1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
104 // Or the more-general IAllocator interface.
105 void* __cdecl operator new(size_t n, IAllocator* alloc);
106 void* __cdecl operator new[](size_t n, IAllocator* alloc);
108 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
109 // caused these to be ambiguous with the global placement new operators.
110 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
111 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
114 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
115 #include "loopcloning.h"
117 /*****************************************************************************/
119 /* This is included here and not earlier as it needs the definition of "CSE"
120 * which is defined in the section above */
122 /*****************************************************************************/
124 unsigned genLog2(unsigned value);
125 unsigned genLog2(unsigned __int64 value);
127 var_types genActualType(var_types type);
128 var_types genUnsignedType(var_types type);
129 var_types genSignedType(var_types type);
131 unsigned ReinterpretHexAsDecimal(unsigned);
133 /*****************************************************************************/
136 #ifdef FEATURE_AVX_SUPPORT
137 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
138 #else // !FEATURE_AVX_SUPPORT
139 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
140 #endif // !FEATURE_AVX_SUPPORT
141 #else // !FEATURE_SIMD
142 const unsigned TEMP_MAX_SIZE = sizeof(double);
143 #endif // !FEATURE_SIMD
144 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
146 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
149 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
152 // The following holds the Local var info (scope information)
153 typedef const char* VarName; // Actual ASCII string
156 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
157 IL_OFFSET vsdLifeEnd; // instr offset of end of life
158 unsigned vsdVarNum; // (remapped) LclVarDsc number
161 VarName vsdName; // name of the var
164 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
165 // Also, it is the index of this entry in the info.compVarScopes array,
166 // which is useful since the array is also accessed via the
167 // compEnterScopeList and compExitScopeList sorted arrays.
170 /*****************************************************************************
172 * The following holds the local variable counts and the descriptor table.
175 // This is the location of a definition.
181 DefLoc() : m_blk(nullptr), m_tree(nullptr)
186 // This class encapsulates all info about a local variable that may vary for different SSA names
191 ValueNumPair m_vnPair;
199 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
204 // The constructor. Most things can just be zero'ed.
205 LclVarDsc(Compiler* comp);
207 // note this only packs because var_types is a typedef of unsigned char
208 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
210 unsigned char lvIsParam : 1; // is this a parameter?
211 unsigned char lvIsRegArg : 1; // is this a register argument?
212 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
214 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
215 // is to help determine whether to use block init in the prolog.
216 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
217 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
218 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
219 // variable is in the same register for the entire function.
220 unsigned char lvTracked : 1; // is this a tracked variable?
221 bool lvTrackedNonStruct()
223 return lvTracked && lvType != TYP_STRUCT;
225 unsigned char lvPinned : 1; // is this a pinned variable?
227 unsigned char lvMustInit : 1; // must be initialized
228 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
229 // global location, etc.
230 // We cannot reason reliably about the value of the variable.
231 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
232 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
236 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
238 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
239 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
240 // read/write its value.
241 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
242 // the variable to be
243 // in the stack (at least at those boundaries.)
244 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
245 // particular byte from an int).
246 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
247 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
249 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
250 unsigned char lvRefAssign : 1; // involved in pointer assignment
251 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
252 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
255 unsigned char lvArgWrite : 1; // variable is a parameter and STARG was used on it
256 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp
258 unsigned char lvIsBoolean : 1; // set if variable is boolean
260 unsigned char lvRngOptDone : 1; // considered for range check opt?
261 unsigned char lvLoopInc : 1; // incremented in the loop?
262 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
263 unsigned char lvArrIndx : 1; // used as an array index?
264 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
265 unsigned char lvArrIndxDom : 1; // index dominates loop exit
267 unsigned char lvSingleDef : 1; // variable has a single def
268 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
269 unsigned char lvVolatileHint : 1; // hint for AssertionProp
272 unsigned char lvSpilled : 1; // enregistered variable was spilled
273 #ifndef _TARGET_64BIT_
274 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
275 #endif // !_TARGET_64BIT_
276 #ifdef _TARGET_64BIT_
277 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
280 unsigned char lvKeepType : 1; // Don't change the type of this variable
281 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
283 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
285 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
286 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
288 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
289 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
290 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
291 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
292 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
294 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
295 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
298 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
299 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
300 // with (lvIsRegArg && lvIsHfa())
301 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
302 #endif // FEATURE_HFA
305 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
306 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
307 // Consider cleaning this up so this workaround is not required.
308 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
309 // I.e. there is no longer any reference to the struct directly.
310 // In this case we can simply remove this struct local.
312 #ifndef LEGACY_BACKEND
313 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
314 #endif // !LEGACY_BACKEND
317 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
318 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
319 unsigned char lvSIMDType : 1; // This is a SIMD struct
320 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
321 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
322 #endif // FEATURE_SIMD
323 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
326 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
328 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
329 // Valid on promoted struct local fields.
332 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
333 unsigned char lvFldOffset;
334 unsigned char lvFldOrdinal;
336 #if FEATURE_MULTIREG_ARGS
337 regNumber lvRegNumForSlot(unsigned slotNum)
343 else if (slotNum == 1)
345 return lvOtherArgReg;
349 assert(false && "Invalid slotNum!");
354 #endif // FEATURE_MULTIREG_ARGS
372 bool lvIsHfaRegArg() const
375 return _lvIsHfaRegArg;
381 void lvSetIsHfaRegArg()
384 _lvIsHfaRegArg = true;
388 bool lvHfaTypeIsFloat() const
391 return _lvHfaTypeIsFloat;
397 void lvSetHfaTypeIsFloat(bool value)
400 _lvHfaTypeIsFloat = value;
404 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
405 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
407 unsigned lvHfaSlots() const
410 assert(lvType == TYP_STRUCT);
412 return lvExactSize / sizeof(float);
413 #else // _TARGET_ARM64_
414 if (lvHfaTypeIsFloat())
416 return lvExactSize / sizeof(float);
420 return lvExactSize / sizeof(double);
422 #endif // _TARGET_ARM64_
425 // lvIsMultiRegArgOrRet()
426 // returns true if this is a multireg LclVar struct used in an argument context
427 // or if this is a multireg LclVar struct assigned from a multireg call
428 bool lvIsMultiRegArgOrRet()
430 return lvIsMultiRegArg || lvIsMultiRegRet;
434 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
435 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
436 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
437 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
438 // to non-zero if the variable gets the same register assignment for its entire
440 #if !defined(_TARGET_64BIT_)
441 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
442 #endif // !defined(_TARGET_64BIT_)
444 regNumberSmall _lvArgReg; // The register in which this argument is passed.
446 #if FEATURE_MULTIREG_ARGS
447 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
448 // Note this is defined but not used by ARM32
449 #endif // FEATURE_MULTIREG_ARGS
451 #ifndef LEGACY_BACKEND
453 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
454 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
456 #endif // !LEGACY_BACKEND
459 // The register number is stored in a small format (8 bits), but the getters return and the setters take
460 // a full-size (unsigned) format, to localize the casts here.
462 /////////////////////
464 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
466 regNumber GetRegNum() const
468 return (regNumber)_lvRegNum;
471 void SetRegNum(regNumber reg)
473 _lvRegNum = (regNumberSmall)reg;
474 assert(_lvRegNum == reg);
477 /////////////////////
479 #if defined(_TARGET_64BIT_)
480 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
482 regNumber GetOtherReg() const
484 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
485 // "unreachable code" warnings
489 void SetOtherReg(regNumber reg)
491 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
492 // "unreachable code" warnings
494 #else // !_TARGET_64BIT_
495 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
497 regNumber GetOtherReg() const
499 return (regNumber)_lvOtherReg;
502 void SetOtherReg(regNumber reg)
504 _lvOtherReg = (regNumberSmall)reg;
505 assert(_lvOtherReg == reg);
507 #endif // !_TARGET_64BIT_
509 /////////////////////
511 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
513 regNumber GetArgReg() const
515 return (regNumber)_lvArgReg;
518 void SetArgReg(regNumber reg)
520 _lvArgReg = (regNumberSmall)reg;
521 assert(_lvArgReg == reg);
524 #if FEATURE_MULTIREG_ARGS
525 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
527 regNumber GetOtherArgReg() const
529 return (regNumber)_lvOtherArgReg;
532 void SetOtherArgReg(regNumber reg)
534 _lvOtherArgReg = (regNumberSmall)reg;
535 assert(_lvOtherArgReg == reg);
537 #endif // FEATURE_MULTIREG_ARGS
540 // Is this is a SIMD struct?
541 bool lvIsSIMDType() const
546 // Is this is a SIMD struct which is used for SIMD intrinsic?
547 bool lvIsUsedInSIMDIntrinsic() const
549 return lvUsedInSIMDIntrinsic;
552 // If feature_simd not enabled, return false
553 bool lvIsSIMDType() const
557 bool lvIsUsedInSIMDIntrinsic() const
563 /////////////////////
565 #ifndef LEGACY_BACKEND
566 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
568 regNumber GetArgInitReg() const
570 return (regNumber)_lvArgInitReg;
573 void SetArgInitReg(regNumber reg)
575 _lvArgInitReg = (regNumberSmall)reg;
576 assert(_lvArgInitReg == reg);
579 /////////////////////
581 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
583 regPairNo GetArgInitRegPair() const
585 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
586 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
590 void SetArgInitRegPair(regPairNo regPair)
592 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
593 _lvArgInitRegPair = (regPairNoSmall)regPair;
594 assert(_lvArgInitRegPair == regPair);
597 /////////////////////
599 bool lvIsRegCandidate() const
601 return lvLRACandidate != 0;
604 bool lvIsInReg() const
606 return lvIsRegCandidate() && (lvRegNum != REG_STK);
609 #else // LEGACY_BACKEND
611 bool lvIsRegCandidate() const
613 return lvTracked != 0;
616 bool lvIsInReg() const
618 return lvRegister != 0;
621 #endif // LEGACY_BACKEND
623 regMaskTP lvRegMask() const
625 regMaskTP regMask = RBM_NONE;
626 if (varTypeIsFloating(TypeGet()))
628 if (lvRegNum != REG_STK)
630 regMask = genRegMaskFloat(lvRegNum, TypeGet());
635 if (lvRegNum != REG_STK)
637 regMask = genRegMask(lvRegNum);
640 // For longs we may have two regs
641 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
643 regMask |= genRegMask(lvOtherReg);
649 regMaskSmall lvPrefReg; // set of regs it prefers to live in
651 unsigned short lvVarIndex; // variable tracking index
652 unsigned short lvRefCnt; // unweighted (real) reference count
653 unsigned lvRefCntWtd; // weighted reference count
654 int lvStkOffs; // stack offset of home
655 unsigned lvExactSize; // (exact) size of the type in bytes
657 // Is this a promoted struct?
658 // This method returns true only for structs (including SIMD structs), not for
659 // locals that are split on a 32-bit target.
660 // It is only necessary to use this:
661 // 1) if only structs are wanted, and
662 // 2) if Lowering has already been done.
663 // Otherwise lvPromoted is valid.
664 bool lvPromotedStruct()
666 #if !defined(_TARGET_64BIT_)
667 return (lvPromoted && !varTypeIsLong(lvType));
668 #else // defined(_TARGET_64BIT_)
670 #endif // defined(_TARGET_64BIT_)
673 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
675 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
676 // where the struct itself is no longer used because all access is via its member fields.
677 // When that happens, the struct is marked as unused and its type has been changed to
678 // TYP_INT (to keep the GC tracking code from looking at it).
679 // See Compiler::raAssignVars() for details. For example:
680 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
681 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
682 // float V03.f1 (offs=0x00) -> V12 tmp7
683 // f8 (last use) (last use) $345
684 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
685 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
687 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
689 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
690 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
691 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
692 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
693 if ((lvType == TYP_SIMD12) && !lvIsParam)
695 assert(lvExactSize == 12);
698 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
700 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
703 unsigned lvSlotNum; // original slot # (if remapped)
705 typeInfo lvVerTypeInfo; // type info needed for verification
707 BYTE* lvGcLayout; // GC layout info for structs
710 BlockSet lvRefBlks; // Set of blocks that contain refs
711 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
712 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
714 var_types TypeGet() const
716 return (var_types)lvType;
718 bool lvStackAligned() const
720 assert(lvIsStructField);
721 return ((lvFldOffset % sizeof(void*)) == 0);
723 bool lvNormalizeOnLoad() const
725 return varTypeIsSmall(TypeGet()) &&
726 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
727 (lvIsParam || lvAddrExposed || lvIsStructField);
730 bool lvNormalizeOnStore()
732 return varTypeIsSmall(TypeGet()) &&
733 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
734 !(lvIsParam || lvAddrExposed || lvIsStructField);
737 void lvaResetSortAgainFlag(Compiler* pComp);
738 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
739 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
740 void setPrefReg(regNumber regNum, Compiler* pComp);
741 void addPrefReg(regMaskTP regMask, Compiler* pComp);
742 bool IsFloatRegType() const
744 return isFloatRegType(lvType) || lvIsHfaRegArg();
746 var_types GetHfaType() const
748 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
750 void SetHfaType(var_types type)
752 assert(varTypeIsFloating(type));
753 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
756 #ifndef LEGACY_BACKEND
757 var_types lvaArgType();
760 PerSsaArray lvPerSsaData;
763 // Keep track of the # of SsaNames, for a bounds check.
764 unsigned lvNumSsaNames;
767 // Returns the address of the per-Ssa data for the given ssaNum (which is required
768 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
769 // not an SSA variable).
770 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
772 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
773 assert(SsaConfig::RESERVED_SSA_NUM == 0);
774 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
775 assert(zeroBased < lvNumSsaNames);
776 return &lvPerSsaData.GetRef(zeroBased);
781 void PrintVarReg() const
783 if (isRegPairType(TypeGet()))
785 printf("%s:%s", getRegName(lvOtherReg), // hi32
786 getRegName(lvRegNum)); // lo32
790 printf("%s", getRegName(lvRegNum));
795 }; // class LclVarDsc
798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
799 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
803 XX The temporary lclVars allocated by the compiler for code generation XX
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
809 /*****************************************************************************
811 * The following keeps track of temporaries allocated in the stack frame
812 * during code-generation (after register allocation). These spill-temps are
813 * only used if we run out of registers while evaluating a tree.
815 * These are different from the more common temps allocated by lvaGrabTemp().
826 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
834 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
838 0); // temps must have a negative number (so they have a different number from all local variables)
839 tdOffs = BAD_TEMP_OFFSET;
843 IMPL_LIMITATION("too many spill temps");
848 bool tdLegalOffset() const
850 return tdOffs != BAD_TEMP_OFFSET;
854 int tdTempOffs() const
856 assert(tdLegalOffset());
859 void tdSetTempOffs(int offs)
862 assert(tdLegalOffset());
864 void tdAdjustTempOffs(int offs)
867 assert(tdLegalOffset());
870 int tdTempNum() const
875 unsigned tdTempSize() const
879 var_types tdTempType() const
885 // interface to hide linearscan implementation from rest of compiler
886 class LinearScanInterface
889 virtual void doLinearScan() = 0;
890 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
893 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
895 // Information about arrays: their element type and size, and the offset of the first element.
896 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
897 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
898 // for example, in value numbering of array index expressions.
901 var_types m_elemType;
902 CORINFO_CLASS_HANDLE m_elemStructType;
904 unsigned m_elemOffset;
906 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
910 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
911 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
916 // This enumeration names the phases into which we divide compilation. The phases should completely
917 // partition a compilation.
920 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
921 #include "compphases.h"
925 extern const char* PhaseNames[];
926 extern const char* PhaseEnums[];
927 extern const LPCWSTR PhaseShortNames[];
929 // The following enum provides a simple 1:1 mapping to CLR API's
930 enum API_ICorJitInfo_Names
932 #define DEF_CLR_API(name) API_##name,
933 #include "ICorJitInfo_API_names.h"
937 //---------------------------------------------------------------
941 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
942 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
943 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
944 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
945 // by "m_timerFailure" being true.
946 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
949 #ifdef FEATURE_JIT_METHOD_PERF
950 // The string names of the phases.
951 static const char* PhaseNames[];
953 static bool PhaseHasChildren[];
954 static int PhaseParent[];
956 unsigned m_byteCodeBytes;
957 unsigned __int64 m_totalCycles;
958 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
959 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
960 #if MEASURE_CLRAPI_CALLS
961 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
962 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
964 // For better documentation, we call EndPhase on
965 // non-leaf phases. We should also call EndPhase on the
966 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
967 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
968 // We add all such "redundant end phase" intervals to this variable below; we print
969 // it out in a report, so we can verify that it is, indeed, very small. If it ever
970 // isn't, this means that we're doing something significant between the end of the last
971 // declared subphase and the end of its parent.
972 unsigned __int64 m_parentPhaseEndSlop;
975 #if MEASURE_CLRAPI_CALLS
976 // The following measures the time spent inside each individual CLR API call.
977 unsigned m_allClrAPIcalls;
978 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
979 unsigned __int64 m_allClrAPIcycles;
980 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
981 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
982 #endif // MEASURE_CLRAPI_CALLS
984 CompTimeInfo(unsigned byteCodeBytes);
988 #ifdef FEATURE_JIT_METHOD_PERF
990 #if MEASURE_CLRAPI_CALLS
991 struct WrapICorJitInfo;
994 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
995 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
996 // The operation of adding a single method's timing to the summary may be performed concurrently by several
997 // threads, so it is protected by a lock.
998 // This class is intended to be used as a singleton type, with only a single instance.
999 class CompTimeSummaryInfo
1001 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1002 static CritSecObject s_compTimeSummaryLock;
1006 CompTimeInfo m_total;
1007 CompTimeInfo m_maximum;
1009 int m_numFilteredMethods;
1010 CompTimeInfo m_filtered;
1012 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1013 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1014 // If any OS interaction fails, returns 0.0.
1015 double CyclesPerSecond();
1017 // This can use what ever data you want to determine if the value to be added
1018 // belongs in the filtered section (it's always included in the unfiltered section)
1019 bool IncludedInFilteredData(CompTimeInfo& info);
1022 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1023 static CompTimeSummaryInfo s_compTimeSummary;
1025 CompTimeSummaryInfo()
1026 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1030 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1031 // This is thread safe.
1032 void AddInfo(CompTimeInfo& info, bool includePhases);
1034 // Print the summary information to "f".
1035 // This is not thread-safe; assumed to be called by only one thread.
1036 void Print(FILE* f);
1039 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1040 // and when the current phase started. This is intended to be part of a Compilation object. This is
1041 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1045 unsigned __int64 m_start; // Start of the compilation.
1046 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1047 #if MEASURE_CLRAPI_CALLS
1048 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1049 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1050 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1051 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1052 static double s_cyclesPerSec; // Cached for speedier measurements
1055 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1057 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1059 static CritSecObject s_csvLock; // Lock to protect the time log file.
1060 void PrintCsvMethodStats(Compiler* comp);
1063 void* operator new(size_t);
1064 void* operator new[](size_t);
1065 void operator delete(void*);
1066 void operator delete[](void*);
1069 // Initialized the timer instance
1070 JitTimer(unsigned byteCodeSize);
1072 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1074 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1077 static void PrintCsvHeader();
1079 // Ends the current phase (argument is for a redundant check).
1080 void EndPhase(Phases phase);
1082 #if MEASURE_CLRAPI_CALLS
1083 // Start and end a timed CLR API call.
1084 void CLRApiCallEnter(unsigned apix);
1085 void CLRApiCallLeave(unsigned apix);
1086 #endif // MEASURE_CLRAPI_CALLS
1088 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1089 // and adds it to "sum".
1090 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1092 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1093 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1094 // "m_info" to true.
1095 bool GetThreadCycles(unsigned __int64* cycles)
1097 bool res = CycleTimer::GetThreadCyclesS(cycles);
1100 m_info.m_timerFailure = true;
1105 #endif // FEATURE_JIT_METHOD_PERF
1107 //------------------- Function/Funclet info -------------------------------
1108 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1110 FUNC_ROOT, // The main/root function (always id==0)
1111 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1112 FUNC_FILTER, // a funclet associated with an EH filter
1115 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1122 BYTE funFlags; // Currently unused, just here for padding
1123 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1124 // funclet. It is only valid if funKind field indicates this is a
1125 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1127 #if defined(_TARGET_AMD64_)
1129 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1130 emitLocation* startLoc;
1131 emitLocation* endLoc;
1132 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1133 emitLocation* coldEndLoc;
1134 UNWIND_INFO unwindHeader;
1135 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1136 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1137 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1138 unsigned unwindCodeSlot;
1140 #ifdef UNIX_AMD64_ABI
1141 jitstd::vector<CFI_CODE>* cfiCodes;
1142 #endif // UNIX_AMD64_ABI
1144 #elif defined(_TARGET_ARMARCH_)
1146 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1147 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1148 // Note: we only have a pointer here instead of the actual object,
1149 // to save memory in the JIT case (compared to the NGEN case),
1150 // where we don't have any cold section.
1151 // Note 2: we currently don't support hot/cold splitting in functions
1152 // with EH, so uwiCold will be NULL for all funclets.
1154 #endif // _TARGET_ARMARCH_
1156 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1157 // that isn't shared between the main function body and funclets.
1160 struct fgArgTabEntry
1163 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1166 otherRegNum = REG_NA;
1167 isStruct = false; // is this a struct arg
1169 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1171 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1173 // it will point at the actual argument in the gtCallLateArgs list.
1174 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1176 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1178 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1180 unsigned numRegs; // Count of number of registers that this argument uses
1182 // A slot is a pointer sized region in the OutArg area.
1183 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1184 unsigned numSlots; // Count of number of slots that this argument uses
1186 unsigned alignment; // 1 or 2 (slots/registers)
1187 unsigned lateArgInx; // index into gtCallLateArgs list
1188 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1189 #if defined(UNIX_X86_ABI)
1190 unsigned padStkAlign; // Count of number of padding slots for stack alignment. For each Call, only the first
1191 // argument may have a value to emit "sub esp, n" to adjust the stack before pushing
1195 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1196 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1197 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1198 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1199 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1200 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1201 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1202 // previous arguments.
1203 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1204 // to be on the stack despite its arg list position.
1206 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1207 bool isStruct : 1; // True if this is a struct arg
1209 regNumber otherRegNum; // The (second) register to use when passing this argument.
