);
9500 void cBlockIR(Compiler* comp, BasicBlock* block);
9501 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9502 void cTreeIR(Compiler* comp, GenTree* tree);
9503 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9504 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9505 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9506 int cOperandIR(Compiler* comp, GenTree* operand);
9507 int cLeafIR(Compiler* comp, GenTree* tree);
9508 int cIndirIR(Compiler* comp, GenTree* tree);
9509 int cListIR(Compiler* comp, GenTree* list);
9510 int cSsaNumIR(Compiler* comp, GenTree* tree);
9511 int cValNumIR(Compiler* comp, GenTree* tree);
9512 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9514 void dBlock(BasicBlock* block);
9517 void dTree(GenTree* tree);
9520 void dVar(unsigned lclNum);
9521 void dVarDsc(LclVarDsc* varDsc);
9524 void dBlockPreds(BasicBlock* block);
9528 void dCVarSet(VARSET_VALARG_TP vars);
9530 void dVarSet(VARSET_VALARG_TP vars);
9531 void dRegMask(regMaskTP mask);
9534 void dBlockIR(BasicBlock* block);
9535 void dTreeIR(GenTree* tree);
9536 void dLoopIR(Compiler::LoopDsc* loop);
9537 void dLoopNumIR(unsigned loopNum);
9538 int dTabStopIR(int curr, int tabstop);
9539 int dTreeTypeIR(GenTree* tree);
9540 int dTreeKindsIR(GenTree* tree);
9541 int dTreeFlagsIR(GenTree* tree);
9542 int dOperandIR(GenTree* operand);
9543 int dLeafIR(GenTree* tree);
9544 int dIndirIR(GenTree* tree);
9545 int dListIR(GenTree* list);
9546 int dSsaNumIR(GenTree* tree);
9547 int dValNumIR(GenTree* tree);
9548 int dDependsIR(GenTree* comma);
9551 GenTree* dFindTree(GenTree* tree, unsigned id);
9552 GenTree* dFindTree(unsigned id);
9553 GenTreeStmt* dFindStmt(unsigned id);
9554 BasicBlock* dFindBlock(unsigned bbNum);
9558 #include "compiler.hpp" // All the shared inline functions
9560 /*****************************************************************************/
9561 #endif //_COMPILER_H_
9562 /*****************************************************************************/