1211 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1212 #elif defined(_TARGET_X86_)
1213 __declspec(property(get = getIsStruct)) bool isStruct;
1216 return varTypeIsStruct(node);
1218 #endif // _TARGET_X86_
1221 void SetIsHfaRegArg(bool hfaRegArg)
1223 isHfaRegArg = hfaRegArg;
1226 void SetIsBackFilled(bool backFilled)
1228 isBackFilled = backFilled;
1231 bool IsBackFilled() const
1233 return isBackFilled;
1235 #else // !_TARGET_ARM_
1236 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1238 void SetIsHfaRegArg(bool hfaRegArg)
1242 void SetIsBackFilled(bool backFilled)
1246 bool IsBackFilled() const
1250 #endif // !_TARGET_ARM_
1256 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1258 //-------------------------------------------------------------------------
1260 // The class fgArgInfo is used to handle the arguments
1261 // when morphing a GT_CALL node.
1266 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1267 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1268 unsigned argCount; // Updatable arg count value
1269 unsigned nextSlotNum; // Updatable slot count value
1270 unsigned stkLevel; // Stack depth when we make this call (for x86)
1271 #if defined(UNIX_X86_ABI)
1272 unsigned padStkAlign; // Count of number of padding slots for stack alignment. This value is used to turn back
1273 // stack pointer before it was adjusted after each Call
1276 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1277 bool hasRegArgs; // true if we have one or more register arguments
1278 bool hasStackArgs; // true if we have one or more stack arguments
1279 bool argsComplete; // marker for state
1280 bool argsSorted; // marker for state
1281 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1284 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1287 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1288 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1290 fgArgTabEntryPtr AddRegArg(
1291 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1293 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1294 fgArgTabEntryPtr AddRegArg(
1301 const bool isStruct,
1302 const regNumber otherRegNum = REG_NA,
1303 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1304 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1306 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1310 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1312 void RemorphReset();
1313 fgArgTabEntryPtr RemorphRegArg(
1314 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1316 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1318 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1320 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1322 void ArgsComplete();
1324 #if defined(UNIX_X86_ABI)
1325 void ArgsAlignPadding();
1330 void EvalArgsToTemps();
1332 void RecordStkLevel(unsigned stkLvl);
1333 unsigned RetrieveStkLevel();
1339 fgArgTabEntryPtr* ArgTable()
1343 unsigned GetNextSlotNum()
1347 #if defined(UNIX_X86_ABI)
1348 unsigned GetPadStackAlign()
1359 return hasStackArgs;
1361 bool AreArgsComplete() const
1363 return argsComplete;
1368 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1369 // We have the ability to mark source expressions with "Test Labels."
1370 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1371 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1373 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1376 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1377 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1378 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1379 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1380 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1383 struct TestLabelAndNum
1388 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1393 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1395 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1398 // This class implements the "IAllocator" interface, so that we can use
1399 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1401 class CompAllocator : public IAllocator
1404 #if MEASURE_MEM_ALLOC
1408 CompAllocator(Compiler* comp, CompMemKind cmk)
1410 #if MEASURE_MEM_ALLOC
1416 inline void* Alloc(size_t sz);
1418 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1420 // For the compiler's no-release allocator, free operations are no-ops.
1427 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1428 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1430 XX The big guy. The sections are currently organized as : XX
1432 XX o GenTree and BasicBlock XX
1444 XX o PrologScopeInfo XX
1445 XX o CodeGenerator XX
1450 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1451 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1456 friend class emitter;
1457 friend class UnwindInfo;
1458 friend class UnwindFragmentInfo;
1459 friend class UnwindEpilogInfo;
1460 friend class JitTimer;
1461 friend class LinearScan;
1462 friend class fgArgInfo;
1463 friend class Rationalizer;
1465 friend class Lowering;
1466 friend class CSE_DataFlow;
1467 friend class CSE_Heuristic;
1468 friend class CodeGenInterface;
1469 friend class CodeGen;
1470 friend class LclVarDsc;
1471 friend class TempDsc;
1473 friend class ObjectAllocator;
1475 #ifndef _TARGET_64BIT_
1476 friend class DecomposeLongs;
1477 #endif // !_TARGET_64BIT_
1480 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1481 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1483 XX Misc structs definitions XX
1485 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1486 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1490 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1509 bool dumpIRDataflow;
1510 bool dumpIRBlockHeaders;
1512 LPCWSTR dumpIRPhase;
1513 LPCWSTR dumpIRFormat;
1515 bool shouldUseVerboseTrees();
1516 bool asciiTrees; // If true, dump trees using only ASCII characters
1517 bool shouldDumpASCIITrees();
1518 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1519 bool shouldUseVerboseSsa();
1520 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1521 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1523 const char* VarNameToStr(VarName name)
1528 DWORD expensiveDebugCheckLevel;
1531 #if FEATURE_MULTIREG_RET
1532 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1533 #endif // FEATURE_MULTIREG_RET
1536 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1537 #endif // ARM_SOFTFP
1539 //-------------------------------------------------------------------------
1540 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1541 // HFAs are one to four element structs where each element is the same
1542 // type, either all float or all double. They are treated specially
1543 // in the ARM Procedure Call Standard, specifically, they are passed in
1544 // floating-point registers instead of the general purpose registers.
1547 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1548 bool IsHfa(GenTreePtr tree);
1550 var_types GetHfaType(GenTreePtr tree);
1551 unsigned GetHfaCount(GenTreePtr tree);
1553 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1554 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1556 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1557 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1559 //-------------------------------------------------------------------------
1560 // The following is used for validating format of EH table
1564 typedef struct EHNodeDsc* pEHNodeDsc;
1566 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1567 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1580 EHBlockType ehnBlockType; // kind of EH block
1581 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1582 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1583 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1585 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1586 pEHNodeDsc ehnChild; // leftmost nested block
1588 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1589 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1591 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1592 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1594 inline void ehnSetTryNodeType()
1596 ehnBlockType = TryNode;
1598 inline void ehnSetFilterNodeType()
1600 ehnBlockType = FilterNode;
1602 inline void ehnSetHandlerNodeType()
1604 ehnBlockType = HandlerNode;
1606 inline void ehnSetFinallyNodeType()
1608 ehnBlockType = FinallyNode;
1610 inline void ehnSetFaultNodeType()
1612 ehnBlockType = FaultNode;
1615 inline BOOL ehnIsTryBlock()
1617 return ehnBlockType == TryNode;
1619 inline BOOL ehnIsFilterBlock()
1621 return ehnBlockType == FilterNode;
1623 inline BOOL ehnIsHandlerBlock()
1625 return ehnBlockType == HandlerNode;
1627 inline BOOL ehnIsFinallyBlock()
1629 return ehnBlockType == FinallyNode;
1631 inline BOOL ehnIsFaultBlock()
1633 return ehnBlockType == FaultNode;
1636 // returns true if there is any overlap between the two nodes
1637 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1639 if (node1->ehnStartOffset < node2->ehnStartOffset)
1641 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1645 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1649 // fails with BADCODE if inner is not completely nested inside outer
1650 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1652 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1656 //-------------------------------------------------------------------------
1657 // Exception handling functions
1660 #if !FEATURE_EH_FUNCLETS
1662 bool ehNeedsShadowSPslots()
1664 return (info.compXcptnsCount || opts.compDbgEnC);
1667 // 0 for methods with no EH
1668 // 1 for methods with non-nested EH, or where only the try blocks are nested
1669 // 2 for a method with a catch within a catch
1671 unsigned ehMaxHndNestingCount;
1673 #endif // !FEATURE_EH_FUNCLETS
1675 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1676 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1678 bool bbInCatchHandlerILRange(BasicBlock* blk);
1679 bool bbInFilterILRange(BasicBlock* blk);
1680 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1681 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1682 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1683 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1684 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1686 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1687 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1689 // Returns true if "block" is the start of a try region.
1690 bool bbIsTryBeg(BasicBlock* block);
1692 // Returns true if "block" is the start of a handler or filter region.
1693 bool bbIsHandlerBeg(BasicBlock* block);
1695 // Returns true iff "block" is where control flows if an exception is raised in the
1696 // try region, and sets "*regionIndex" to the index of the try for the handler.
1697 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1698 // block of the filter, but not for the filter's handler.
1699 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1701 bool ehHasCallableHandlers();
1703 // Return the EH descriptor for the given region index.
1704 EHblkDsc* ehGetDsc(unsigned regionIndex);
1706 // Return the EH index given a region descriptor.
1707 unsigned ehGetIndex(EHblkDsc* ehDsc);
1709 // Return the EH descriptor index of the enclosing try, for the given region index.
1710 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1712 // Return the EH descriptor index of the enclosing handler, for the given region index.
1713 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1715 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1716 // block is not in a 'try' region).
1717 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1719 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1720 // if this block is not in a filter or handler region).
1721 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1723 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1724 // nullptr if this block's exceptions propagate to caller).
1725 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1727 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1728 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1729 bool ehIsBlockEHLast(BasicBlock* block);
1731 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1733 // Return the region index of the most nested EH region this block is in.
1734 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1736 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1737 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1739 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1740 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1741 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1742 // (It can never be a filter.)
1743 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1745 // A block has been deleted. Update the EH table appropriately.
1746 void ehUpdateForDeletedBlock(BasicBlock* block);
1748 // Determine whether a block can be deleted while preserving the EH normalization rules.
1749 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1751 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1752 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1754 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1755 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1756 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1757 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1758 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1759 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1760 // lives in a filter.)
1761 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1763 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1764 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1765 // (nullptr if the last block is the last block in the program).
1766 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1767 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1770 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1771 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1772 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1775 #if FEATURE_EH_FUNCLETS
1776 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1777 // if there is a filter that protects a region with a nested EH clause (such as a
1778 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1779 // genFuncletProlog() for more details. However, the VM seems to use it for more
1780 // purposes, maybe including debugging. Until we are sure otherwise, always create
1781 // a PSPSym for functions with any EH.
1782 bool ehNeedsPSPSym() const
1786 #else // _TARGET_X86_
1787 return compHndBBtabCount > 0;
1788 #endif // _TARGET_X86_
1791 bool ehAnyFunclets(); // Are there any funclets in this function?
1792 unsigned ehFuncletCount(); // Return the count of funclets in the function
1794 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1795 #else // !FEATURE_EH_FUNCLETS
1796 bool ehAnyFunclets()
1800 unsigned ehFuncletCount()
1805 unsigned bbThrowIndex(BasicBlock* blk)
1807 return blk->bbTryIndex;
1808 } // Get the index to use as the cache key for sharing throw blocks
1809 #endif // !FEATURE_EH_FUNCLETS
1811 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1812 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1813 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1814 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1815 // convenient to also consider it a predecessor.)
1816 flowList* BlockPredsWithEH(BasicBlock* blk);
1818 // This table is useful for memoization of the method above.
1819 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1821 BlockToFlowListMap* m_blockToEHPreds;
1822 BlockToFlowListMap* GetBlockToEHPreds()
1824 if (m_blockToEHPreds == nullptr)
1826 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1828 return m_blockToEHPreds;
1831 void* ehEmitCookie(BasicBlock* block);
1832 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1834 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1836 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1838 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1840 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1842 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1844 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1846 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1848 void fgAllocEHTable();
1850 void fgRemoveEHTableEntry(unsigned XTnum);
1852 #if FEATURE_EH_FUNCLETS
1854 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1856 #endif // FEATURE_EH_FUNCLETS
1860 #endif // !FEATURE_EH
1862 void fgSortEHTable();
1864 // Causes the EH table to obey some well-formedness conditions, by inserting
1865 // empty BB's when necessary:
1866 // * No block is both the first block of a handler and the first block of a try.
1867 // * No block is the first block of multiple 'try' regions.
1868 // * No block is the last block of multiple EH regions.
1869 void fgNormalizeEH();
1870 bool fgNormalizeEHCase1();
1871 bool fgNormalizeEHCase2();
1872 bool fgNormalizeEHCase3();
1875 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1876 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1877 void fgVerifyHandlerTab();
1878 void fgDispHandlerTab();
1881 bool fgNeedToSortEHTable;
1883 void verInitEHTree(unsigned numEHClauses);
1884 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1885 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1886 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1887 void verCheckNestingLevel(EHNodeDsc* initRoot);
1890 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1891 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1893 XX GenTree and BasicBlock XX
1895 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1897 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1898 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1901 // Functions to create nodes
1902 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1905 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1907 // For binary opers.
1908 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1910 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1912 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1913 var_types type = TYP_I_IMPL,
1914 GenTreePtr op1 = nullptr,
1915 GenTreePtr op2 = nullptr);
1917 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1919 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1921 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1923 GenTreePtr gtNewJmpTableNode();
1924 GenTreePtr gtNewIconHandleNode(
1925 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1927 unsigned gtTokenToIconFlags(unsigned token);
1929 GenTreePtr gtNewIconEmbHndNode(void* value,
1932 unsigned handle1 = 0,
1933 void* handle2 = nullptr,
1934 void* compileTimeHandle = nullptr);
1936 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1937 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1938 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1939 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1941 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1943 GenTreePtr gtNewLconNode(__int64 value);
1945 GenTreePtr gtNewDconNode(double value);
1947 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1949 GenTreePtr gtNewZeroConNode(var_types type);
1951 GenTreePtr gtNewOneConNode(var_types type);
1954 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1955 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1958 GenTreeBlk* gtNewBlkOpNode(
1959 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1961 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1964 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1967 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1968 void gtSetObjGcInfo(GenTreeObj* objNode);
1969 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1970 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1972 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1974 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1976 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1977 CORINFO_METHOD_HANDLE handle,
1979 GenTreeArgList* args,
1980 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1982 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1984 GenTreeArgList* args,
1985 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1987 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1990 GenTreeArgList* args = nullptr);
1992 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1995 GenTreeSIMD* gtNewSIMDNode(
1996 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1997 GenTreeSIMD* gtNewSIMDNode(var_types type,
2000 SIMDIntrinsicID simdIntrinsicID,
2003 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2006 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2007 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2008 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2010 GenTreePtr gtNewCodeRef(BasicBlock* block);
2012 GenTreePtr gtNewFieldRef(
2013 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2015 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2017 GenTreeArgList* gtNewArgList(GenTreePtr op);
2018 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2019 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2021 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2022 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2023 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2024 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2026 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2028 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2030 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2031 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2032 CORINFO_ACCESS_FLAGS access,
2033 CORINFO_FIELD_INFO* pFieldInfo,
2035 CORINFO_CLASS_HANDLE structType,
2038 GenTreePtr gtNewNothingNode();
2040 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2042 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2044 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2046 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2048 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2050 //------------------------------------------------------------------------
2051 // Other GenTree functions
2053 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2055 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2056 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2057 // IntCnses with value `deepVarVal`.
2058 GenTreePtr gtCloneExpr(
2059 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2061 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2062 // `varNum` to int constants with value `varVal`.
2063 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2065 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2068 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2070 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2072 // Returns "true" iff the complexity (not formally defined, but first interpretation
2073 // is #of nodes in subtree) of "tree" is greater than "limit".
2074 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2075 // before they have been set.)
2076 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2078 bool gtCompareTree(GenTree* op1, GenTree* op2);
2080 GenTreePtr gtReverseCond(GenTree* tree);
2082 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2084 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2086 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2088 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2091 unsigned gtHashValue(GenTree* tree);
2093 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2096 void gtPrepareCost(GenTree* tree);
2097 bool gtIsLikelyRegVar(GenTree* tree);
2099 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2101 // Returns true iff the secondNode can be swapped with firstNode.
2102 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2104 unsigned gtSetEvalOrder(GenTree* tree);
2106 #if FEATURE_STACK_FP_X87
2108 void gtComputeFPlvls(GenTreePtr tree);
2109 #endif // FEATURE_STACK_FP_X87
2111 void gtSetStmtInfo(GenTree* stmt);
2113 // Returns "true" iff "node" has any of the side effects in "flags".
2114 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2116 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2117 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2119 // Appends 'expr' in front of 'list'
2120 // 'list' will typically start off as 'nullptr'
2121 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2122 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2124 void gtExtractSideEffList(GenTreePtr expr,
2126 unsigned flags = GTF_SIDE_EFFECT,
2127 bool ignoreRoot = false);
2129 GenTreePtr gtGetThisArg(GenTreePtr call);
2131 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2132 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2133 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2134 // the given "fldHnd", is such an object pointer.
2135 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2137 // Return true if call is a recursive call; return false otherwise.
2138 bool gtIsRecursiveCall(GenTreeCall* call)
2140 return (call->gtCallMethHnd == info.compMethodHnd);
2143 //-------------------------------------------------------------------------
2145 GenTreePtr gtFoldExpr(GenTreePtr tree);
2148 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2149 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2150 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2151 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2152 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2153 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2154 // optimizations for now.
2155 __attribute__((optnone))
2157 gtFoldExprConst(GenTreePtr tree);
2158 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2159 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2161 //-------------------------------------------------------------------------
2162 // Get the handle, if any.
2163 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2164 // Get the handle, and assert if not found.
2165 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2167 //-------------------------------------------------------------------------
2168 // Functions to display the trees
2171 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2173 void gtDispVN(GenTreePtr tree);
2174 void gtDispConst(GenTreePtr tree);
2175 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2176 void gtDispNodeName(GenTreePtr tree);
2177 void gtDispRegVal(GenTreePtr tree);
2189 void gtDispChild(GenTreePtr child,
2190 IndentStack* indentStack,
2192 __in_opt const char* msg = nullptr,
2193 bool topOnly = false);
2194 void gtDispTree(GenTreePtr tree,
2195 IndentStack* indentStack = nullptr,
2196 __in_opt const char* msg = nullptr,
2197 bool topOnly = false,
2198 bool isLIR = false);
2199 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2200 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2201 char* gtGetLclVarName(unsigned lclNum);
2202 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2203 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2204 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2205 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2206 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2207 void gtDispFieldSeq(FieldSeqNode* pfsn);
2209 void gtDispRange(LIR::ReadOnlyRange const& range);
2211 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2213 void gtDispLIRNode(GenTree* node);
2225 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2226 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2229 static fgWalkPreFn gtAssertColonCond;
2231 static fgWalkPreFn gtMarkColonCond;
2232 static fgWalkPreFn gtClearColonCond;
2234 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2235 bool gtHasCatchArg(GenTreePtr tree);
2236 bool gtHasUnmanagedCall(GenTreePtr tree);
2238 typedef ArrayStack<GenTree*> GenTreeStack;
2240 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2241 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2243 //=========================================================================
2244 // BasicBlock functions
2246 // This is a debug flag we will use to assert when creating block during codegen
2247 // as this interferes with procedure splitting. If you know what you're doing, set
2248 // it to true before creating the block. (DEBUG only)
2249 bool fgSafeBasicBlockCreation;
2252 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2255 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2256 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2260 XX The variables to be used by the code generator. XX
2262 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2263 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2267 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2268 // be placed in the stack frame and it's fields must be laid out sequentially.
2270 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2271 // a local variable that can be enregistered or placed in the stack frame.
2272 // The fields do not need to be laid out sequentially
2274 enum lvaPromotionType
2276 PROMOTION_TYPE_NONE, // The struct local is not promoted
2277 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2278 // and its field locals are independent of its parent struct local.
2279 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2280 // but its field locals depend on its parent struct local.
2283 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2284 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2286 /*****************************************************************************/
2288 enum FrameLayoutState
2291 INITIAL_FRAME_LAYOUT,
2292 PRE_REGALLOC_FRAME_LAYOUT,
2293 REGALLOC_FRAME_LAYOUT,
2294 TENTATIVE_FRAME_LAYOUT,
2299 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2300 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2301 bool lvaSortAgain; // true: We need to sort the lvaTable
2302 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2303 unsigned lvaCount; // total number of locals
2305 unsigned lvaRefCount; // total number of references to locals
2306 LclVarDsc* lvaTable; // variable descriptor table
2307 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2309 LclVarDsc** lvaRefSorted; // table sorted by refcount
2311 unsigned short lvaTrackedCount; // actual # of locals being tracked
2312 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2314 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2315 // Only for AMD64 System V cache the first caller stack homed argument.
2316 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2317 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2320 VARSET_TP lvaTrackedVars; // set of tracked variables
2322 #ifndef _TARGET_64BIT_
2323 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2325 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2327 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2328 // It that changes, this changes. VarSets from different epochs
2329 // cannot be meaningfully combined.
2331 unsigned GetCurLVEpoch()
2336 // reverse map of tracked number to var number
2337 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2339 #ifdef LEGACY_BACKEND
2340 // variable interference graph
2341 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2344 // variable preference graph
2345 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2349 // # of procs compiled a with double-aligned stack
2350 static unsigned s_lvaDoubleAlignedProcsCount;
2354 // Getters and setters for address-exposed and do-not-enregister local var properties.
2355 bool lvaVarAddrExposed(unsigned varNum);
2356 void lvaSetVarAddrExposed(unsigned varNum);
2357 bool lvaVarDoNotEnregister(unsigned varNum);
2359 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2360 enum DoNotEnregisterReason
2365 DNER_VMNeedsStackAddr,
2366 DNER_LiveInOutOfHandler,
2367 DNER_LiveAcrossUnmanagedCall,
2368 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2369 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2370 #ifdef JIT32_GCENCODER
2375 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2377 unsigned lvaVarargsHandleArg;
2379 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2381 #endif // _TARGET_X86_
2383 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2384 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2385 #if FEATURE_FIXED_OUT_ARGS
2386 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2388 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2389 // that tracks whether the lock has been taken
2391 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2392 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2393 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2395 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2396 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2398 #if FEATURE_FIXED_OUT_ARGS
2399 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2400 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2401 #endif // FEATURE_FIXED_OUT_ARGS
2404 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2405 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2406 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2407 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2408 // this variable to be this scratch word whenever struct promotion occurs.
2409 unsigned lvaPromotedStructAssemblyScratchVar;
2410 #endif // _TARGET_ARM_
2413 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2414 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2417 bool lvaGenericsContextUsed;
2419 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2420 // CORINFO_GENERICS_CTXT_FROM_THIS?
2421 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2423 //-------------------------------------------------------------------------
2424 // All these frame offsets are inter-related and must be kept in sync
2426 #if !FEATURE_EH_FUNCLETS
2427 // This is used for the callable handlers
2428 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2429 #endif // FEATURE_EH_FUNCLETS
2431 unsigned lvaCachedGenericContextArgOffs;
2432 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2435 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2437 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2439 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2440 // after the reg predict we will use a computed maxTmpSize
2441 // which is based upon the number of spill temps predicted by reg predict
2442 // All this is necessary because if we under-estimate the size of the spill
2443 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2445 // Pre codegen max spill temp size.
2446 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2448 //-------------------------------------------------------------------------
2450 unsigned lvaGetMaxSpillTempSize();
2452 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2453 #endif // _TARGET_ARM_
2454 void lvaAssignFrameOffsets(FrameLayoutState curState);
2455 void lvaFixVirtualFrameOffsets();
2457 #ifndef LEGACY_BACKEND
2458 void lvaUpdateArgsWithInitialReg();
2459 #endif // !LEGACY_BACKEND
2461 void lvaAssignVirtualFrameOffsetsToArgs();
2462 #ifdef UNIX_AMD64_ABI
2463 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2464 #else // !UNIX_AMD64_ABI
2465 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2466 #endif // !UNIX_AMD64_ABI
2467 void lvaAssignVirtualFrameOffsetsToLocals();
2468 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2469 #ifdef _TARGET_AMD64_
2470 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2471 bool lvaIsCalleeSavedIntRegCountEven();
2473 void lvaAlignFrame();
2474 void lvaAssignFrameOffsetsToPromotedStructs();
2475 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2478 void lvaDumpRegLocation(unsigned lclNum);
2479 void lvaDumpFrameLocation(unsigned lclNum);
2480 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2481 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2482 // layout state defined by lvaDoneFrameLayout
2485 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2486 // to avoid bugs from borderline cases.
2487 #define MAX_FrameSize 0x3FFFFFFF
2488 void lvaIncrementFrameSize(unsigned size);
2490 unsigned lvaFrameSize(FrameLayoutState curState);
2492 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2493 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2495 // Returns the caller-SP-relative offset for the local variable "varNum."
2496 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2498 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2499 int lvaGetSPRelativeOffset(unsigned varNum);
2501 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2502 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2504 //------------------------ For splitting types ----------------------------
2506 void lvaInitTypeRef();
2508 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2509 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2510 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2511 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2512 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2513 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2515 void lvaInitVarDsc(LclVarDsc* varDsc,
2517 CorInfoType corInfoType,
2518 CORINFO_CLASS_HANDLE typeHnd,
2519 CORINFO_ARG_LIST_HANDLE varList,
2520 CORINFO_SIG_INFO* varSig);
2522 static unsigned lvaTypeRefMask(var_types type);
2524 var_types lvaGetActualType(unsigned lclNum);
2525 var_types lvaGetRealType(unsigned lclNum);
2527 //-------------------------------------------------------------------------
2531 unsigned lvaLclSize(unsigned varNum);
2532 unsigned lvaLclExactSize(unsigned varNum);
2534 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2536 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2537 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2538 // the return result.
2539 bool lvaLclVarRefsAccum(
2540 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2542 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2543 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2544 // and (destructively) unions "trkedVars" into "*result".
2545 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2547 ALLVARSET_VALARG_TP allVars,
2548 VARSET_VALARG_TP trkdVars);
2550 bool lvaHaveManyLocals() const;
2552 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2553 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2554 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2557 void lvaSortByRefCount();
2558 void lvaDumpRefCounts();
2560 void lvaMarkLocalVars(BasicBlock* block);
2562 void lvaMarkLocalVars(); // Local variable ref-counting
2564 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2566 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2568 static fgWalkPreFn lvaIncRefCntsCB;
2569 void lvaIncRefCnts(GenTreePtr tree);
2571 static fgWalkPreFn lvaDecRefCntsCB;
2572 void lvaDecRefCnts(GenTreePtr tree);
2573 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2574 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2575 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2578 struct lvaStressLclFldArgs
2580 Compiler* m_pCompiler;
2584 static fgWalkPreFn lvaStressLclFldCB;
2585 void lvaStressLclFld();
2587 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2588 void lvaDispVarSet(VARSET_VALARG_TP set);
2593 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2595 int lvaFrameAddress(int varNum, bool* pFPbased);
2598 bool lvaIsParameter(unsigned varNum);
2599 bool lvaIsRegArgument(unsigned varNum);
2600 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2601 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2602 // that writes to arg0
2604 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2605 // (this is an overload of lvIsTemp because there are no temp parameters).
2606 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2607 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2608 bool lvaIsImplicitByRefLocal(unsigned varNum)
2610 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2611 LclVarDsc* varDsc = &(lvaTable[varNum]);
2612 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2614 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2617 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2621 // Returns true if this local var is a multireg struct
2622 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2624 // If the class is a TYP_STRUCT, get/set a class handle describing it
2626 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2627 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2629 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2631 // Info about struct fields
2632 struct lvaStructFieldInfo
2634 CORINFO_FIELD_HANDLE fldHnd;
2635 unsigned char fldOffset;
2636 unsigned char fldOrdinal;
2639 CORINFO_CLASS_HANDLE fldTypeHnd;
2642 // Info about struct to be promoted.
2643 struct lvaStructPromotionInfo
2645 CORINFO_CLASS_HANDLE typeHnd;
2647 bool requiresScratchVar;
2650 unsigned char fieldCnt;
2651 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2653 lvaStructPromotionInfo()
2654 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2659 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2660 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2661 lvaStructPromotionInfo* StructPromotionInfo,
2663 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2664 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2665 #if !defined(_TARGET_64BIT_)
2666 void lvaPromoteLongVars();
2667 #endif // !defined(_TARGET_64BIT_)
2668 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2669 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2670 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2671 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2672 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2673 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2674 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2676 #if defined(FEATURE_SIMD)
2677 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2679 assert(varDsc->lvType == TYP_SIMD12);
2680 assert(varDsc->lvExactSize == 12);
2682 #if defined(_TARGET_64BIT_)
2683 assert(varDsc->lvSize() == 16);
2685 #else // !defined(_TARGET_64BIT_)
2687 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2688 // already does this calculation. However, we also need to prevent mapping types if the var is a
2689 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2690 // However, we don't know this until late, so we may have already pretended the field is bigger
2692 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2701 #endif // !defined(_TARGET_64BIT_)
2703 #endif // defined(FEATURE_SIMD)
2705 BYTE* lvaGetGcLayout(unsigned varNum);
2706 bool lvaTypeIsGC(unsigned varNum);
2707 unsigned lvaGSSecurityCookie; // LclVar number
2708 bool lvaTempsHaveLargerOffsetThanVars();
2710 unsigned lvaSecurityObject; // variable representing the security object on the stack
2711 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2713 #if FEATURE_EH_FUNCLETS
2714 unsigned lvaPSPSym; // variable representing the PSPSym
2717 InlineInfo* impInlineInfo;
2718 InlineStrategy* m_inlineStrategy;
2720 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2721 Compiler* impInlineRoot();
2723 #if defined(DEBUG) || defined(INLINE_DATA)
2724 unsigned __int64 getInlineCycleCount()
2726 return m_compCycles;
2728 #endif // defined(DEBUG) || defined(INLINE_DATA)
2730 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2731 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2733 //=========================================================================
2735 //=========================================================================
2738 //---------------- Local variable ref-counting ----------------------------
2741 BasicBlock* lvaMarkRefsCurBlock;
2742 GenTreePtr lvaMarkRefsCurStmt;
2744 BasicBlock::weight_t lvaMarkRefsWeight;
2746 static fgWalkPreFn lvaMarkLclRefsCallback;
2747 void lvaMarkLclRefs(GenTreePtr tree);
2749 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2750 PerSsaArray lvHeapPerSsaData;
2751 unsigned lvHeapNumSsaNames;
2754 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2755 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2756 // not an SSA variable).
2757 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2759 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2760 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2762 assert(ssaNum < lvHeapNumSsaNames);
2763 return &lvHeapPerSsaData.GetRef(ssaNum);
2767 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2768 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2772 XX Imports the given method and converts it to semantic trees XX
2774 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2775 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2781 void impImport(BasicBlock* method);
2783 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2784 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2785 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2786 CORINFO_CLASS_HANDLE impGetStringClass();
2787 CORINFO_CLASS_HANDLE impGetObjectClass();
2789 //=========================================================================
2791 //=========================================================================
2794 //-------------------- Stack manipulation ---------------------------------
2796 unsigned impStkSize; // Size of the full stack
2798 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2800 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2802 struct SavedStack // used to save/restore stack contents.
2804 unsigned ssDepth; // number of values on stack
2805 StackEntry* ssTrees; // saved tree values
2808 bool impIsPrimitive(CorInfoType type);
2809 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2811 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2812 void impPushOnStackNoType(GenTreePtr tree);
2814 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2815 void impPushNullObjRefOnStack();
2816 StackEntry impPopStack();
2817 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2818 GenTreePtr impPopStack(typeInfo& ti);
2819 StackEntry& impStackTop(unsigned n = 0);
2821 void impSaveStackState(SavedStack* savePtr, bool copy);
2822 void impRestoreStackState(SavedStack* savePtr);
2824 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2825 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2826 CORINFO_CALL_INFO* pCallInfo);
2828 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2830 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2832 bool impCanPInvokeInline();
2833 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2834 void impCheckForPInvokeCall(
2835 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2836 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2837 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2839 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2840 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2841 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2843 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2844 CORINFO_METHOD_HANDLE calleeMethodHnd,
2845 CORINFO_CLASS_HANDLE delegateTypeHnd);
2847 var_types impImportCall(OPCODE opcode,
2848 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2849 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2851 GenTreePtr newobjThis,
2853 CORINFO_CALL_INFO* callInfo,
2854 IL_OFFSET rawILOffset);
2856 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2858 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2860 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2863 var_types impImportJitTestLabelMark(int numArgs);
2866 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2868 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2870 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2871 CORINFO_ACCESS_FLAGS access,
2872 CORINFO_FIELD_INFO* pFieldInfo,
2875 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2877 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2879 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2881 void impImportLeave(BasicBlock* block);
2882 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2883 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2884 CORINFO_CLASS_HANDLE clsHnd,
2885 CORINFO_METHOD_HANDLE method,
2886 CORINFO_SIG_INFO* sig,
2890 CorInfoIntrinsics* pIntrinsicID);
2891 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2892 CORINFO_SIG_INFO* sig,
2895 CorInfoIntrinsics intrinsicID);
2896 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2898 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2900 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2901 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2902 CORINFO_THIS_TRANSFORM transform);
2904 //----------------- Manipulating the trees and stmts ----------------------
2906 GenTreePtr impTreeList; // Trees for the BB being imported
2907 GenTreePtr impTreeLast; // The last tree for the current BB
2911 CHECK_SPILL_ALL = -1,
2912 CHECK_SPILL_NONE = -2
2916 void impBeginTreeList();
2917 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2918 void impEndTreeList(BasicBlock* block);
2919 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2920 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2921 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2922 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2923 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2924 void impAssignTempGen(unsigned tmp,
2927 GenTreePtr* pAfterStmt = nullptr,
2928 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2929 BasicBlock* block = nullptr);
2930 void impAssignTempGen(unsigned tmpNum,
2932 CORINFO_CLASS_HANDLE structHnd,
2934 GenTreePtr* pAfterStmt = nullptr,
2935 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2936 BasicBlock* block = nullptr);
2937 GenTreePtr impCloneExpr(GenTreePtr tree,
2939 CORINFO_CLASS_HANDLE structHnd,
2941 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2942 GenTreePtr impAssignStruct(GenTreePtr dest,
2944 CORINFO_CLASS_HANDLE structHnd,
2946 GenTreePtr* pAfterStmt = nullptr,
2947 BasicBlock* block = nullptr);
2948 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2950 CORINFO_CLASS_HANDLE structHnd,
2952 GenTreePtr* pAfterStmt = nullptr,
2953 BasicBlock* block = nullptr);
2955 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2956 CORINFO_CLASS_HANDLE structHnd,
2960 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2961 BYTE* gcLayout = nullptr,
2962 unsigned* numGCVars = nullptr,
2963 var_types* simdBaseType = nullptr);
2965 GenTreePtr impNormStructVal(GenTreePtr structVal,
2966 CORINFO_CLASS_HANDLE structHnd,
2968 bool forceNormalization = false);
2970 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2971 BOOL* pRuntimeLookup = nullptr,
2972 BOOL mustRestoreHandle = FALSE,
2973 BOOL importParent = FALSE);
2975 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2976 BOOL* pRuntimeLookup = nullptr,
2977 BOOL mustRestoreHandle = FALSE)
2979 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2982 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2983 CORINFO_LOOKUP* pLookup,
2985 void* compileTimeHandle);
2987 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2989 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2990 CORINFO_LOOKUP* pLookup,
2991 void* compileTimeHandle);
2993 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2995 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2996 CorInfoHelpFunc helper,
2998 GenTreeArgList* arg = nullptr,
2999 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3001 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3003 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3006 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3007 CORINFO_CLASS_HANDLE typeClass,
3011 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3012 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3013 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3014 static bool IsMathIntrinsic(GenTreePtr tree);
3017 //----------------- Importing the method ----------------------------------
3019 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3022 unsigned impCurOpcOffs;
3023 const char* impCurOpcName;
3024 bool impNestedStackSpill;
3026 // For displaying instrs with generated native code (-n:B)
3027 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3028 void impNoteLastILoffs();
3031 /* IL offset of the stmt currently being imported. It gets set to
3032 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3033 updated at IL offsets for which we have to report mapping info.
3034 It also includes flag bits, so use jitGetILoffs()
3035 to get the actual IL offset value.
3038 IL_OFFSETX impCurStmtOffs;
3039 void impCurStmtOffsSet(IL_OFFSET offs);
3041 void impNoteBranchOffs();
3043 unsigned impInitBlockLineInfo();
3045 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3046 bool impIsThis(GenTreePtr obj);
3047 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3048 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3049 bool impIsAnySTLOC(OPCODE opcode)
3051 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3052 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3055 GenTreeArgList* impPopList(unsigned count,
3057 CORINFO_SIG_INFO* sig,
3058 GenTreeArgList* prefixTree = nullptr);
3060 GenTreeArgList* impPopRevList(unsigned count,
3062 CORINFO_SIG_INFO* sig,
3063 unsigned skipReverseCount = 0);
3066 * Get current IL offset with stack-empty info incoporated
3068 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3070 //---------------- Spilling the importer stack ----------------------------
3076 SavedStack pdSavedStack;
3077 ThisInitState pdThisPtrInit;
3080 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3081 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3083 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3084 ExpandArray<BYTE> impPendingBlockMembers;
3086 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3087 // Operates on the map in the top-level ancestor.
3088 BYTE impGetPendingBlockMember(BasicBlock* blk)
3090 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3093 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3094 // Operates on the map in the top-level ancestor.
3095 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3097 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3100 bool impCanReimport;
3102 bool impSpillStackEntry(unsigned level,
3106 bool bAssertOnRecursion,
3111 void impSpillStackEnsure(bool spillLeaves = false);
3112 void impEvalSideEffects();
3113 void impSpillSpecialSideEff();
3114 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3115 void impSpillValueClasses();
3116 void impSpillEvalStack();
3117 static fgWalkPreFn impFindValueClasses;
3118 void impSpillLclRefs(ssize_t lclNum);
3120 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3122 void impImportBlockCode(BasicBlock* block);
3124 void impReimportMarkBlock(BasicBlock* block);
3125 void impReimportMarkSuccessors(BasicBlock* block);
3127 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3129 void impImportBlockPending(BasicBlock* block);
3131 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3132 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3133 // for the block, but instead, just re-uses the block's existing EntryState.
3134 void impReimportBlockPending(BasicBlock* block);
3136 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3138 void impImportBlock(BasicBlock* block);
3140 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3141 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3142 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3143 // the variables that will be used -- and for all the predecessors of those successors, and the
3144 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3145 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3146 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3147 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3148 // of local variable numbers, so we represent them with the base local variable number), returns that.
3149 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3150 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3151 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3152 // on which kind of member of the clique the block is).
3153 unsigned impGetSpillTmpBase(BasicBlock* block);
3155 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3156 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3157 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3158 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3159 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3160 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3161 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3162 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3163 // successors receive a native int. Similarly float and double are unified to double.
3164 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3165 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3166 // predecessors, so they insert an upcast if needed).
3167 void impReimportSpillClique(BasicBlock* block);
3169 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3170 // block, and represent the predecessor and successor members of the clique currently being computed.
3171 // *** Access to these will need to be locked in a parallel compiler.
3172 ExpandArray<BYTE> impSpillCliquePredMembers;
3173 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3181 // Abstract class for receiving a callback while walking a spill clique
3182 class SpillCliqueWalker
3185 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3188 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3189 class SetSpillTempsBase : public SpillCliqueWalker
3194 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3197 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3200 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3201 class ReimportSpillClique : public SpillCliqueWalker
3206 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3209 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3212 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3213 // predecessor or successor within the spill clique
3214 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3216 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3217 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3218 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3219 void impRetypeEntryStateTemps(BasicBlock* blk);
3221 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3222 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3224 void impPushVar(GenTree* op, typeInfo tiRetVal);
3225 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3226 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3228 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3230 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3231 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3232 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3235 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3238 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3239 struct BlockListNode
3242 BlockListNode* m_next;
3243 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3246 void* operator new(size_t sz, Compiler* comp);
3248 BlockListNode* impBlockListNodeFreeList;
3250 BlockListNode* AllocBlockListNode();
3251 void FreeBlockListNode(BlockListNode* node);
3253 bool impIsValueType(typeInfo* pTypeInfo);
3254 var_types mangleVarArgsType(var_types type);
3257 regNumber getCallArgIntRegister(regNumber floatReg);
3258 regNumber getCallArgFloatRegister(regNumber intReg);
3259 #endif // FEATURE_VARARG
3262 static unsigned jitTotalMethodCompiled;
3266 static LONG jitNestingLevel;
3269 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3271 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3273 // STATIC inlining decision based on the IL code.
3274 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3275 CORINFO_METHOD_INFO* methInfo,
3277 InlineResult* inlineResult);
3279 void impCheckCanInline(GenTreePtr call,
3280 CORINFO_METHOD_HANDLE fncHandle,
3282 CORINFO_CONTEXT_HANDLE exactContextHnd,
3283 InlineCandidateInfo** ppInlineCandidateInfo,
3284 InlineResult* inlineResult);
3286 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3287 GenTreePtr curArgVal,
3289 InlineResult* inlineResult);
3291 void impInlineInitVars(InlineInfo* pInlineInfo);
3293 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3295 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3297 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3299 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3300 GenTreePtr variableBeingDereferenced,
3301 InlArgInfo* inlArgInfo);
3303 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3305 bool impTailCallRetTypeCompatible(var_types callerRetType,
3306 CORINFO_CLASS_HANDLE callerRetTypeClass,
3307 var_types calleeRetType,
3308 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3310 bool impIsTailCallILPattern(bool tailPrefixed,
3312 const BYTE* codeAddrOfNextOpcode,
3313 const BYTE* codeEnd,
3315 bool* IsCallPopRet = nullptr);
3317 bool impIsImplicitTailCallCandidate(
3318 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3321 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3322 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3326 XX Info about the basic-blocks, their contents and the flow analysis XX
3328 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3329 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3333 BasicBlock* fgFirstBB; // Beginning of the basic block list
3334 BasicBlock* fgLastBB; // End of the basic block list
3335 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3336 #if FEATURE_EH_FUNCLETS
3337 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3339 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3341 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3342 unsigned fgEdgeCount; // # of control flow edges between the BBs
3343 unsigned fgBBcount; // # of BBs in the method
3345 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3347 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3348 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3349 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3350 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3352 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3353 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3354 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3355 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3356 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3357 // index). The arrays are of size fgBBNumMax + 1.
3358 unsigned* fgDomTreePreOrder;
3359 unsigned* fgDomTreePostOrder;
3361 bool fgBBVarSetsInited;
3363 // Allocate array like T* a = new T[fgBBNumMax + 1];
3364 // Using helper so we don't keep forgetting +1.
3365 template <typename T>
3366 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3368 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3371 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3372 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3373 // cannot be meaningfully combined. Note that new blocks can be created with higher
3374 // block numbers without changing the basic block epoch. These blocks *cannot*
3375 // participate in a block set until the blocks are all renumbered, causing the epoch
3376 // to change. This is useful if continuing to use previous block sets is valuable.
3377 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3378 unsigned fgCurBBEpoch;
3380 unsigned GetCurBasicBlockEpoch()
3382 return fgCurBBEpoch;
3385 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3386 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3387 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3388 unsigned fgCurBBEpochSize;
3390 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3391 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3392 unsigned fgBBSetCountInSizeTUnits;
3394 void NewBasicBlockEpoch()
3396 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3398 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3400 fgCurBBEpochSize = fgBBNumMax + 1;
3401 fgBBSetCountInSizeTUnits =
3402 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3405 // All BlockSet objects are now invalid!
3406 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3407 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3411 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3412 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3413 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3414 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3416 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3417 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3418 // array of size_t bitsets), then print that out.
3419 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3426 void EnsureBasicBlockEpoch()
3428 if (fgCurBBEpochSize != fgBBNumMax + 1)
3430 NewBasicBlockEpoch();
3434 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3435 void fgEnsureFirstBBisScratch();
3436 bool fgFirstBBisScratch();
3437 bool fgBBisScratch(BasicBlock* block);
3439 void fgExtendEHRegionBefore(BasicBlock* block);
3440 void fgExtendEHRegionAfter(BasicBlock* block);
3442 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3444 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3446 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3449 BasicBlock* nearBlk,
3450 bool putInFilter = false,
3451 bool runRarely = false,
3452 bool insertAtEnd = false);
3454 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3456 bool runRarely = false,
3457 bool insertAtEnd = false);
3459 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3461 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3462 BasicBlock* afterBlk,
3463 unsigned xcptnIndex,
3464 bool putInTryRegion);
3466 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3467 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3468 void fgUnlinkBlock(BasicBlock* block);
3470 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3471 bool fgMultipleNots;
3474 bool fgModified; // True if the flow graph has been modified recently
3475 bool fgComputePredsDone; // Have we computed the bbPreds list
3476 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3477 bool fgDomsComputed; // Have we computed the dominator sets?
3478 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3480 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3481 bool fgHasPostfix; // any postfix ++/-- found?
3482 unsigned fgIncrCount; // number of increment nodes found
3484 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3488 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3489 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3492 bool fgRemoveRestOfBlock; // true if we know that we will throw
3493 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3495 // There are two modes for ordering of the trees.
3496 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3497 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3498 // by traversing the tree according to the order of the operands.
3499 // - In FGOrderLinear, the dominant ordering is the linear order.
3506 FlowGraphOrder fgOrder;
3508 // The following are boolean flags that keep track of the state of internal data structures
3510 bool fgStmtListThreaded;
3511 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3512 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3513 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3514 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3515 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3516 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3517 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3518 // This is derived from the profile data
3519 // or is BB_UNITY_WEIGHT when we don't have profile data
3521 #if FEATURE_EH_FUNCLETS
3522 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3523 #endif // FEATURE_EH_FUNCLETS
3525 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3526 // since fgMorphTree can be called from several places
3527 bool fgExpandInline; // indicates that we are creating tree for the inliner
3529 bool impBoxTempInUse; // the temp below is valid and available
3530 unsigned impBoxTemp; // a temporary that is used for boxing
3533 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3534 // and we are trying to compile again in a "safer", minopts mode?
3538 unsigned impInlinedCodeSize;
3541 //-------------------------------------------------------------------------
3549 void fgRemoveEmptyTry();
3551 void fgRemoveEmptyFinally();
3553 void fgCloneFinally();
3555 void fgCleanupContinuation(BasicBlock* continuation);
3557 void fgUpdateFinallyTargetFlags();
3559 GenTreePtr fgGetCritSectOfStaticMethod();
3561 #if !defined(_TARGET_X86_)
3563 void fgAddSyncMethodEnterExit();
3565 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3567 void fgConvertSyncReturnToLeave(BasicBlock* block);
3569 #endif // !_TARGET_X86_
3571 void fgAddReversePInvokeEnterExit();
3573 bool fgMoreThanOneReturnBlock();
3575 // The number of separate return points in the method.
3576 unsigned fgReturnCount;
3578 void fgAddInternal();
3580 bool fgFoldConditional(BasicBlock* block);
3582 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3583 void fgMorphBlocks();
3585 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3587 void fgCheckArgCnt();
3588 void fgSetOptions();
3591 static fgWalkPreFn fgAssertNoQmark;
3592 void fgPreExpandQmarkChecks(GenTreePtr expr);
3593 void fgPostExpandQmarkChecks();
3594 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3597 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3599 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3600 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3601 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3602 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3603 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3605 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3606 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3607 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3608 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3610 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3611 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3612 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3613 void fgExpandQmarkNodes();
3617 // Do "simple lowering." This functionality is (conceptually) part of "general"
3618 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3619 void fgSimpleLowering();
3621 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3623 GenTreePtr fgInitThisClass();
3625 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3627 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3629 void fgLocalVarLiveness();
3631 void fgLocalVarLivenessInit();
3633 #ifdef LEGACY_BACKEND
3634 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3636 void fgPerNodeLocalVarLiveness(GenTree* node);
3638 void fgPerBlockLocalVarLiveness();
3640 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3642 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3644 // This is used in the liveness computation, as a temporary. When we use the
3645 // arbitrary-length VarSet representation, it is better not to allocate a new one
3647 VARSET_TP fgMarkIntfUnionVS;
3649 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3651 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3653 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3655 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3657 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3659 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3661 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3662 GenTreePtr startNode,
3664 VARSET_VALARG_TP volatileVars,
3665 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3667 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3669 bool fgRemoveDeadStore(GenTree** pTree,
3673 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3675 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3677 // For updating liveset during traversal AFTER fgComputeLife has completed
3678 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3679 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3681 // Returns the set of live variables after endTree,
3682 // assuming that liveSet is the set of live variables BEFORE tree.
3683 // Requires that fgComputeLife has completed, and that tree is in the same
3684 // statement as endTree, and that it comes before endTree in execution order
3686 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3688 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3689 while (tree != nullptr && tree != endTree->gtNext)
3691 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3692 tree = tree->gtNext;
3694 assert(tree == endTree->gtNext);
3698 void fgInterBlockLocalVarLiveness();
3700 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3701 // "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
3702 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3703 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3704 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3705 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3706 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3708 if (m_opAsgnVarDefSsaNums == nullptr)
3710 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3712 return m_opAsgnVarDefSsaNums;
3715 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3716 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3717 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3719 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3721 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3722 // Except: assumes that lcl is a def, and if it is
3723 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3724 // rather than the "use" SSA number recorded in the tree "lcl".
3725 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3727 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3728 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3729 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3730 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3731 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3733 // (byref addrS1 = &s1,
3734 // *(addrS1 * offsetof(f0)) = s2f0,
3736 // *(addrS1 * offsetof(fn)) = s2fn)
3738 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3739 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3740 // give it SSA names and value numbers?
3742 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3743 // end with an instance of the structure below, whose fields are described in the declaration.
3744 struct IndirectAssignmentAnnotation
3746 unsigned m_lclNum; // The local num that is being indirectly assigned.
3747 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3748 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3749 // be the singleton field sequence "g". The individual assignments would
3750 // further append the fields of "s.g" to that.
3751 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3752 // structure has a single field).
3753 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3754 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3757 IndirectAssignmentAnnotation(unsigned lclNum,
3758 FieldSeqNode* fldSeq,
3760 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3761 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3762 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3766 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3767 NodeToIndirAssignMap;
3768 NodeToIndirAssignMap* m_indirAssignMap;
3769 NodeToIndirAssignMap* GetIndirAssignMap()
3771 if (m_indirAssignMap == nullptr)
3773 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3774 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3775 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3777 return m_indirAssignMap;
3780 // Performs SSA conversion.
3783 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3784 void fgResetForSsa();
3786 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3788 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3789 inline bool fgExcludeFromSsa(unsigned lclNum);
3791 // The value numbers for this compilation.
3792 ValueNumStore* vnStore;
3795 ValueNumStore* GetValueNumStore()
3800 // Do value numbering (assign a value number to each
3802 void fgValueNumber();
3804 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3805 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3806 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3807 // match the element type of the array or fldSeq. When this type doesn't match
3808 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3810 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3813 FieldSeqNode* fldSeq,
3817 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3818 // has been parsed to yield the other input arguments. If evaluation of the address
3819 // can raise exceptions, those should be captured in the exception set "excVN."
3820 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3821 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3822 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3823 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3824 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3826 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3827 CORINFO_CLASS_HANDLE elemTypeEq,
3831 FieldSeqNode* fldSeq);
3833 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3834 // by evaluating the array index expression "tree". Returns the value number resulting from
3835 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3836 // "GT_IND" that does the dereference, and it is given the returned value number.
3837 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3839 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3841 // Utility functions for fgValueNumber.
3843 // Perform value-numbering for the trees in "blk".
3844 void fgValueNumberBlock(BasicBlock* blk);
3846 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3847 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3848 // assumed for the heap at the start "entryBlk".
3849 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3851 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3852 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3854 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3855 // value in that SSA #.
3856 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3858 // The input 'tree' is a leaf node that is a constant
3859 // Assign the proper value number to the tree
3860 void fgValueNumberTreeConst(GenTreePtr tree);
3862 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3863 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3865 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3867 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3869 // Does value-numbering for a block assignment.
3870 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3872 // Does value-numbering for a cast tree.
3873 void fgValueNumberCastTree(GenTreePtr tree);
3875 // Does value-numbering for an intrinsic tree.
3876 void fgValueNumberIntrinsic(GenTreePtr tree);
3878 // Does value-numbering for a call. We interpret some helper calls.
3879 void fgValueNumberCall(GenTreeCall* call);
3881 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3882 void fgUpdateArgListVNs(GenTreeArgList* args);
3884 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3885 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3887 // Requires "helpCall" to be a helper call. Assigns it a value number;
3888 // we understand the semantics of some of the calls. Returns "true" if
3889 // the call may modify the heap (we assume arbitrary memory side effects if so).
3890 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3892 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3893 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3895 // This is the current value number for the "Heap" implicit variable while
3896 // doing value numbering. This is the value number under the "liberal" interpretation
3897 // of heap values; the "conservative" interpretation needs no VN, since every access of
3898 // the heap yields an unknown value.
3899 ValueNum fgCurHeapVN;
3901 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3902 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3903 // is 1, and the rest is an encoding of "elemTyp".
3904 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3906 if (elemStructType != nullptr)
3908 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3909 varTypeIsIntegral(elemTyp));
3910 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3911 return elemStructType;
3915 elemTyp = varTypeUnsignedToSigned(elemTyp);
3916 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3919 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3920 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3921 // the struct type of the element).
3922 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3924 size_t clsHndVal = size_t(clsHnd);
3925 if (clsHndVal & 0x1)
3927 return var_types(clsHndVal >> 1);
3935 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3936 var_types getJitGCType(BYTE gcType);
3938 enum structPassingKind
3940 SPK_Unknown, // Invalid value, never returned
3941 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3942 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3943 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3944 // parameters registers are used, then the stack will be used)
3945 // for X86 passed on the stack, for ARM32 passed in registers
3946 // or the stack or split between registers and the stack.
3947 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3949 }; // The struct is passed/returned by reference to a copy/buffer.
3951 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3952 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3953 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3954 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3956 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3958 // Get the type that is used to pass values of the given struct type.
3959 // If you have already retrieved the struct size then pass it as the optional third argument
3961 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3962 structPassingKind* wbPassStruct,
3963 unsigned structSize = 0);
3965 // Get the type that is used to return values of the given struct type.
3966 // If you have already retrieved the struct size then pass it as the optional third argument
3968 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3969 structPassingKind* wbPassStruct = nullptr,
3970 unsigned structSize = 0);
3973 // Print a representation of "vnp" or "vn" on standard output.
3974 // If "level" is non-zero, we also print out a partial expansion of the value.
3975 void vnpPrint(ValueNumPair vnp, unsigned level);
3976 void vnPrint(ValueNum vn, unsigned level);
3979 // Dominator computation member functions
3980 // Not exposed outside Compiler
3982 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3984 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3986 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3987 // flow graph. We first assume the fields bbIDom on each
3988 // basic block are invalid. This computation is needed later
3989 // by fgBuildDomTree to build the dominance tree structure.
3990 // Based on: A Simple, Fast Dominance Algorithm
3991 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3993 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3994 // Note: this is relatively slow compared to calling fgDominate(),
3995 // especially if dealing with a single block versus block check.
3997 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3999 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4001 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4003 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4005 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4007 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4008 // processed in topological sort, this function takes care of that.
4010 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4012 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4013 // Returns this as a set.
4015 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4016 // root nodes. Returns this as a set.
4019 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4022 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4023 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4026 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4027 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4028 // && postOrder(A) >= postOrder(B) making the computation O(1).
4029 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4031 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4033 void fgUpdateChangedFlowGraph();
4036 // Compute the predecessors of the blocks in the control flow graph.
4037 void fgComputePreds();
4039 // Remove all predecessor information.
4040 void fgRemovePreds();
4042 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4043 // before the full predecessors lists are computed.
4044 void fgComputeCheapPreds();
4047 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4049 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4059 // Initialize the per-block variable sets (used for liveness analysis).
4060 void fgInitBlockVarSets();
4062 // true if we've gone through and created GC Poll calls.
4063 bool fgGCPollsCreated;
4064 void fgMarkGCPollBlocks();
4065 void fgCreateGCPolls();
4066 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4068 // Requires that "block" is a block that returns from
4069 // a finally. Returns the number of successors (jump targets of
4070 // of blocks in the covered "try" that did a "LEAVE".)
4071 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4073 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4074 // a finally. Returns its "i"th successor (jump targets of
4075 // of blocks in the covered "try" that did a "LEAVE".)
4076 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4077 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4080 // Factor out common portions of the impls of the methods above.
4081 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4084 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4085 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4086 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4087 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4088 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4089 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4090 // we leave the entry associated with the block, but it will no longer be accessed.)
4091 struct SwitchUniqueSuccSet
4093 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4094 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4097 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4098 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4099 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4100 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4103 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4104 BlockToSwitchDescMap;
4107 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4108 // iteration over only the distinct successors.
4109 BlockToSwitchDescMap* m_switchDescMap;
4112 BlockToSwitchDescMap* GetSwitchDescMap()
4114 if (m_switchDescMap == nullptr)
4116 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4118 return m_switchDescMap;
4121 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4122 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4123 // we don't accidentally look up and return the wrong switch data.
4124 void InvalidateUniqueSwitchSuccMap()
4126 m_switchDescMap = nullptr;
4129 // Requires "switchBlock" to be a block that ends in a switch. Returns
4130 // the corresponding SwitchUniqueSuccSet.
4131 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4133 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4134 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4135 // remove it from "this", and ensure that "to" is a member.
4136 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4138 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4139 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4141 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4143 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4145 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4147 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4149 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4151 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4153 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4155 void fgRemoveBlockAsPred(BasicBlock* block);
4157 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4159 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4161 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4163 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4165 flowList* fgAddRefPred(BasicBlock* block,
4166 BasicBlock* blockPred,
4167 flowList* oldEdge = nullptr,
4168 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4171 void fgFindBasicBlocks();
4173 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4175 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4177 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4178 bool putInTryRegion,
4179 BasicBlock* startBlk,
4181 BasicBlock* nearBlk,
4182 BasicBlock* jumpBlk,
4185 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4187 void fgRemoveEmptyBlocks();
4189 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4191 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4193 void fgCreateLoopPreHeader(unsigned lnum);
4195 void fgUnreachableBlock(BasicBlock* block);
4197 void fgRemoveConditionalJump(BasicBlock* block);
4199 BasicBlock* fgLastBBInMainFunction();
4201 BasicBlock* fgEndBBAfterMainFunction();
4203 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4205 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4207 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4209 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4211 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4213 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4215 bool fgRenumberBlocks();
4217 bool fgExpandRarelyRunBlocks();
4219 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4221 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4223 enum FG_RELOCATE_TYPE
4225 FG_RELOCATE_TRY, // relocate the 'try' region
4226 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4228 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4230 #if FEATURE_EH_FUNCLETS
4231 #if defined(_TARGET_ARM_)
4232 void fgClearFinallyTargetBit(BasicBlock* block);
4233 #endif // defined(_TARGET_ARM_)
4234 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4235 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4236 void fgInsertFuncletPrologBlock(BasicBlock* block);
4237 void fgCreateFuncletPrologBlocks();
4238 void fgCreateFunclets();
4239 #else // !FEATURE_EH_FUNCLETS
4240 bool fgRelocateEHRegions();
4241 #endif // !FEATURE_EH_FUNCLETS
4243 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4245 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4247 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4249 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4251 bool fgOptimizeEmptyBlock(BasicBlock* block);
4253 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4255 bool fgOptimizeBranch(BasicBlock* bJump);
4257 bool fgOptimizeSwitchBranches(BasicBlock* block);
4259 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4261 bool fgOptimizeSwitchJumps();
4263 void fgPrintEdgeWeights();
4265 void fgComputeEdgeWeights();
4267 void fgReorderBlocks();
4269 void fgDetermineFirstColdBlock();
4271 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4273 bool fgUpdateFlowGraph(bool doTailDup = false);
4275 void fgFindOperOrder();
4277 // method that returns if you should split here
4278 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4280 void fgSetBlockOrder();
4282 void fgRemoveReturnBlock(BasicBlock* block);
4284 /* Helper code that has been factored out */
4285 inline void fgConvertBBToThrowBB(BasicBlock* block);
4287 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4288 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4289 GenTreePtr fgMakeTmpArgNode(
4290 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4292 // The following check for loops that don't execute calls
4293 bool fgLoopCallMarked;
4295 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4296 void fgLoopCallMark();
4298 void fgMarkLoopHead(BasicBlock* block);
4300 unsigned fgGetCodeEstimate(BasicBlock* block);
4303 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4304 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4305 bool fgDumpFlowGraph(Phases phase);
4307 #endif // DUMP_FLOWGRAPHS
4312 void fgDispBBLiveness(BasicBlock* block);
4313 void fgDispBBLiveness();
4314 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4315 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4316 void fgDispBasicBlocks(bool dumpTrees = false);
4317 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4318 void fgDumpBlock(BasicBlock* block);
4319 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4321 static fgWalkPreFn fgStress64RsltMulCB;
4322 void fgStress64RsltMul();
4323 void fgDebugCheckUpdate();
4324 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4325 void fgDebugCheckBlockLinks();
4326 void fgDebugCheckLinks(bool morphTrees = false);
4327 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4328 void fgDebugCheckFlags(GenTreePtr tree);
4329 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4330 void fgDebugCheckTryFinallyExits();
4333 #ifdef LEGACY_BACKEND
4334 static void fgOrderBlockOps(GenTreePtr tree,
4338 GenTreePtr* opsPtr, // OUT
4339 regMaskTP* regsPtr); // OUT
4340 #endif // LEGACY_BACKEND
4342 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4343 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4345 inline bool fgIsInlining()
4347 return fgExpandInline;
4350 void fgTraverseRPO();
4352 //--------------------- Walking the trees in the IR -----------------------
4357 fgWalkPreFn* wtprVisitorFn;
4358 fgWalkPostFn* wtpoVisitorFn;
4359 void* pCallbackData; // user-provided data
4360 bool wtprLclsOnly; // whether to only visit lclvar nodes
4361 GenTreePtr parent; // parent of current node, provided to callback
4362 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4364 bool printModified; // callback can use this
4368 template <bool computeStack>
4369 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4371 // general purpose tree-walker that is capable of doing pre- and post- order
4372 // callbacks at the same time
4373 template <bool doPreOrder, bool doPostOrder>
4374 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4376 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4377 fgWalkPreFn* visitor,
4378 void* pCallBackData = nullptr,
4379 bool lclVarsOnly = false,
4380 bool computeStack = false);
4382 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4383 fgWalkPreFn* preVisitor,
4384 fgWalkPostFn* postVisitor,
4385 void* pCallBackData = nullptr);
4387 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4391 template <bool computeStack>
4392 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4394 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4395 fgWalkPostFn* visitor,
4396 void* pCallBackData = nullptr,
4397 bool computeStack = false);
4399 // An fgWalkPreFn that looks for expressions that have inline throws in
4400 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4401 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4402 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4403 // properly propagated to parent trees). It returns WALK_CONTINUE
4405 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4406 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4407 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4409 /**************************************************************************
4411 *************************************************************************/
4414 friend class SsaBuilder;
4415 friend struct ValueNumberState;
4417 //--------------------- Detect the basic blocks ---------------------------
4419 BasicBlock** fgBBs; // Table of pointers to the BBs
4421 void fgInitBBLookup();
4422 BasicBlock* fgLookupBB(unsigned addr);
4424 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4426 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4428 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4430 void fgLinkBasicBlocks();
4432 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4434 void fgCheckBasicBlockControlFlow();
4436 void fgControlFlowPermitted(BasicBlock* blkSrc,
4437 BasicBlock* blkDest,
4438 BOOL IsLeave = false /* is the src a leave block */);
4440 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4442 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4444 void fgAdjustForAddressExposedOrWrittenThis();
4446 bool fgProfileData_ILSizeMismatch;
4447 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4448 ULONG fgProfileBufferCount;
4449 ULONG fgNumProfileRuns;
4451 unsigned fgStressBBProf()
4454 unsigned result = JitConfig.JitStressBBProf();
4457 if (compStressCompile(STRESS_BB_PROFILE, 15))
4468 bool fgHaveProfileData();
4469 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4471 bool fgIsUsingProfileWeights()
4473 return (fgHaveProfileData() || fgStressBBProf());
4475 void fgInstrumentMethod();
4477 //-------- Insert a statement at the start or end of a basic block --------
4481 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4485 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4487 public: // Used by linear scan register allocation
4488 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4491 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4492 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4494 public: // Used by linear scan register allocation
4495 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4498 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4500 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4502 // Create a new temporary variable to hold the result of *ppTree,
4503 // and transform the graph accordingly.
4504 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4505 GenTree* fgMakeMultiUse(GenTree** ppTree);
4508 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4509 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4510 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4512 //-------- Determine the order in which the trees will be evaluated -------
4514 unsigned fgTreeSeqNum;
4515 GenTree* fgTreeSeqLst;
4516 GenTree* fgTreeSeqBeg;
4518 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4519 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4520 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4521 void fgSetStmtSeq(GenTree* tree);
4522 void fgSetBlockOrder(BasicBlock* block);
4524 //------------------------- Morphing --------------------------------------
4526 unsigned fgPtrArgCntCur;
4527 unsigned fgPtrArgCntMax;
4528 hashBv* fgOutgoingArgTemps;
4529 hashBv* fgCurrentlyInUseArgTemps;
4531 bool compCanEncodePtrArgCntMax();
4533 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4536 void fgMoveOpsLeft(GenTreePtr tree);
4539 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4541 bool fgIsThrow(GenTreePtr tree);
4543 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4544 bool fgIsBlockCold(BasicBlock* block);
4546 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4548 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4550 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4552 bool fgMorphRelopToQmark(GenTreePtr tree);
4554 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4555 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4556 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4557 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4558 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4559 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4560 // small; hence the other fields of MorphAddrContext.
4561 enum MorphAddrContextKind
4566 struct MorphAddrContext
4568 MorphAddrContextKind m_kind;
4569 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4570 // top-level indirection and here have been constants.
4571 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4572 // In that case, is the sum of those constant offsets.
4574 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4579 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4580 static MorphAddrContext s_CopyBlockMAC;
4583 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4584 var_types* baseTypeOut,
4586 unsigned* simdSizeOut,
4587 bool ignoreUsedInSIMDIntrinsic = false);
4588 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4589 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4590 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4591 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4593 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4594 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4595 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4597 #endif // FEATURE_SIMD
4598 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4599 GenTreePtr fgMorphCast(GenTreePtr tree);
4600 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4601 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4603 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4606 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4607 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4609 void fgFixupStructReturn(GenTreePtr call);
4610 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4611 bool fgAddrCouldBeNull(GenTreePtr addr);
4612 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4613 bool fgCanFastTailCall(GenTreeCall* call);
4614 void fgMorphTailCall(GenTreeCall* call);
4615 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4616 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4617 fgArgTabEntryPtr argTabEntry,
4619 IL_OFFSETX callILOffset,
4620 GenTreePtr tmpAssignmentInsertionPoint,
4621 GenTreePtr paramAssignmentInsertionPoint);
4622 static int fgEstimateCallStackSize(GenTreeCall* call);
4623 GenTreePtr fgMorphCall(GenTreeCall* call);
4624 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4625 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4627 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4628 static fgWalkPreFn fgFindNonInlineCandidate;
4630 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4631 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4632 void fgAssignSetVarDef(GenTreePtr tree);
4633 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4634 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4635 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4636 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4637 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4638 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4639 void fgMorphUnsafeBlk(GenTreeObj* obj);
4640 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4641 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4642 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4643 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4644 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4645 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4646 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4648 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4649 GenTreePtr fgMorphConst(GenTreePtr tree);
4652 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4655 #if LOCAL_ASSERTION_PROP
4656 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4657 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4659 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4661 GenTreeStmt* fgMorphStmt;
4663 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4664 // used when morphing big offset.
4666 //----------------------- Liveness analysis -------------------------------
4668 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4669 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4671 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4672 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4673 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4675 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4677 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4678 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4680 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4681 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4683 void fgExtendDbgScopes();
4684 void fgExtendDbgLifetimes();
4687 void fgDispDebugScopes();
4690 //-------------------------------------------------------------------------
4692 // The following keeps track of any code we've added for things like array
4693 // range checking or explicit calls to enable GC, and so on.
4698 AddCodeDsc* acdNext;
4699 BasicBlock* acdDstBlk; // block to which we jump
4701 SpecialCodeKind acdKind; // what kind of a special block is this?
4702 unsigned short acdStkLvl;
4706 static unsigned acdHelper(SpecialCodeKind codeKind);
4708 AddCodeDsc* fgAddCodeList;
4710 bool fgRngChkThrowAdded;
4711 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4713 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4715 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4718 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4721 bool fgIsCodeAdded();
4723 bool fgIsThrowHlpBlk(BasicBlock* block);
4724 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4726 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4728 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4729 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4730 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4731 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4732 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4734 #if FEATURE_MULTIREG_RET
4735 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4736 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4737 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4738 #endif // FEATURE_MULTIREG_RET
4740 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4743 static fgWalkPreFn fgDebugCheckInlineCandidates;
4746 void fgPromoteStructs();
4747 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4748 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4749 void fgMarkImplicitByRefArgs();
4750 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4751 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4752 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4753 void fgMarkAddressExposedLocals();
4754 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4756 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4758 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4760 // The given local variable, required to be a struct variable, is being assigned via
4761 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4762 // the variable is not enregistered, and is therefore not promoted independently.
4763 void fgLclFldAssign(unsigned lclNum);
4765 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4766 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4767 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4768 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4771 bool fgPrintInlinedMethods;
4774 bool fgIsBigOffset(size_t offset);
4776 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4777 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4778 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4779 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4780 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4783 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4784 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4788 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4789 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4796 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4799 void optRemoveRangeCheck(
4800 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4801 bool optIsRangeCheckRemovable(GenTreePtr tree);
4804 static fgWalkPreFn optValidRangeCheckIndex;
4805 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4808 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4810 /**************************************************************************
4812 *************************************************************************/
4815 // Do hoisting for all loops.
4816 void optHoistLoopCode();
4818 // To represent sets of VN's that have already been hoisted in outer loops.
4819 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4820 typedef VNToBoolMap VNSet;
4822 struct LoopHoistContext
4825 // The set of variables hoisted in the current loop (or nullptr if there are none).
4826 VNSet* m_pHoistedInCurLoop;
4829 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4830 VNSet m_hoistedInParentLoops;
4831 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4832 // Previous decisions on loop-invariance of value numbers in the current loop.
4833 VNToBoolMap m_curLoopVnInvariantCache;
4835 VNSet* GetHoistedInCurLoop(Compiler* comp)
4837 if (m_pHoistedInCurLoop == nullptr)
4839 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4841 return m_pHoistedInCurLoop;
4844 VNSet* ExtractHoistedInCurLoop()
4846 VNSet* res = m_pHoistedInCurLoop;
4847 m_pHoistedInCurLoop = nullptr;
4851 LoopHoistContext(Compiler* comp)
4852 : m_pHoistedInCurLoop(nullptr)
4853 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4854 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4859 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4860 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4861 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4862 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4864 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4865 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4866 // "m_hoistedInParentLoops".
4868 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4870 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4871 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4872 // expressions to "hoistInLoop".
4873 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4875 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4876 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4878 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4879 // that are invariant in loop "lnum" (an index into the optLoopTable)
4880 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4881 // expressions to "hoistInLoop".
4882 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4883 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4884 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4885 bool optHoistLoopExprsForTree(GenTreePtr tree,
4887 LoopHoistContext* hoistCtxt,
4888 bool* firstBlockAndBeforeSideEffect,
4891 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4892 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4894 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4895 // Constants and init values are always loop invariant.
4896 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4897 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4899 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4900 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4901 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4902 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4903 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4905 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4906 // in the loop table.
4907 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4909 // Records the set of "side effects" of all loops: fields (object instance and static)
4910 // written to, and SZ-array element type equivalence classes updated.
4911 void optComputeLoopSideEffects();
4914 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4915 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4916 // static) written to, and SZ-array element type equivalence classes updated.
4917 void optComputeLoopNestSideEffects(unsigned lnum);
4919 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4920 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4922 // Hoist the expression "expr" out of loop "lnum".
4923 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4926 void optOptimizeBools();
4929 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4931 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4934 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4936 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4937 // the loop into a "do-while" loop
4938 // Also finds all natural loops and records them in the loop table
4940 // Optionally clone loops in the loop table.
4941 void optCloneLoops();
4943 // Clone loop "loopInd" in the loop table.
4944 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4946 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4947 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4948 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4950 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4952 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4955 // This enumeration describes what is killed by a call.
4959 CALLINT_NONE, // no interference (most helpers)
4960 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4961 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4962 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4963 CALLINT_ALL, // kills everything (normal method call)
4967 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4968 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4969 // in bbNext order; we use comparisons on the bbNum to decide order.)
4970 // The blocks that define the body are
4971 // first <= top <= entry <= bottom .
4972 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4973 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4974 // Compiler::optFindNaturalLoops().
4977 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4978 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4979 // loop, but not the outer loop.)
4980 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4982 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4983 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4984 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4986 callInterf lpAsgCall; // "callInterf" for calls in the loop
4987 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4988 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4990 unsigned short lpFlags; // Mask of the LPFLG_* constants
4992 unsigned char lpExitCnt; // number of exits from the loop
4994 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4995 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4996 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4997 // (Actually, an "immediately" nested loop --
4998 // no other child of this loop is a parent of lpChild.)
4999 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5000 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5001 // by following "lpChild" then "lpSibling" links.
5003 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5004 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5006 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5007 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5008 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5010 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5011 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5013 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5014 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5015 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5016 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5018 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5019 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5020 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5022 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5023 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5024 // type are assigned to.
5026 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
5027 // If this is set, the fields below may not be accurate (since they become irrelevant.)
5028 bool lpContainsCall; // True if executing the loop body *may* execute a call
5030 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5031 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5033 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5035 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5036 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5038 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5040 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5041 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5043 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5044 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5046 JitSimplerHashBehavior>
5048 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5049 // instance fields modified
5052 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5053 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5055 JitSimplerHashBehavior>
5057 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5058 // arrays of that type are modified
5061 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5062 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5064 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5065 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5066 // (shifted left, with a low-order bit set to distinguish.)
5067 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5068 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5070 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5072 GenTreePtr lpIterTree; // The "i <op>= const" tree
5073 unsigned lpIterVar(); // iterator variable #
5074 int lpIterConst(); // the constant with which the iterator is incremented
5075 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5076 void VERIFY_lpIterTree();
5078 var_types lpIterOperType(); // For overflow instructions
5081 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5082 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5086 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5088 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5089 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5090 void VERIFY_lpTestTree();
5092 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5093 GenTreePtr lpIterator(); // the iterator node in the loop test
5094 GenTreePtr lpLimit(); // the limit node in the loop test
5096 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5097 // LPFLG_CONST_LIMIT
5098 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5100 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5101 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5102 // LPFLG_ARRLEN_LIMIT
5104 // Returns "true" iff "*this" contains the blk.
5105 bool lpContains(BasicBlock* blk)
5107 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5109 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5110 // to be equal, but requiring bottoms to be different.)
5111 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5113 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5116 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5117 // bottoms to be different.)
5118 bool lpContains(const LoopDsc& lp2)
5120 return lpContains(lp2.lpFirst, lp2.lpBottom);
5123 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5124 // (allowing firsts to be equal, but requiring bottoms to be different.)
5125 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5127 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5130 // Returns "true" iff "*this" is (properly) contained by "lp2"
5131 // (allowing firsts to be equal, but requiring bottoms to be different.)
5132 bool lpContainedBy(const LoopDsc& lp2)
5134 return lpContains(lp2.lpFirst, lp2.lpBottom);
5137 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5138 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5140 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5142 // Returns "true" iff "*this" is disjoint from "lp2".
5143 bool lpDisjoint(const LoopDsc& lp2)
5145 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5147 // Returns "true" iff the loop is well-formed (see code for defn).
5150 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5151 lpEntry->bbNum <= lpBottom->bbNum &&
5152 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5157 bool fgMightHaveLoop(); // returns true if there are any backedges
5158 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5161 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5162 unsigned char optLoopCount; // number of tracked loops
5165 unsigned optCallCount; // number of calls made in the method
5166 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5167 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5168 unsigned optLoopsCloned; // number of loops cloned in the current method.
5171 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5172 void optPrintLoopInfo(unsigned loopNum,
5174 BasicBlock* lpFirst,
5176 BasicBlock* lpEntry,
5177 BasicBlock* lpBottom,
5178 unsigned char lpExitCnt,
5180 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5181 void optPrintLoopInfo(unsigned lnum);
5182 void optPrintLoopRecording(unsigned lnum);
5184 void optCheckPreds();
5187 void optSetBlockWeights();
5189 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5191 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5193 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5195 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5196 unsigned optIsLoopIncrTree(GenTreePtr incr);
5197 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5198 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5199 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5200 bool optExtractInitTestIncr(BasicBlock* head,
5205 GenTreePtr* ppIncr);
5207 void optRecordLoop(BasicBlock* head,
5213 unsigned char exitCnt);
5215 void optFindNaturalLoops();
5217 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5218 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5219 bool optCanonicalizeLoopNest(unsigned char loopInd);
5221 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5222 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5223 bool optCanonicalizeLoop(unsigned char loopInd);
5225 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5226 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5227 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5228 bool optLoopContains(unsigned l1, unsigned l2);
5230 // Requires "loopInd" to be a valid index into the loop table.
5231 // Updates the loop table by changing loop "loopInd", whose head is required
5232 // to be "from", to be "to". Also performs this transformation for any
5233 // loop nested in "loopInd" that shares the same head as "loopInd".
5234 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5236 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5237 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5238 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5240 // Marks the containsCall information to "lnum" and any parent loops.
5241 void AddContainsCallAllContainingLoops(unsigned lnum);
5242 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5243 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5244 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5245 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5246 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5247 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5249 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5250 // of "from".) Copies the jump destination from "from" to "to".
5251 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5253 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5254 unsigned optLoopDepth(unsigned lnum)
5256 unsigned par = optLoopTable[lnum].lpParent;
5257 if (par == BasicBlock::NOT_IN_LOOP)
5263 return 1 + optLoopDepth(par);
5267 void fgOptWhileLoop(BasicBlock* block);
5269 bool optComputeLoopRep(int constInit,
5272 genTreeOps iterOper,
5274 genTreeOps testOper,
5277 unsigned* iterCount);
5278 #if FEATURE_STACK_FP_X87
5281 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5282 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5283 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5284 #endif // FEATURE_STACK_FP_X87
5287 static fgWalkPreFn optIsVarAssgCB;
5290 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5292 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5294 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5296 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5298 /**************************************************************************
5299 * Optimization conditions
5300 *************************************************************************/
5302 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5303 bool optPentium4(void);
5304 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5305 bool optAvoidIntMult(void);
5310 // The following is the upper limit on how many expressions we'll keep track
5311 // of for the CSE analysis.
5313 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5315 static const int MIN_CSE_COST = 2;
5317 // Keeps tracked cse indices
5318 BitVecTraits* cseTraits;
5322 /* Generic list of nodes - used by the CSE logic */
5330 typedef struct treeLst* treeLstPtr;
5334 treeStmtLst* tslNext;
5335 GenTreePtr tslTree; // tree node
5336 GenTreePtr tslStmt; // statement containing the tree
5337 BasicBlock* tslBlock; // block containing the statement
5340 typedef struct treeStmtLst* treeStmtLstPtr;
5342 // The following logic keeps track of expressions via a simple hash table.
5346 CSEdsc* csdNextInBucket; // used by the hash table
5348 unsigned csdHashValue; // the orginal hashkey
5350 unsigned csdIndex; // 1..optCSECandidateCount
5351 char csdLiveAcrossCall; // 0 or 1
5353 unsigned short csdDefCount; // definition count
5354 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5356 unsigned csdDefWtCnt; // weighted def count
5357 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5359 GenTreePtr csdTree; // treenode containing the 1st occurance
5360 GenTreePtr csdStmt; // stmt containing the 1st occurance
5361 BasicBlock* csdBlock; // block containing the 1st occurance
5363 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5364 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5367 static const size_t s_optCSEhashSize;
5368 CSEdsc** optCSEhash;
5373 CSEdsc* optCSEfindDsc(unsigned index);
5374 void optUnmarkCSE(GenTreePtr tree);
5376 // user defined callback data for the tree walk function optCSE_MaskHelper()
5377 struct optCSE_MaskData
5379 EXPSET_TP CSE_defMask;
5380 EXPSET_TP CSE_useMask;
5383 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5384 static fgWalkPreFn optCSE_MaskHelper;
5386 // This function walks all the node for an given tree
5387 // and return the mask of CSE definitions and uses for the tree
5389 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5391 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5392 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5393 bool optCSE_canSwap(GenTree* tree);
5395 static fgWalkPostFn optPropagateNonCSE;
5396 static fgWalkPreFn optHasNonCSEChild;
5398 static fgWalkPreFn optUnmarkCSEs;
5400 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5401 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5403 void optCleanupCSEs();
5406 void optEnsureClearCSEInfo();
5409 #endif // FEATURE_ANYCSE
5411 #if FEATURE_VALNUM_CSE
5412 /**************************************************************************
5413 * Value Number based CSEs
5414 *************************************************************************/
5417 void optOptimizeValnumCSEs();
5420 void optValnumCSE_Init();
5421 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5422 unsigned optValnumCSE_Locate();
5423 void optValnumCSE_InitDataFlow();
5424 void optValnumCSE_DataFlow();
5425 void optValnumCSE_Availablity();
5426 void optValnumCSE_Heuristic();
5427 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5429 #endif // FEATURE_VALNUM_CSE
5432 bool optDoCSE; // True when we have found a duplicate CSE tree
5433 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5434 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5435 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5436 unsigned optCSEstart; // The first local variable number that is a CSE
5437 unsigned optCSEcount; // The total count of CSE's introduced.
5438 unsigned optCSEweight; // The weight of the current block when we are
5439 // scanning for CSE expressions
5441 bool optIsCSEcandidate(GenTreePtr tree);
5443 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5445 bool lclNumIsTrueCSE(unsigned lclNum) const
5447 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5450 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5452 bool lclNumIsCSE(unsigned lclNum) const
5454 return lvaTable[lclNum].lvIsCSE;
5458 bool optConfigDisableCSE();
5459 bool optConfigDisableCSE2();
5461 void optOptimizeCSEs();
5463 #endif // FEATURE_ANYCSE
5471 unsigned ivaVar; // Variable we are interested in, or -1
5472 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5473 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5474 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5475 callInterf ivaMaskCall; // What kind of calls are there?
5478 static callInterf optCallInterf(GenTreePtr call);
5481 // VN based copy propagation.
5482 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5483 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5484 LclNumToGenTreePtrStack;
5486 // Kill set to track variables with intervening definitions.
5487 VARSET_TP optCopyPropKillSet;
5489 // Copy propagation functions.
5490 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5491 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5492 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5493 bool optIsSsaLocal(GenTreePtr tree);
5494 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5495 void optVnCopyProp();
5497 /**************************************************************************
5498 * Early value propagation
5499 *************************************************************************/
5505 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5509 static unsigned GetHashCode(SSAName ssaNm)
5511 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5514 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5516 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5520 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5521 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5522 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5523 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5524 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5526 unsigned optMethodFlags;
5528 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5529 // No throughput diff was found with backward walk bound between 3-8.
5530 static const int optEarlyPropRecurBound = 5;
5532 enum class optPropKind
5540 bool gtIsVtableRef(GenTreePtr tree);
5541 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5542 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5543 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5544 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5545 bool optEarlyPropRewriteTree(GenTreePtr tree);
5546 bool optDoEarlyPropForBlock(BasicBlock* block);
5547 bool optDoEarlyPropForFunc();
5548 void optEarlyProp();
5549 void optFoldNullCheck(GenTreePtr tree);
5550 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5553 /**************************************************************************
5554 * Value/Assertion propagation
5555 *************************************************************************/
5557 // Data structures for assertion prop
5558 BitVecTraits* apTraits;
5562 enum optAssertionKind
5577 O1K_ARRLEN_OPER_BND,
5578 O1K_ARRLEN_LOOP_BND,
5579 O1K_CONSTANT_LOOP_BND,
5600 optAssertionKind assertionKind;
5603 unsigned lclNum; // assigned to or property of this local var number
5611 struct AssertionDscOp1
5613 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5620 struct AssertionDscOp2
5622 optOp2Kind kind; // a const or copy assignment
5626 ssize_t iconVal; // integer
5627 unsigned iconFlags; // gtFlags
5629 struct Range // integer subrange
5643 bool IsArrLenArithBound()
5645 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5647 bool IsArrLenBound()
5649 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5651 bool IsConstantBound()
5653 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5654 op1.kind == O1K_CONSTANT_LOOP_BND);
5656 bool IsBoundsCheckNoThrow()
5658 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5661 bool IsCopyAssertion()
5663 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5666 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5668 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5669 a1->op2.kind == a2->op2.kind;
5672 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5674 if (kind == OAK_EQUAL)
5676 return kind2 == OAK_NOT_EQUAL;
5678 else if (kind == OAK_NOT_EQUAL)
5680 return kind2 == OAK_EQUAL;
5685 static ssize_t GetLowerBoundForIntegralType(var_types type)
5705 static ssize_t GetUpperBoundForIntegralType(var_types type)
5729 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5731 return (op1.kind == that->op1.kind) &&
5732 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5735 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5737 if (op2.kind != that->op2.kind)
5743 case O2K_IND_CNS_INT:
5745 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5747 case O2K_CONST_LONG:
5748 return (op2.lconVal == that->op2.lconVal);
5750 case O2K_CONST_DOUBLE:
5751 // exact match because of positive and negative zero.
5752 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5754 case O2K_LCLVAR_COPY:
5756 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5757 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5760 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5763 // we will return false
5767 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5773 bool Complementary(AssertionDsc* that, bool vnBased)
5775 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5776 HasSameOp2(that, vnBased);
5779 bool Equals(AssertionDsc* that, bool vnBased)
5781 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5785 typedef unsigned short AssertionIndex;
5788 static fgWalkPreFn optAddCopiesCallback;
5789 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5790 unsigned optAddCopyLclNum;
5791 GenTreePtr optAddCopyAsgnNode;
5793 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5794 bool optAssertionPropagated; // set to true if we modified the trees
5795 bool optAssertionPropagatedCurrentStmt;
5797 GenTreePtr optAssertionPropCurrentTree;
5799 AssertionIndex* optComplementaryAssertionMap;
5800 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5801 // using the value of a local var) for each local var
5802 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5803 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5804 AssertionIndex optMaxAssertionCount;
5807 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5808 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5809 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5810 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5811 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5812 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5814 AssertionIndex GetAssertionCount()
5816 return optAssertionCount;
5818 ASSERT_TP* bbJtrueAssertionOut;
5819 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5820 ValueNumToAssertsMap;
5821 ValueNumToAssertsMap* optValueNumToAsserts;
5823 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5825 // Assertion prop helpers.
5826 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5827 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5828 void optAssertionInit(bool isLocalProp);
5829 void optAssertionTraitsInit(AssertionIndex assertionCount);
5830 #if LOCAL_ASSERTION_PROP
5831 void optAssertionReset(AssertionIndex limit);
5832 void optAssertionRemove(AssertionIndex index);
5835 // Assertion prop data flow functions.
5836 void optAssertionPropMain();
5837 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5838 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5839 ASSERT_TP* optInitAssertionDataflowFlags();
5840 ASSERT_TP* optComputeAssertionGen();
5842 // Assertion Gen functions.
5843 void optAssertionGen(GenTreePtr tree);
5844 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5845 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5846 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5847 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5848 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5849 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5851 // Assertion creation functions.
5852 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5853 AssertionIndex optCreateAssertion(GenTreePtr op1,
5855 optAssertionKind assertionKind,
5856 AssertionDsc* assertion);
5857 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5859 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5860 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5861 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5863 void optPrintVnAssertionMapping();
5865 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5867 // Used for respective assertion propagations.
5868 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5869 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5870 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5871 bool optAssertionIsNonNull(GenTreePtr op,
5872 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5874 // Used for Relop propagation.
5875 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5876 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5877 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5879 // Assertion prop for lcl var functions.
5880 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5881 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5883 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5884 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5885 const GenTreePtr tree,
5886 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5887 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5889 // Assertion propagation functions.
5890 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5891 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5892 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5893 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5894 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5895 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5896 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5897 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5898 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5899 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5900 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5901 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5903 // Implied assertion functions.
5904 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5905 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5906 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5907 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5909 ASSERT_VALRET_TP optNewFullAssertSet();
5910 ASSERT_VALRET_TP optNewEmptyAssertSet();
5913 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5914 void optDebugCheckAssertion(AssertionDsc* assertion);
5915 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5917 void optAddCopies();
5918 #endif // ASSERTION_PROP
5920 /**************************************************************************
5922 *************************************************************************/
5925 struct LoopCloneVisitorInfo
5927 LoopCloneContext* context;
5930 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5931 : context(context), loopNum(loopNum), stmt(nullptr)
5936 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5937 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5938 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5939 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5940 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5941 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5942 void optObtainLoopCloningOpts(LoopCloneContext* context);
5943 bool optIsLoopClonable(unsigned loopInd);
5945 bool optCanCloneLoops();
5948 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5950 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5951 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5952 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5953 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5957 void optInsertLoopCloningStress(BasicBlock* head);
5959 #if COUNT_RANGECHECKS
5960 static unsigned optRangeChkRmv;
5961 static unsigned optRangeChkAll;
5970 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5975 RngChkDsc* rcdNextInBucket; // used by the hash table
5977 unsigned short rcdHashValue; // to make matching faster
5978 unsigned short rcdIndex; // 0..optRngChkCount-1
5980 GenTreePtr rcdTree; // the array index tree
5983 unsigned optRngChkCount;
5984 static const size_t optRngChkHashSize;
5986 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5987 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5989 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5992 bool optLoopsMarked;
5995 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5996 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6000 XX Does the register allocation and puts the remaining lclVars on the stack XX
6002 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6003 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6007 #ifndef LEGACY_BACKEND
6012 #else // LEGACY_BACKEND
6017 #endif // LEGACY_BACKEND
6019 #ifdef LEGACY_BACKEND
6021 void raAssignVars(); // register allocation
6022 #endif // LEGACY_BACKEND
6024 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6026 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6028 void raMarkStkVars();
6031 // Some things are used by both LSRA and regpredict allocators.
6033 FrameType rpFrameType;
6034 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6036 #ifdef LEGACY_BACKEND
6037 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6039 #endif // LEGACY_BACKEND
6041 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6043 #if FEATURE_FP_REGALLOC
6044 enum enumConfigRegisterFP
6046 CONFIG_REGISTER_FP_NONE = 0x0,
6047 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6048 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6049 CONFIG_REGISTER_FP_FULL = 0x3,
6051 enumConfigRegisterFP raConfigRegisterFP();
6052 #endif // FEATURE_FP_REGALLOC
6055 regMaskTP raConfigRestrictMaskFP();
6058 #ifndef LEGACY_BACKEND
6059 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6060 #else // LEGACY_BACKEND
6061 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6062 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6063 bool raNewBlocks; // True is we added killing blocks for FPU registers
6064 unsigned rpPasses; // Number of passes made by the register predicter
6065 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6066 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6067 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6068 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6069 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6070 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6071 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6072 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6073 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6074 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6075 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6076 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6078 bool rpRegAllocDone; // Set to true after we have completed register allocation
6080 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6082 void raSetupArgMasks(RegState* r);
6084 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6086 void raDumpVarIntf(); // Dump the variable to variable interference graph
6087 void raDumpRegIntf(); // Dump the variable to register interference graph
6089 void raAdjustVarIntf();
6091 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6093 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6095 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6096 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6098 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6100 static fgWalkPreFn rpMarkRegIntf;
6102 regMaskTP rpPredictAddressMode(
6103 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6105 void rpPredictRefAssign(unsigned lclNum);
6107 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6109 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6111 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6113 void rpPredictRegUse(); // Entry point
6115 unsigned raPredictTreeRegUse(GenTreePtr tree);
6116 unsigned raPredictListRegUse(GenTreePtr list);
6118 void raSetRegVarOrder(var_types regType,
6119 regNumber* customVarOrder,
6120 unsigned* customVarOrderSize,
6122 regMaskTP avoidReg);
6124 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6125 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6126 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6127 void raAddToStkPredict(unsigned val)
6129 unsigned newStkPredict = rpStkPredict + val;
6130 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6131 rpStkPredict = UINT_MAX - 1;
6133 rpStkPredict = newStkPredict;
6137 #if !FEATURE_FP_REGALLOC
6138 void raDispFPlifeInfo();
6142 regMaskTP genReturnRegForTree(GenTreePtr tree);
6143 #endif // LEGACY_BACKEND
6145 /* raIsVarargsStackArg is called by raMaskStkVars and by
6146 lvaSortByRefCount. It identifies the special case
6147 where a varargs function has a parameter passed on the
6148 stack, other than the special varargs handle. Such parameters
6149 require special treatment, because they cannot be tracked
6150 by the GC (their offsets in the stack are not known
6154 bool raIsVarargsStackArg(unsigned lclNum)
6158 LclVarDsc* varDsc = &lvaTable[lclNum];
6160 assert(varDsc->lvIsParam);
6162 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6164 #else // _TARGET_X86_
6168 #endif // _TARGET_X86_
6171 #ifdef LEGACY_BACKEND
6172 // Records the current prediction, if it's better than any previous recorded prediction.
6173 void rpRecordPrediction();
6174 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6175 void rpUseRecordedPredictionIfBetter();
6177 // Data members used in the methods above.
6178 unsigned rpBestRecordedStkPredict;
6179 struct VarRegPrediction
6181 bool m_isEnregistered;
6182 regNumberSmall m_regNum;
6183 regNumberSmall m_otherReg;
6185 VarRegPrediction* rpBestRecordedPrediction;
6186 #endif // LEGACY_BACKEND
6189 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6190 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6194 XX Get to the class and method info from the Execution Engine given XX
6195 XX tokens for the class and method XX
6197 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6198 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6202 /* These are the different addressing modes used to access a local var.
6203 * The JIT has to report the location of the locals back to the EE
6204 * for debugging purposes.
6210 VLT_REG_BYREF, // this type is currently only used for value types on X64
6213 VLT_STK_BYREF, // this type is currently only used for value types on X64
6227 siVarLocType vlType;
6230 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6232 // VLT_REG_BYREF -- the specified register contains the address of the variable
6240 // VLT_STK -- Any 32 bit value which is on the stack
6241 // eg. [ESP+0x20], or [EBP-0x28]
6242 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6243 // eg. mov EAX, [ESP+0x20]; [EAX]
6247 regNumber vlsBaseReg;
6248 NATIVE_OFFSET vlsOffset;
6251 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6260 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6261 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6269 regNumber vlrssBaseReg;
6270 NATIVE_OFFSET vlrssOffset;
6274 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6275 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6281 regNumber vlsrsBaseReg;
6282 NATIVE_OFFSET vlsrsOffset;
6288 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6289 // eg 2 DWords at [ESP+0x10]
6293 regNumber vls2BaseReg;
6294 NATIVE_OFFSET vls2Offset;
6297 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6298 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6305 // VLT_FIXED_VA -- fixed argument of a varargs function.
6306 // The argument location depends on the size of the variable
6307 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6308 // location of the first arg. This argument can then be accessed
6309 // relative to the position of the first arg
6313 unsigned vlfvOffset;
6320 void* rpValue; // pointer to the in-process
6321 // location of the value.
6327 bool vlIsInReg(regNumber reg);
6328 bool vlIsOnStk(regNumber reg, signed offset);
6331 /*************************************************************************/
6336 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6337 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6338 CORINFO_CALLINFO_FLAGS flags,
6339 CORINFO_CALL_INFO* pResult);
6340 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6342 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6343 CORINFO_ACCESS_FLAGS flags,
6344 CORINFO_FIELD_INFO* pResult);
6348 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6350 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6352 bool IsSuperPMIException(unsigned code)
6354 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6356 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6357 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6358 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6359 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6360 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6361 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6362 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6363 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6367 case EXCEPTIONCODE_DebugBreakorAV:
6368 case EXCEPTIONCODE_MC:
6369 case EXCEPTIONCODE_LWM:
6370 case EXCEPTIONCODE_SASM:
6371 case EXCEPTIONCODE_SSYM:
6372 case EXCEPTIONCODE_CALLUTILS:
6373 case EXCEPTIONCODE_TYPEUTILS:
6374 case EXCEPTIONCODE_ASSERT:
6381 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6382 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6384 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6385 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6388 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6389 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6390 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6392 // VOM info, method sigs
6394 void eeGetSig(unsigned sigTok,
6395 CORINFO_MODULE_HANDLE scope,
6396 CORINFO_CONTEXT_HANDLE context,
6397 CORINFO_SIG_INFO* retSig);
6399 void eeGetCallSiteSig(unsigned sigTok,
6400 CORINFO_MODULE_HANDLE scope,
6401 CORINFO_CONTEXT_HANDLE context,
6402 CORINFO_SIG_INFO* retSig);
6404 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6406 // Method entry-points, instrs
6408 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6410 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6412 CORINFO_EE_INFO eeInfo;
6413 bool eeInfoInitialized;
6415 CORINFO_EE_INFO* eeGetEEInfo();
6417 // Gets the offset of a SDArray's first element
6418 unsigned eeGetArrayDataOffset(var_types type);
6419 // Gets the offset of a MDArray's first element
6420 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6422 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6424 // Returns the page size for the target machine as reported by the EE.
6425 inline size_t eeGetPageSize()
6427 #if COR_JIT_EE_VERSION > 460
6428 return eeGetEEInfo()->osPageSize;
6429 #else // COR_JIT_EE_VERSION <= 460
6430 return CORINFO_PAGE_SIZE;
6431 #endif // COR_JIT_EE_VERSION > 460
6434 // Returns the frame size at which we will generate a loop to probe the stack.
6435 inline size_t getVeryLargeFrameSize()
6438 // The looping probe code is 40 bytes, whereas the straight-line probing for
6439 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6440 // or greater, to generate smaller code.
6441 return 2 * eeGetPageSize();
6443 return 3 * eeGetPageSize();
6447 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6449 #if COR_JIT_EE_VERSION > 460
6450 return eeGetEEInfo()->targetAbi == abi;
6452 return CORINFO_DESKTOP_ABI == abi;
6456 inline bool generateCFIUnwindCodes()
6458 #ifdef UNIX_AMD64_ABI
6459 return IsTargetAbi(CORINFO_CORERT_ABI);
6467 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6469 // Debugging support - Line number info
6471 void eeGetStmtOffsets();
6473 unsigned eeBoundariesCount;
6475 struct boundariesDsc
6477 UNATIVE_OFFSET nativeIP;
6479 unsigned sourceReason;
6480 } * eeBoundaries; // Boundaries to report to EE
6481 void eeSetLIcount(unsigned count);
6482 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6486 static void eeDispILOffs(IL_OFFSET offs);
6487 static void eeDispLineInfo(const boundariesDsc* line);
6488 void eeDispLineInfos();
6491 // Debugging support - Local var info
6495 unsigned eeVarsCount;
6497 struct VarResultInfo
6499 UNATIVE_OFFSET startOffset;
6500 UNATIVE_OFFSET endOffset;
6504 void eeSetLVcount(unsigned count);
6505 void eeSetLVinfo(unsigned which,
6506 UNATIVE_OFFSET startOffs,
6507 UNATIVE_OFFSET length,
6512 const siVarLoc& loc);
6516 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6517 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6520 // ICorJitInfo wrappers
6522 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6524 void eeAllocUnwindInfo(BYTE* pHotCode,
6530 CorJitFuncKind funcKind);
6532 void eeSetEHcount(unsigned cEH);
6534 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6536 WORD eeGetRelocTypeHint(void* target);
6538 // ICorStaticInfo wrapper functions
6540 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6542 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6544 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6547 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6548 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6549 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6550 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6552 template <typename ParamType>
6553 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6555 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6558 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6560 // Utility functions
6562 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6565 const wchar_t* eeGetCPString(size_t stringHandle);
6568 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6570 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6571 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6573 static fgWalkPreFn CountSharedStaticHelper;
6574 static bool IsSharedStaticHelper(GenTreePtr tree);
6575 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6577 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6578 // returns true/false if 'field' is a Jit Data offset
6579 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6580 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6581 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6583 /*****************************************************************************/
6588 enum TEMP_USAGE_TYPE
6594 static var_types tmpNormalizeType(var_types type);
6595 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6596 void tmpRlsTemp(TempDsc* temp);
6597 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6600 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6601 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6605 bool tmpAllFree() const;
6608 #ifndef LEGACY_BACKEND
6609 void tmpPreAllocateTemps(var_types type, unsigned count);
6610 #endif // !LEGACY_BACKEND
6613 #ifdef LEGACY_BACKEND
6614 unsigned tmpIntSpillMax; // number of int-sized spill temps
6615 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6616 #endif // LEGACY_BACKEND
6618 unsigned tmpCount; // Number of temps
6619 unsigned tmpSize; // Size of all the temps
6622 // Used by RegSet::rsSpillChk()
6623 unsigned tmpGetCount; // Temps which haven't been released yet
6626 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6628 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6629 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6632 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6633 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6637 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6638 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6642 CodeGenInterface* codeGen;
6644 // The following holds information about instr offsets in terms of generated code.
6648 IPmappingDsc* ipmdNext; // next line# record
6649 IL_OFFSETX ipmdILoffsx; // the instr offset
6650 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6651 bool ipmdIsLabel; // Can this code be a branch label?
6654 // Record the instr offset mapping to the generated code
6656 IPmappingDsc* genIPmappingList;
6657 IPmappingDsc* genIPmappingLast;
6659 // Managed RetVal - A side hash table meant to record the mapping from a
6660 // GT_CALL node to its IL offset. This info is used to emit sequence points
6661 // that can be used by debugger to determine the native offset at which the
6662 // managed RetVal will be available.
6664 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6665 // favor of a side table for two reasons: 1) We need IL offset for only those
6666 // GT_CALL nodes (created during importation) that correspond to an IL call and
6667 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6668 // structure and IL offset is needed only when generating debuggable code. Therefore
6669 // it is desirable to avoid memory size penalty in retail scenarios.
6670 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6671 CallSiteILOffsetTable;
6672 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6674 unsigned genReturnLocal; // Local number for the return value when applicable.
6675 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6677 // The following properties are part of CodeGenContext. Getters are provided here for
6678 // convenience and backward compatibility, but the properties can only be set by invoking
6679 // the setter on CodeGenContext directly.
6681 __declspec(property(get = getEmitter)) emitter* genEmitter;
6682 emitter* getEmitter()
6684 return codeGen->getEmitter();
6687 const bool isFramePointerUsed()
6689 return codeGen->isFramePointerUsed();
6692 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6693 bool getInterruptible()
6695 return codeGen->genInterruptible;
6697 void setInterruptible(bool value)
6699 codeGen->setInterruptible(value);
6703 const bool genDoubleAlign()
6705 return codeGen->doDoubleAlign();
6707 DWORD getCanDoubleAlign();
6708 bool shouldDoubleAlign(unsigned refCntStk,
6710 unsigned refCntWtdReg,
6711 unsigned refCntStkParam,
6712 unsigned refCntWtdStkDbl);
6713 #endif // DOUBLE_ALIGN
6715 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6716 bool getFullPtrRegMap()
6718 return codeGen->genFullPtrRegMap;
6720 void setFullPtrRegMap(bool value)
6722 codeGen->setFullPtrRegMap(value);
6725 // Things that MAY belong either in CodeGen or CodeGenContext
6727 #if FEATURE_EH_FUNCLETS
6728 FuncInfoDsc* compFuncInfos;
6729 unsigned short compCurrFuncIdx;
6730 unsigned short compFuncInfoCount;
6732 unsigned short compFuncCount()
6734 assert(fgFuncletsCreated);
6735 return compFuncInfoCount;
6738 #else // !FEATURE_EH_FUNCLETS
6740 // This is a no-op when there are no funclets!
6741 void genUpdateCurrentFunclet(BasicBlock* block)
6746 FuncInfoDsc compFuncInfoRoot;
6748 static const unsigned compCurrFuncIdx = 0;
6750 unsigned short compFuncCount()
6755 #endif // !FEATURE_EH_FUNCLETS
6757 FuncInfoDsc* funCurrentFunc();
6758 void funSetCurrentFunc(unsigned funcIdx);
6759 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6760 unsigned int funGetFuncIdx(BasicBlock* block);
6764 VARSET_TP compCurLife; // current live variables
6765 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6767 template <bool ForCodeGen>
6768 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6770 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6772 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6775 template <bool ForCodeGen>
6776 void compUpdateLife(GenTreePtr tree);
6778 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6779 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6780 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6781 template <bool ForCodeGen>
6782 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6784 template <bool ForCodeGen>
6785 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6787 // Gets a register mask that represent the kill set for a helper call since
6788 // not all JIT Helper calls follow the standard ABI on the target architecture.
6789 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6791 // Gets a register mask that represent the kill set for a NoGC helper call.
6792 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6795 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6796 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6797 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6798 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6799 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6800 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6801 #endif // _TARGET_ARM_
6803 // 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
6805 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6807 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6808 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6809 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6810 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6811 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6812 // for the tracked var indices of the field vars, as in a live var set).
6813 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6815 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6817 if (m_promotedStructDeathVars == nullptr)
6819 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6821 return m_promotedStructDeathVars;
6825 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6826 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6830 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6831 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6834 #if !defined(__GNUC__)
6835 #pragma region Unwind information
6840 // Infrastructure functions: start/stop/reserve/emit.
6843 void unwindBegProlog();
6844 void unwindEndProlog();
6845 void unwindBegEpilog();
6846 void unwindEndEpilog();
6847 void unwindReserve();
6848 void unwindEmit(void* pHotCode, void* pColdCode);
6851 // Specific unwind information functions: called by code generation to indicate a particular
6852 // prolog or epilog unwindable instruction has been generated.
6855 void unwindPush(regNumber reg);
6856 void unwindAllocStack(unsigned size);
6857 void unwindSetFrameReg(regNumber reg, unsigned offset);
6858 void unwindSaveReg(regNumber reg, unsigned offset);
6860 #if defined(_TARGET_ARM_)
6861 void unwindPushMaskInt(regMaskTP mask);
6862 void unwindPushMaskFloat(regMaskTP mask);
6863 void unwindPopMaskInt(regMaskTP mask);
6864 void unwindPopMaskFloat(regMaskTP mask);
6865 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6866 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6867 // called via unwindPadding().
6868 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6869 // instruction and the current location.
6870 #endif // _TARGET_ARM_
6872 #if defined(_TARGET_ARM64_)
6874 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6875 // instruction and the current location.
6876 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6877 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6878 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6879 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6880 void unwindSaveNext(); // unwind code: save_next
6881 void unwindReturn(regNumber reg); // ret lr
6882 #endif // defined(_TARGET_ARM64_)
6885 // Private "helper" functions for the unwind implementation.
6889 #if FEATURE_EH_FUNCLETS
6890 void unwindGetFuncLocations(FuncInfoDsc* func,
6891 bool getHotSectionData,
6892 /* OUT */ emitLocation** ppStartLoc,
6893 /* OUT */ emitLocation** ppEndLoc);
6894 #endif // FEATURE_EH_FUNCLETS
6896 void unwindReserveFunc(FuncInfoDsc* func);
6897 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6899 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6901 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6902 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6904 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6906 #if defined(_TARGET_AMD64_)
6908 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6910 void unwindBegPrologWindows();
6911 void unwindPushWindows(regNumber reg);
6912 void unwindAllocStackWindows(unsigned size);
6913 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6914 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6916 #ifdef UNIX_AMD64_ABI
6917 void unwindBegPrologCFI();
6918 void unwindPushCFI(regNumber reg);
6919 void unwindAllocStackCFI(unsigned size);
6920 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6921 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6922 int mapRegNumToDwarfReg(regNumber reg);
6923 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6924 #endif // UNIX_AMD64_ABI
6925 #elif defined(_TARGET_ARM_)
6927 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6928 void unwindPushPopMaskFloat(regMaskTP mask);
6929 void unwindSplit(FuncInfoDsc* func);
6931 #endif // _TARGET_ARM_
6933 #if !defined(__GNUC__)
6934 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6938 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6939 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6943 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6944 XX that contains the distinguished, well-known SIMD type definitions). XX
6946 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6947 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6950 // Get highest available instruction set for floating point codegen
6951 InstructionSet getFloatingPointInstructionSet()
6953 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6956 return InstructionSet_AVX;
6961 return InstructionSet_SSE3_4;
6965 assert(canUseSSE2());
6966 return InstructionSet_SSE2;
6968 assert(!"getFPInstructionSet() is not implemented for target arch");
6970 return InstructionSet_NONE;
6974 // Get highest available instruction set for SIMD codegen
6975 InstructionSet getSIMDInstructionSet()
6977 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6978 return getFloatingPointInstructionSet();
6980 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6982 return InstructionSet_NONE;
6988 // Should we support SIMD intrinsics?
6991 // Have we identified any SIMD types?
6992 // This is currently used by struct promotion to avoid getting type information for a struct
6993 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6995 bool _usesSIMDTypes;
6996 bool usesSIMDTypes()
6998 return _usesSIMDTypes;
7000 void setUsesSIMDTypes(bool value)
7002 _usesSIMDTypes = value;
7005 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7006 // that require indexed access to the individual fields of the vector, which is not well supported
7007 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7008 unsigned lvaSIMDInitTempVarNum;
7011 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7012 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7013 CORINFO_CLASS_HANDLE SIMDIntHandle;
7014 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7015 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7016 CORINFO_CLASS_HANDLE SIMDShortHandle;
7017 CORINFO_CLASS_HANDLE SIMDByteHandle;
7018 CORINFO_CLASS_HANDLE SIMDLongHandle;
7019 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7020 CORINFO_CLASS_HANDLE SIMDULongHandle;
7021 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7022 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7023 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7024 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7026 // Get the handle for a SIMD type.
7027 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7029 if (simdBaseType == TYP_FLOAT)
7034 return SIMDVector2Handle;
7036 return SIMDVector3Handle;
7038 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7040 return SIMDVector4Handle;
7049 assert(simdType == getSIMDVectorType());
7050 switch (simdBaseType)
7053 return SIMDFloatHandle;
7055 return SIMDDoubleHandle;
7057 return SIMDIntHandle;
7059 return SIMDUShortHandle;
7061 return SIMDUShortHandle;
7063 return SIMDUByteHandle;
7065 return SIMDShortHandle;
7067 return SIMDByteHandle;
7069 return SIMDLongHandle;
7071 return SIMDUIntHandle;
7073 return SIMDULongHandle;
7075 assert(!"Didn't find a class handle for simdType");
7077 return NO_CLASS_HANDLE;
7081 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7082 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7083 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7085 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7086 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7087 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7088 bool isSIMDTypeLocal(GenTree* tree)
7090 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7093 // Returns true if the type of the tree is a byref of TYP_SIMD
7094 bool isAddrOfSIMDType(GenTree* tree)
7096 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7098 switch (tree->OperGet())
7101 return varTypeIsSIMD(tree->gtGetOp1());
7103 case GT_LCL_VAR_ADDR:
7104 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7107 return isSIMDTypeLocal(tree);
7114 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7116 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7117 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7118 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7121 // Returns base type of a TYP_SIMD local.
7122 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7123 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7125 if (isSIMDTypeLocal(tree))
7127 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7133 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7135 return info.compCompHnd->isInSIMDModule(clsHnd);
7138 bool isSIMDClass(typeInfo* pTypeInfo)
7140 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7143 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7144 // if it is not a SIMD type or is an unsupported base type.
7145 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7147 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7149 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7152 // Get SIMD Intrinsic info given the method handle.
7153 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7154 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7155 CORINFO_METHOD_HANDLE methodHnd,
7156 CORINFO_SIG_INFO* sig,
7159 var_types* baseType,
7160 unsigned* sizeBytes);
7162 // Pops and returns GenTree node from importers type stack.
7163 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7164 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7166 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7167 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7169 // Creates a GT_SIMD tree for Select operation
7170 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7172 unsigned simdVectorSize,
7177 // Creates a GT_SIMD tree for Min/Max operation
7178 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7179 CORINFO_CLASS_HANDLE typeHnd,
7181 unsigned simdVectorSize,
7185 // Transforms operands and returns the SIMD intrinsic to be applied on
7186 // transformed operands to obtain given relop result.
7187 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7188 CORINFO_CLASS_HANDLE typeHnd,
7189 unsigned simdVectorSize,
7190 var_types* baseType,
7194 // Creates a GT_SIMD tree for Abs intrinsic.
7195 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7197 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7198 // Transforms operands and returns the SIMD intrinsic to be applied on
7199 // transformed operands to obtain == comparison result.
7200 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7201 unsigned simdVectorSize,
7205 // Transforms operands and returns the SIMD intrinsic to be applied on
7206 // transformed operands to obtain > comparison result.
7207 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7208 unsigned simdVectorSize,
7212 // Transforms operands and returns the SIMD intrinsic to be applied on
7213 // transformed operands to obtain >= comparison result.
7214 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7215 unsigned simdVectorSize,
7219 // Transforms operands and returns the SIMD intrinsic to be applied on
7220 // transformed operands to obtain >= comparison result in case of int32
7221 // and small int base type vectors.
7222 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7223 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7224 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7226 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7227 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7228 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7229 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7230 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7232 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7233 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7234 GenTreePtr newobjThis,
7235 CORINFO_CLASS_HANDLE clsHnd,
7236 CORINFO_METHOD_HANDLE method,
7237 CORINFO_SIG_INFO* sig,
7240 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7242 // Whether SIMD vector occupies part of SIMD register.
7243 // SSE2: vector2f/3f are considered sub register SIMD types.
7244 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7245 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7247 unsigned sizeBytes = 0;
7248 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7249 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7252 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7254 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7257 // Get the type for the hardware SIMD vector.
7258 // This is the maximum SIMD type supported for this target.
7259 var_types getSIMDVectorType()
7261 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7268 assert(canUseSSE2());
7272 assert(!"getSIMDVectorType() unimplemented on target arch");
7277 // Get the size of the SIMD type in bytes
7278 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7280 unsigned sizeBytes = 0;
7281 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7285 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7286 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7288 // Get the the number of elements of basetype of SIMD vector given by its type handle
7289 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7291 // Get preferred alignment of SIMD type.
7292 int getSIMDTypeAlignment(var_types simdType);
7294 // Get the number of bytes in a SIMD Vector for the current compilation.
7295 unsigned getSIMDVectorRegisterByteLength()
7297 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7300 return YMM_REGSIZE_BYTES;
7304 assert(canUseSSE2());
7305 return XMM_REGSIZE_BYTES;
7308 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7313 // The minimum and maximum possible number of bytes in a SIMD vector.
7314 unsigned int maxSIMDStructBytes()
7316 return getSIMDVectorRegisterByteLength();
7318 unsigned int minSIMDStructBytes()
7320 return emitTypeSize(TYP_SIMD8);
7323 #ifdef FEATURE_AVX_SUPPORT
7324 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7325 static const unsigned maxPossibleSIMDStructBytes = 32;
7326 #else // !FEATURE_AVX_SUPPORT
7327 static const unsigned maxPossibleSIMDStructBytes = 16;
7328 #endif // !FEATURE_AVX_SUPPORT
7330 // Returns the codegen type for a given SIMD size.
7331 var_types getSIMDTypeForSize(unsigned size)
7333 var_types simdType = TYP_UNDEF;
7336 simdType = TYP_SIMD8;
7338 else if (size == 12)
7340 simdType = TYP_SIMD12;
7342 else if (size == 16)
7344 simdType = TYP_SIMD16;
7346 #ifdef FEATURE_AVX_SUPPORT
7347 else if (size == 32)
7349 simdType = TYP_SIMD32;
7351 #endif // FEATURE_AVX_SUPPORT
7354 noway_assert(!"Unexpected size for SIMD type");
7359 unsigned getSIMDInitTempVarNum()
7361 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7363 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7364 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7366 return lvaSIMDInitTempVarNum;
7369 #endif // FEATURE_SIMD
7372 //------------------------------------------------------------------------
7373 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7375 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7376 // candidate for enregistration.
7378 unsigned largestEnregisterableStructSize()
7381 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7382 if (vectorRegSize > TARGET_POINTER_SIZE)
7384 return vectorRegSize;
7387 #endif // FEATURE_SIMD
7389 return TARGET_POINTER_SIZE;
7394 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7395 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7396 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7398 // Is this var is of type simd struct?
7399 bool lclVarIsSIMDType(unsigned varNum)
7401 LclVarDsc* varDsc = lvaTable + varNum;
7402 return varDsc->lvIsSIMDType();
7405 // Is this Local node a SIMD local?
7406 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7408 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7411 // Returns true if the TYP_SIMD locals on stack are aligned at their
7412 // preferred byte boundary specified by getSIMDTypeAlignment().
7414 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7415 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7416 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7417 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7418 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7419 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7420 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7423 bool isSIMDTypeLocalAligned(unsigned varNum)
7425 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7426 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7429 int off = lvaFrameAddress(varNum, &ebpBased);
7430 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7431 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7432 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7435 #endif // FEATURE_SIMD
7440 // Whether SSE2 is available
7441 bool canUseSSE2() const
7443 #ifdef _TARGET_XARCH_
7444 return opts.compCanUseSSE2;
7450 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7451 bool CanUseSSE3_4() const
7453 #ifdef _TARGET_XARCH_
7454 return opts.compCanUseSSE3_4;
7460 bool canUseAVX() const
7462 #ifdef FEATURE_AVX_SUPPORT
7463 return opts.compCanUseAVX;
7470 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7471 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7475 XX Generic info about the compilation and the method being compiled. XX
7476 XX It is responsible for driving the other phases. XX
7477 XX It is also responsible for all the memory management. XX
7479 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7480 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7484 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7486 InlineResult* compInlineResult; // The result of importing the inlinee method.
7488 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7489 bool compJmpOpUsed; // Does the method do a JMP
7490 bool compLongUsed; // Does the method use TYP_LONG
7491 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7492 bool compTailCallUsed; // Does the method do a tailcall
7493 bool compLocallocUsed; // Does the method use localloc.
7494 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7495 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7496 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7498 // NOTE: These values are only reliable after
7499 // the importing is completely finished.
7501 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7502 // we can iterate over these efficiently.
7504 #if CPU_USES_BLOCK_MOVE
7505 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7509 // State information - which phases have completed?
7510 // These are kept together for easy discoverability
7512 bool bRangeAllowStress;
7513 bool compCodeGenDone;
7514 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7515 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7516 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7517 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7520 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7521 bool fgLocalVarLivenessChanged;
7523 bool compStackProbePrologDone;
7525 #ifndef LEGACY_BACKEND
7527 #endif // !LEGACY_BACKEND
7528 bool compRationalIRForm;
7530 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7532 bool compGeneratingProlog;
7533 bool compGeneratingEpilog;
7534 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7535 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7536 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7537 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7538 bool getNeedsGSSecurityCookie() const
7540 return compNeedsGSSecurityCookie;
7542 void setNeedsGSSecurityCookie()
7544 compNeedsGSSecurityCookie = true;
7547 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7548 // frame layout calculations, this is the level we are currently
7551 //---------------------------- JITing options -----------------------------
7564 JitFlags* jitFlags; // all flags passed from the EE
7565 unsigned compFlags; // method attributes
7567 codeOptimize compCodeOpt; // what type of code optimizations
7571 #ifdef _TARGET_XARCH_
7572 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7573 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7575 #ifdef FEATURE_AVX_SUPPORT
7576 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7577 #endif // FEATURE_AVX_SUPPORT
7578 #endif // _TARGET_XARCH_
7580 // optimize maximally and/or favor speed over size?
7582 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7583 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7584 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7585 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7586 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7588 // Maximun number of locals before turning off the inlining
7589 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7592 unsigned instrCount;
7593 unsigned lvRefCount;
7594 bool compMinOptsIsSet;
7596 bool compMinOptsIsUsed;
7598 inline bool MinOpts()
7600 assert(compMinOptsIsSet);
7601 compMinOptsIsUsed = true;
7604 inline bool IsMinOptsSet()
7606 return compMinOptsIsSet;
7609 inline bool MinOpts()
7613 inline bool IsMinOptsSet()
7615 return compMinOptsIsSet;
7618 inline void SetMinOpts(bool val)
7620 assert(!compMinOptsIsUsed);
7621 assert(!compMinOptsIsSet || (compMinOpts == val));
7623 compMinOptsIsSet = true;
7626 // true if the CLFLG_* for an optimization is set.
7627 inline bool OptEnabled(unsigned optFlag)
7629 return !!(compFlags & optFlag);
7632 #ifdef FEATURE_READYTORUN_COMPILER
7633 inline bool IsReadyToRun()
7635 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7638 inline bool IsReadyToRun()
7644 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7645 // PInvoke transitions inline (e.g. when targeting CoreRT).
7646 inline bool ShouldUsePInvokeHelpers()
7648 #if COR_JIT_EE_VERSION > 460
7649 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7655 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7657 inline bool IsReversePInvoke()
7659 #if COR_JIT_EE_VERSION > 460
7660 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7666 // true if we must generate code compatible with JIT32 quirks
7667 inline bool IsJit32Compat()
7669 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7670 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7676 // true if we must generate code compatible with Jit64 quirks
7677 inline bool IsJit64Compat()
7679 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7680 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7681 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7688 bool compScopeInfo; // Generate the LocalVar info ?
7689 bool compDbgCode; // Generate debugger-friendly code?
7690 bool compDbgInfo; // Gather debugging info?
7693 #ifdef PROFILING_SUPPORTED
7694 bool compNoPInvokeInlineCB;
7696 static const bool compNoPInvokeInlineCB;
7700 bool compGcChecks; // Check arguments and return values to ensure they are sane
7701 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7702 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7706 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7707 // to be allocated on the stack.
7708 // It will be set to true in the following cases:
7709 // 1. When the method being compiled has a declarative security
7710 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7711 // This is also the case when we inject a prolog and epilog in the method.
7713 // 2. When the method being compiled has imperative security (i.e. the method
7714 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7716 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7718 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7719 // which gets reported as a GC root to stackwalker.
7720 // (See also ICodeManager::GetAddrOfSecurityObject.)
7727 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7728 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7732 #ifdef UNIX_AMD64_ABI
7733 // This flag is indicating if there is a need to align the frame.
7734 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7735 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7736 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7737 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7738 // there are calls and making sure the frame alignment logic is executed.
7739 bool compNeedToAlignFrame;
7740 #endif // UNIX_AMD64_ABI
7742 bool compProcedureSplitting; // Separate cold code from hot code
7744 bool genFPorder; // Preserve FP order (operations are non-commutative)
7745 bool genFPopt; // Can we do frame-pointer-omission optimization?
7746 bool altJit; // True if we are an altjit and are compiling this method
7749 bool optRepeat; // Repeat optimizer phases k times
7750 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7751 bool dspCode; // Display native code generated
7752 bool dspEHTable; // Display the EH table reported to the VM
7753 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7754 bool dspEmit; // Display emitter output
7755 bool dspLines; // Display source-code lines intermixed with native code output
7756 bool dmpHex; // Display raw bytes in hex of native code output
7757 bool varNames; // Display variables names in native code output
7758 bool disAsm; // Display native code as it is generated
7759 bool disAsmSpilled; // Display native code when any register spilling occurs
7760 bool disDiffable; // Makes the Disassembly code 'diff-able'
7761 bool disAsm2; // Display native code after it is generated using external disassembler
7762 bool dspOrder; // Display names of each of the methods that we ngen/jit
7763 bool dspUnwind; // Display the unwind info output
7764 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7765 bool compLongAddress; // Force using large pseudo instructions for long address
7766 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7767 bool dspGCtbls; // Display the GC tables
7771 bool doLateDisasm; // Run the late disassembler
7772 #endif // LATE_DISASM
7774 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7775 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7776 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7777 static const bool dspGCtbls = true;
7780 // We need stack probes to guarantee that we won't trigger a stack overflow
7781 // when calling unmanaged code until they get a chance to set up a frame, because
7782 // the EE will have no idea where it is.
7784 // We will only be doing this currently for hosted environments. Unfortunately
7785 // we need to take care of stubs, so potentially, we will have to do the probes
7786 // for any call. We have a plan for not needing for stubs though
7787 bool compNeedStackProbes;
7789 #ifdef PROFILING_SUPPORTED
7790 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7791 // This option helps make the JIT behave as if it is running under a profiler.
7792 bool compJitELTHookEnabled;
7793 #endif // PROFILING_SUPPORTED
7795 #if FEATURE_TAILCALL_OPT
7796 // Whether opportunistic or implicit tail call optimization is enabled.
7797 bool compTailCallOpt;
7798 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7799 bool compTailCallLoopOpt;
7803 static const bool compUseSoftFP = true;
7804 #else // !ARM_SOFTFP
7805 static const bool compUseSoftFP = false;
7808 GCPollType compGCPollType;
7812 static bool s_pAltJitExcludeAssembliesListInitialized;
7813 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7818 template <typename T>
7821 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7824 template <typename T>
7827 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7830 static int dspTreeID(GenTree* tree)
7832 return tree->gtTreeID;
7834 static void printTreeID(GenTree* tree)
7836 if (tree == nullptr)
7842 printf("[%06d]", dspTreeID(tree));
7849 #define STRESS_MODES \
7853 /* "Variations" stress areas which we try to mix up with each other. */ \
7854 /* These should not be exhaustively used as they might */ \
7855 /* hide/trivialize other areas */ \
7857 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7858 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7859 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7860 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7861 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7862 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7863 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7864 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7865 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7866 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7867 STRESS_MODE(NULL_OBJECT_CHECK) \
7868 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7869 STRESS_MODE(RANDOM_INLINE) \
7871 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7873 /* "Check" stress areas that can be exhaustively used if we */ \
7874 /* dont care about performance at all */ \
7876 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7877 STRESS_MODE(CHK_FLOW_UPDATE) \
7878 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7880 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7884 #define STRESS_MODE(mode) STRESS_##mode,
7891 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7892 BYTE compActiveStressModes[STRESS_COUNT];
7895 #define MAX_STRESS_WEIGHT 100
7897 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7901 bool compInlineStress()
7903 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7906 bool compRandomInlineStress()
7908 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7913 bool compTailCallStress()
7916 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7922 codeOptimize compCodeOpt()
7925 // Switching between size & speed has measurable throughput impact
7926 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7927 // DEBUG, but should generate identical code between CHK & RET builds,
7928 // so that's not acceptable.
7929 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7930 // Investigate the cause of the throughput regression.
7932 return opts.compCodeOpt;
7934 return BLENDED_CODE;
7938 //--------------------- Info about the procedure --------------------------
7942 COMP_HANDLE compCompHnd;
7943 CORINFO_MODULE_HANDLE compScopeHnd;
7944 CORINFO_CLASS_HANDLE compClassHnd;
7945 CORINFO_METHOD_HANDLE compMethodHnd;
7946 CORINFO_METHOD_INFO* compMethodInfo;
7948 BOOL hasCircularClassConstraints;
7949 BOOL hasCircularMethodConstraints;
7951 #if defined(DEBUG) || defined(LATE_DISASM)
7952 const char* compMethodName;
7953 const char* compClassName;
7954 const char* compFullName;
7955 #endif // defined(DEBUG) || defined(LATE_DISASM)
7957 #if defined(DEBUG) || defined(INLINE_DATA)
7958 // Method hash is logcally const, but computed
7960 mutable unsigned compMethodHashPrivate;
7961 unsigned compMethodHash() const;
7962 #endif // defined(DEBUG) || defined(INLINE_DATA)
7964 #ifdef PSEUDORANDOM_NOP_INSERTION
7965 // things for pseudorandom nop insertion
7966 unsigned compChecksum;
7970 // The following holds the FLG_xxxx flags for the method we're compiling.
7973 // The following holds the class attributes for the method we're compiling.
7974 unsigned compClassAttr;
7976 const BYTE* compCode;
7977 IL_OFFSET compILCodeSize; // The IL code size
7978 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7979 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7980 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7981 // (2) the code is hot/cold split, and we issued less code than we expected
7982 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7984 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7985 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7986 bool compIsContextful : 1; // contextful method
7987 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7988 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7989 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7990 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7991 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7993 var_types compRetType; // Return type of the method as declared in IL
7994 var_types compRetNativeType; // Normalized return type as per target arch ABI
7995 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7996 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7997 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7998 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7999 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8000 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8001 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8002 unsigned compMaxStack;
8003 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8004 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8006 unsigned compCallUnmanaged; // count of unmanaged calls
8007 unsigned compLvFrameListRoot; // lclNum for the Frame root
8008 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8009 // You should generally use compHndBBtabCount instead: it is the
8010 // current number of EH clauses (after additions like synchronized
8011 // methods and funclets, and removals like unreachable code deletion).
8013 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8014 // and the VM expects that, or the JIT is a "self-host" compiler
8015 // (e.g., x86 hosted targeting x86) and the VM expects that.
8017 /* The following holds IL scope information about local variables.
8020 unsigned compVarScopesCount;
8021 VarScopeDsc* compVarScopes;
8023 /* The following holds information about instr offsets for
8024 * which we need to report IP-mappings
8027 IL_OFFSET* compStmtOffsets; // sorted
8028 unsigned compStmtOffsetsCount;
8029 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8031 #define CPU_X86 0x0100 // The generic X86 CPU
8032 #define CPU_X86_PENTIUM_4 0x0110
8034 #define CPU_X64 0x0200 // The generic x64 CPU
8035 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8036 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8038 #define CPU_ARM 0x0300 // The generic ARM CPU
8040 unsigned genCPU; // What CPU are we running on
8043 // Returns true if the method being compiled returns a non-void and non-struct value.
8044 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8045 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8046 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8047 // Methods returning such structs are considered to return non-struct return value and
8048 // this method returns true in that case.
8049 bool compMethodReturnsNativeScalarType()
8051 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8054 // Returns true if the method being compiled returns RetBuf addr as its return value
8055 bool compMethodReturnsRetBufAddr()
8057 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8058 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8060 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8061 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8062 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8063 // methods with hidden RetBufArg.
8065 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8066 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8067 // returning the address of RetBuf.
8069 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8070 // to be returned in RAX.
8071 CLANG_FORMAT_COMMENT_ANCHOR;
8073 #ifdef _TARGET_AMD64_
8074 return (info.compRetBuffArg != BAD_VAR_NUM);
8075 #else // !_TARGET_AMD64_
8076 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8077 #endif // !_TARGET_AMD64_
8080 // Returns true if the method returns a value in more than one return register
8081 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8082 // TODO-ARM64: Does this apply for ARM64 too?
8083 bool compMethodReturnsMultiRegRetType()
8085 #if FEATURE_MULTIREG_RET
8086 #if defined(_TARGET_X86_)
8087 // On x86 only 64-bit longs are returned in multiple registers
8088 return varTypeIsLong(info.compRetNativeType);
8089 #else // targets: X64-UNIX, ARM64 or ARM32
8090 // On all other targets that support multireg return values:
8091 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8092 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8093 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8094 #endif // TARGET_XXX
8096 #else // not FEATURE_MULTIREG_RET
8098 // For this architecture there are no multireg returns
8101 #endif // FEATURE_MULTIREG_RET
8104 #if FEATURE_MULTIREG_ARGS
8105 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8106 // return the gcPtr layout for the pointers sized fields
8107 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8108 #endif // FEATURE_MULTIREG_ARGS
8110 // Returns true if the method being compiled returns a value
8111 bool compMethodHasRetVal()
8113 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8114 compMethodReturnsMultiRegRetType();
8119 void compDispLocalVars();
8123 //-------------------------- Global Compiler Data ------------------------------------
8126 static unsigned s_compMethodsCount; // to produce unique label names
8127 unsigned compGenTreeID;
8130 BasicBlock* compCurBB; // the current basic block in process
8131 GenTreePtr compCurStmt; // the current statement in process
8133 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8136 // The following is used to create the 'method JIT info' block.
8137 size_t compInfoBlkSize;
8138 BYTE* compInfoBlkAddr;
8140 EHblkDsc* compHndBBtab; // array of EH data
8141 unsigned compHndBBtabCount; // element count of used elements in EH data array
8142 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8144 #if defined(_TARGET_X86_)
8146 //-------------------------------------------------------------------------
8147 // Tracking of region covered by the monitor in synchronized methods
8148 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8149 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8151 #endif // !_TARGET_X86_
8153 Phases previousCompletedPhase; // the most recently completed phase
8155 //-------------------------------------------------------------------------
8156 // The following keeps track of how many bytes of local frame space we've
8157 // grabbed so far in the current function, and how many argument bytes we
8158 // need to pop when we return.
8161 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8163 // Count of callee-saved regs we pushed in the prolog.
8164 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8165 // In case of Amd64 this doesn't include float regs saved on stack.
8166 unsigned compCalleeRegsPushed;
8168 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8169 // Mask of callee saved float regs on stack.
8170 regMaskTP compCalleeFPRegsSavedMask;
8172 #ifdef _TARGET_AMD64_
8173 // Quirk for VS debug-launch scenario to work:
8174 // Bytes of padding between save-reg area and locals.
8175 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8176 unsigned compVSQuirkStackPaddingNeeded;
8177 bool compQuirkForPPPflag;
8180 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8182 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8183 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8184 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8186 //-------------------------------------------------------------------------
8188 static void compStartup(); // One-time initialization
8189 static void compShutdown(); // One-time finalization
8191 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8194 static void compDisplayStaticSizes(FILE* fout);
8196 //------------ Some utility functions --------------
8198 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8199 void** ppIndirection); /* OUT */
8201 // Several JIT/EE interface functions return a CorInfoType, and also return a
8202 // class handle as an out parameter if the type is a value class. Returns the
8203 // size of the type these describe.
8204 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8207 // Components used by the compiler may write unit test suites, and
8208 // have them run within this method. They will be run only once per process, and only
8209 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8210 // These should fail by asserting.
8211 void compDoComponentUnitTestsOnce();
8214 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8215 CORINFO_MODULE_HANDLE classPtr,
8216 COMP_HANDLE compHnd,
8217 CORINFO_METHOD_INFO* methodInfo,
8218 void** methodCodePtr,
8219 ULONG* methodCodeSize,
8220 JitFlags* compileFlags);
8221 void compCompileFinish();
8222 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8223 COMP_HANDLE compHnd,
8224 CORINFO_METHOD_INFO* methodInfo,
8225 void** methodCodePtr,
8226 ULONG* methodCodeSize,
8227 JitFlags* compileFlags,
8228 CorInfoInstantiationVerification instVerInfo);
8230 ArenaAllocator* compGetAllocator();
8232 #if MEASURE_MEM_ALLOC
8234 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8238 unsigned allocCnt; // # of allocs
8239 UINT64 allocSz; // total size of those alloc.
8240 UINT64 allocSzMax; // Maximum single allocation.
8241 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8242 UINT64 nraTotalSizeAlloc;
8243 UINT64 nraTotalSizeUsed;
8245 static const char* s_CompMemKindNames[]; // Names of the kinds.
8247 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8249 for (int i = 0; i < CMK_Count; i++)
8251 allocSzByKind[i] = 0;
8254 MemStats(const MemStats& ms)
8255 : allocCnt(ms.allocCnt)
8256 , allocSz(ms.allocSz)
8257 , allocSzMax(ms.allocSzMax)
8258 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8259 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8261 for (int i = 0; i < CMK_Count; i++)
8263 allocSzByKind[i] = ms.allocSzByKind[i];
8267 // Until we have ubiquitous constructors.
8270 this->MemStats::MemStats();
8273 void AddAlloc(size_t sz, CompMemKind cmk)
8277 if (sz > allocSzMax)
8281 allocSzByKind[cmk] += sz;
8284 void Print(FILE* f); // Print these stats to f.
8285 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8287 MemStats genMemStats;
8289 struct AggregateMemStats : public MemStats
8293 AggregateMemStats() : MemStats(), nMethods(0)
8297 void Add(const MemStats& ms)
8300 allocCnt += ms.allocCnt;
8301 allocSz += ms.allocSz;
8302 allocSzMax = max(allocSzMax, ms.allocSzMax);
8303 for (int i = 0; i < CMK_Count; i++)
8305 allocSzByKind[i] += ms.allocSzByKind[i];
8307 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8308 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8311 void Print(FILE* f); // Print these stats to jitstdout.
8314 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8315 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8316 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8318 #endif // MEASURE_MEM_ALLOC
8320 #if LOOP_HOIST_STATS
8321 unsigned m_loopsConsidered;
8322 bool m_curLoopHasHoistedExpression;
8323 unsigned m_loopsWithHoistedExpressions;
8324 unsigned m_totalHoistedExpressions;
8326 void AddLoopHoistStats();
8327 void PrintPerMethodLoopHoistStats();
8329 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8330 static unsigned s_loopsConsidered;
8331 static unsigned s_loopsWithHoistedExpressions;
8332 static unsigned s_totalHoistedExpressions;
8334 static void PrintAggregateLoopHoistStats(FILE* f);
8335 #endif // LOOP_HOIST_STATS
8337 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8338 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8339 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8340 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8341 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8342 void compFreeMem(void*);
8344 bool compIsForImportOnly();
8345 bool compIsForInlining();
8346 bool compDonotInline();
8349 const char* compLocalVarName(unsigned varNum, unsigned offs);
8350 VarName compVarName(regNumber reg, bool isFloatReg = false);
8351 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8352 const char* compRegPairName(regPairNo regPair);
8353 const char* compRegNameForSize(regNumber reg, size_t size);
8354 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8355 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8356 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8359 //-------------------------------------------------------------------------
8361 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8363 struct VarScopeMapInfo
8365 VarScopeListNode* head;
8366 VarScopeListNode* tail;
8367 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8369 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8376 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8377 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8379 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8380 VarNumToScopeDscMap;
8382 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8383 VarNumToScopeDscMap* compVarScopeMap;
8385 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8387 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8389 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8391 void compInitVarScopeMap();
8393 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8394 // enter scope, sorted by instr offset
8395 unsigned compNextEnterScope;
8397 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8398 // go out of scope, sorted by instr offset
8399 unsigned compNextExitScope;
8401 void compInitScopeLists();
8403 void compResetScopeLists();
8405 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8407 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8409 void compProcessScopesUntil(unsigned offset,
8411 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8412 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8415 void compDispScopeLists();
8418 bool compIsProfilerHookNeeded();
8420 //-------------------------------------------------------------------------
8421 /* Statistical Data Gathering */
8423 void compJitStats(); // call this function and enable
8424 // various ifdef's below for statistical data
8427 void compCallArgStats();
8428 static void compDispCallArgStats(FILE* fout);
8431 //-------------------------------------------------------------------------
8438 ArenaAllocator* compAllocator;
8441 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8442 // suitable for use by utilcode collection types.
8443 IAllocator* compAsIAllocator;
8445 #if MEASURE_MEM_ALLOC
8446 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8447 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8448 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8450 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8452 #endif // MEASURE_MEM_ALLOC
8454 void compFunctionTraceStart();
8455 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8458 size_t compMaxUncheckedOffsetForNullObject;
8460 void compInitOptions(JitFlags* compileFlags);
8462 void compSetProcessor();
8463 void compInitDebuggingInfo();
8464 void compSetOptimizationLevel();
8465 #ifdef _TARGET_ARMARCH_
8466 bool compRsvdRegCheck(FrameLayoutState curState);
8468 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8470 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8471 void ResetOptAnnotations();
8473 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8474 void RecomputeLoopInfo();
8476 #ifdef PROFILING_SUPPORTED
8477 // Data required for generating profiler Enter/Leave/TailCall hooks
8479 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8480 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8481 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8484 #ifdef _TARGET_AMD64_
8485 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8488 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8489 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8491 IAllocator* getAllocator()
8493 return compAsIAllocator;
8496 #if MEASURE_MEM_ALLOC
8497 IAllocator* getAllocatorBitset()
8499 return compAsIAllocatorBitset;
8501 IAllocator* getAllocatorGC()
8503 return compAsIAllocatorGC;
8505 IAllocator* getAllocatorLoopHoist()
8507 return compAsIAllocatorLoopHoist;
8509 #else // !MEASURE_MEM_ALLOC
8510 IAllocator* getAllocatorBitset()
8512 return compAsIAllocator;
8514 IAllocator* getAllocatorGC()
8516 return compAsIAllocator;
8518 IAllocator* getAllocatorLoopHoist()
8520 return compAsIAllocator;
8522 #endif // !MEASURE_MEM_ALLOC
8525 IAllocator* getAllocatorDebugOnly()
8527 #if MEASURE_MEM_ALLOC
8528 return compAsIAllocatorDebugOnly;
8529 #else // !MEASURE_MEM_ALLOC
8530 return compAsIAllocator;
8531 #endif // !MEASURE_MEM_ALLOC
8536 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8537 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8541 XX Checks for type compatibility and merges types XX
8543 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8544 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8548 // Set to TRUE if verification cannot be skipped for this method
8549 // If we detect unverifiable code, we will lazily check
8550 // canSkipMethodVerification() to see if verification is REALLY needed.
8551 BOOL tiVerificationNeeded;
8553 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8554 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8555 BOOL tiIsVerifiableCode;
8557 // Set to TRUE if runtime callout is needed for this method
8558 BOOL tiRuntimeCalloutNeeded;
8560 // Set to TRUE if security prolog/epilog callout is needed for this method
8561 // Note: This flag is different than compNeedSecurityCheck.
8562 // compNeedSecurityCheck means whether or not a security object needs
8563 // to be allocated on the stack, which is currently true for EnC as well.
8564 // tiSecurityCalloutNeeded means whether or not security callouts need
8565 // to be inserted in the jitted code.
8566 BOOL tiSecurityCalloutNeeded;
8568 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8569 // This support is necessary to suport attributes that are not described in
8570 // for example, signatures. For example, the permanent home byref (byref that
8571 // points to the gc heap), isn't a property of method signatures, therefore,
8572 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8573 // but when deciding if we need to reimport a block, we need to take these
8575 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8577 // Returns TRUE if child is equal to or a subtype of parent.
8578 // normalisedForStack indicates that both types are normalised for the stack
8579 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8581 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8582 // *pDest is modified to represent the merged type. Sets "*changed" to true
8583 // if this changes "*pDest".
8584 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8586 // Set pDest from the primitive value type.
8587 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8589 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8592 // <BUGNUM> VSW 471305
8593 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8594 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8595 // We use a "short" as we need to push/pop this scope.
8597 short compRegSetCheckLevel;
8601 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8602 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8604 XX IL verification stuff XX
8607 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8608 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8612 // The following is used to track liveness of local variables, initialization
8613 // of valueclass constructors, and type safe use of IL instructions.
8615 // dynamic state info needed for verification
8616 EntryState verCurrentState;
8618 // this ptr of object type .ctors are considered intited only after
8619 // the base class ctor is called, or an alternate ctor is called.
8620 // An uninited this ptr can be used to access fields, but cannot
8621 // be used to call a member function.
8622 BOOL verTrackObjCtorInitState;
8624 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8626 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8627 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8628 void verInitCurrentState();
8629 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8631 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8632 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8633 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8635 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8636 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8637 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8638 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8639 typeInfo verMakeTypeInfo(CorInfoType ciType,
8640 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8641 BOOL verIsSDArray(typeInfo ti);
8642 typeInfo verGetArrayElemType(typeInfo ti);
8644 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8645 BOOL verNeedsVerification();
8646 BOOL verIsByRefLike(const typeInfo& ti);
8647 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8649 // generic type variables range over types that satisfy IsBoxable
8650 BOOL verIsBoxable(const typeInfo& ti);
8652 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8653 DEBUGARG(unsigned line));
8654 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8655 DEBUGARG(unsigned line));
8656 bool verCheckTailCallConstraint(OPCODE opcode,
8657 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8658 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8659 // on a type parameter?
8660 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8661 // return false to the caller.
8662 // If false, it will throw.
8664 bool verIsBoxedValueType(typeInfo ti);
8666 void verVerifyCall(OPCODE opcode,
8667 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8668 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8670 bool readonlyCall, // is this a "readonly." call?
8671 const BYTE* delegateCreateStart,
8672 const BYTE* codeAddr,
8673 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8675 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8677 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8678 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8679 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8680 const CORINFO_FIELD_INFO& fieldInfo,
8681 const typeInfo* tiThis,
8683 BOOL allowPlainStructAsThis = FALSE);
8684 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8685 void verVerifyThisPtrInitialised();
8686 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8688 // Register allocator
8689 void raInitStackFP();
8690 void raEnregisterVarsPrePassStackFP();
8691 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8692 void raEnregisterVarsPostPassStackFP();
8693 void raGenerateFPRefCounts();
8694 void raEnregisterVarsStackFP();
8695 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8697 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8698 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8700 // returns true if enregistering v1 would save more mem accesses than v2
8701 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8704 void raDumpHeightsStackFP();
8705 void raDumpVariableRegIntfFloat();
8708 #if FEATURE_STACK_FP_X87
8710 // Currently, we use FP transition blocks in only 2 situations:
8712 // -conditional jump on longs where FP stack differs with target: it's not strictly
8713 // necessary, but its low frequency and the code would get complicated if we try to
8714 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8715 // minimize the way we generate the jump code.
8716 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8717 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8719 // However, transition blocks have 2 problems
8721 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8722 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8723 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8724 // in the right place without preordering them), this causes us to have to generate the transition
8725 // blocks in the cold area if we want procedure splitting.
8728 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8729 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8730 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8731 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8732 // a big change in the exception.
8734 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8735 // optimizations. For these 2 cases:
8737 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8738 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8739 // a switch statement.
8741 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8742 // current procedure splitting and exception code have.
8743 bool compMayHaveTransitionBlocks;
8745 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8747 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8749 unsigned raCntStkStackFP;
8750 unsigned raCntWtdStkDblStackFP;
8751 unsigned raCntStkParamDblStackFP;
8753 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8754 // TODO: Do we want to put this in LclVarDsc?
8755 unsigned raPayloadStackFP[lclMAX_TRACKED];
8756 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8758 // Useful for debugging
8759 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8761 #endif // FEATURE_STACK_FP_X87
8764 // One line log function. Default level is 0. Increasing it gives you
8765 // more log information
8767 // levels are currently unused: #define JITDUMP(level,...) ();
8768 void JitLogEE(unsigned level, const char* fmt, ...);
8770 bool compDebugBreak;
8772 bool compJitHaltMethod();
8777 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8778 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8780 XX GS Security checks for unsafe buffers XX
8782 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8783 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8786 struct ShadowParamVarInfo
8788 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8789 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8791 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8793 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8794 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8795 // slots and update all trees to refer to shadow slots is done immediately after
8796 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8797 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8798 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8799 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8800 // creating a shadow slot even though this routine returns true.
8802 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8803 // required. There are two cases under which a reg arg could potentially be used from its
8805 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8806 // b) LSRA spills it
8808 // Possible solution to address case (a)
8809 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8810 // in this routine. Note that live out of exception handler is something we may not be
8811 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8812 // Therefore, for methods with exception handling and need GS cookie check we might have
8813 // to take conservative approach.
8815 // Possible solution to address case (b)
8816 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8817 // create a new spill temp if the method needs GS cookie check.
8818 return varDsc->lvIsParam;
8819 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8820 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8827 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8832 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8833 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8834 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8836 void gsGSChecksInitCookie(); // Grabs cookie variable
8837 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8838 bool gsFindVulnerableParams(); // Shadow param analysis code
8839 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8841 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8842 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8844 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8845 // This can be overwritten by setting complus_JITInlineSize env variable.
8847 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8850 #ifdef FEATURE_JIT_METHOD_PERF
8851 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8852 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8854 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8855 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8857 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8859 #if MEASURE_CLRAPI_CALLS
8860 // Thin wrappers that call into JitTimer (if present).
8861 inline void CLRApiCallEnter(unsigned apix);
8862 inline void CLRApiCallLeave(unsigned apix);
8865 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8866 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8871 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8872 // These variables are associated with maintaining SQM data about compile time.
8873 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8874 // in the current compilation.
8875 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8876 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8877 // the inlining phase in the current compilation.
8878 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8880 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8881 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8882 // type-loading and class initialization).
8883 void RecordStateAtEndOfInlining();
8884 // Assumes being called at the end of compilation. Update the SQM state.
8885 void RecordStateAtEndOfCompilation();
8887 #ifdef FEATURE_CLRSQM
8888 // Does anything SQM related necessary at process shutdown time.
8889 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8890 #endif // FEATURE_CLRSQM
8893 #if FUNC_INFO_LOGGING
8894 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8895 // filename to write it to.
8896 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8897 #endif // FUNC_INFO_LOGGING
8899 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8901 // Is the compilation in a full trust context?
8902 bool compIsFullTrust();
8904 #ifndef FEATURE_TRACELOGGING
8905 // Should we actually fire the noway assert body and the exception handler?
8906 bool compShouldThrowOnNoway();
8907 #else // FEATURE_TRACELOGGING
8908 // Should we actually fire the noway assert body and the exception handler?
8909 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8911 // Telemetry instance to use per method compilation.
8912 JitTelemetry compJitTelemetry;
8914 // Get common parameters that have to be logged with most telemetry data.
8915 void compGetTelemetryDefaults(const char** assemblyName,
8916 const char** scopeName,
8917 const char** methodName,
8918 unsigned* methodHash);
8919 #endif // !FEATURE_TRACELOGGING
8923 NodeToTestDataMap* m_nodeTestData;
8925 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8926 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8927 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8928 // Current kept in this.
8930 NodeToTestDataMap* GetNodeTestData()
8932 Compiler* compRoot = impInlineRoot();
8933 if (compRoot->m_nodeTestData == nullptr)
8935 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8937 return compRoot->m_nodeTestData;
8940 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8942 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8943 // currently occur in the AST graph.
8944 NodeToIntMap* FindReachableNodesInNodeTestData();
8946 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8947 // test data, associate that data with "to".
8948 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8950 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8951 // have annotations, attach similar annotations to the corresponding nodes in "to".
8952 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8954 // These are the methods that test that the various conditions implied by the
8955 // test attributes are satisfied.
8956 void JitTestCheckSSA(); // SSA builder tests.
8957 void JitTestCheckVN(); // Value numbering tests.
8960 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8962 FieldSeqStore* m_fieldSeqStore;
8964 FieldSeqStore* GetFieldSeqStore()
8966 Compiler* compRoot = impInlineRoot();
8967 if (compRoot->m_fieldSeqStore == nullptr)
8969 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8970 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8971 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8973 return compRoot->m_fieldSeqStore;
8976 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8978 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8979 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8980 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8981 // attach the field sequence directly to the address node.
8982 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8984 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8986 // Don't need to worry about inlining here
8987 if (m_zeroOffsetFieldMap == nullptr)
8989 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8991 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8992 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8994 return m_zeroOffsetFieldMap;
8997 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8998 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8999 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9000 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9001 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9002 // record the the field sequence using the ZeroOffsetFieldMap described above.
9004 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9005 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9006 // CoreRT. Such case is handled same as the default case.
9007 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9009 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9011 NodeToArrayInfoMap* m_arrayInfoMap;
9013 NodeToArrayInfoMap* GetArrayInfoMap()
9015 Compiler* compRoot = impInlineRoot();
9016 if (compRoot->m_arrayInfoMap == nullptr)
9018 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9019 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9020 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9022 return compRoot->m_arrayInfoMap;
9025 NodeToUnsignedMap* m_heapSsaMap;
9027 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
9028 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
9029 // all the possible heap states are possible initial states of the corresponding catch block(s).)
9030 NodeToUnsignedMap* GetHeapSsaMap()
9032 Compiler* compRoot = impInlineRoot();
9033 if (compRoot->m_heapSsaMap == nullptr)
9035 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9036 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9037 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
9039 return compRoot->m_heapSsaMap;
9042 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9043 CORINFO_CLASS_HANDLE m_refAnyClass;
9044 CORINFO_FIELD_HANDLE GetRefanyDataField()
9046 if (m_refAnyClass == nullptr)
9048 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9050 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9052 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9054 if (m_refAnyClass == nullptr)
9056 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9058 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9062 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9064 #if ALLVARSET_COUNTOPS
9065 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9068 static HelperCallProperties s_helperCallProperties;
9070 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9071 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9072 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9074 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9077 unsigned __int8* offset0,
9078 unsigned __int8* offset1);
9079 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9080 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9082 void fgMorphMultiregStructArgs(GenTreeCall* call);
9083 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9085 }; // end of class Compiler
9087 // Inline methods of CompAllocator.
9088 void* CompAllocator::Alloc(size_t sz)
9090 #if MEASURE_MEM_ALLOC
9091 return m_comp->compGetMem(sz, m_cmk);
9093 return m_comp->compGetMem(sz);
9097 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9099 #if MEASURE_MEM_ALLOC
9100 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9102 return m_comp->compGetMemArray(elems, elemSize);
9106 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9107 inline LclVarDsc::LclVarDsc(Compiler* comp)
9108 : // Initialize the ArgRegs to REG_STK.
9109 // The morph will do the right thing to change
9110 // to the right register if passed in register.
9113 #if FEATURE_MULTIREG_ARGS
9114 _lvOtherArgReg(REG_STK)
9116 #endif // FEATURE_MULTIREG_ARGS
9118 lvRefBlks(BlockSetOps::UninitVal())
9120 #endif // ASSERTION_PROP
9121 lvPerSsaData(comp->getAllocator())
9126 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9127 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9129 XX Miscellaneous Compiler stuff XX
9131 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9132 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9135 // Values used to mark the types a stack slot is used for
9137 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9138 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9139 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9140 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9141 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9142 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9143 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9144 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9146 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9148 /*****************************************************************************
9150 * Variables to keep track of total code amounts.
9155 extern size_t grossVMsize;
9156 extern size_t grossNCsize;
9157 extern size_t totalNCsize;
9159 extern unsigned genMethodICnt;
9160 extern unsigned genMethodNCnt;
9161 extern size_t gcHeaderISize;
9162 extern size_t gcPtrMapISize;
9163 extern size_t gcHeaderNSize;
9164 extern size_t gcPtrMapNSize;
9166 #endif // DISPLAY_SIZES
9168 /*****************************************************************************
9170 * Variables to keep track of basic block counts (more data on 1 BB methods)
9173 #if COUNT_BASIC_BLOCKS
9174 extern Histogram bbCntTable;
9175 extern Histogram bbOneBBSizeTable;
9178 /*****************************************************************************
9180 * Used by optFindNaturalLoops to gather statistical information such as
9181 * - total number of natural loops
9182 * - number of loops with 1, 2, ... exit conditions
9183 * - number of loops that have an iterator (for like)
9184 * - number of loops that have a constant iterator
9189 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9190 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9191 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9192 extern unsigned totalLoopCount; // counts the total number of natural loops
9193 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9194 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9195 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9196 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9198 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9199 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9200 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9201 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9202 extern Histogram loopCountTable; // Histogram of loop counts
9203 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9205 #endif // COUNT_LOOPS
9207 /*****************************************************************************
9208 * variables to keep track of how many iterations we go in a dataflow pass
9213 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9214 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9216 #endif // DATAFLOW_ITER
9218 #if MEASURE_BLOCK_SIZE
9219 extern size_t genFlowNodeSize;
9220 extern size_t genFlowNodeCnt;
9221 #endif // MEASURE_BLOCK_SIZE
9223 #if MEASURE_NODE_SIZE
9224 struct NodeSizeStats
9229 genTreeNodeSize = 0;
9230 genTreeNodeActualSize = 0;
9233 size_t genTreeNodeCnt;
9234 size_t genTreeNodeSize; // The size we allocate
9235 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9236 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9237 // a smaller node to a larger one. TODO-Cleanup: add stats on
9238 // SetOper()/ChangeOper() usage to quanitfy this.
9240 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9241 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9242 extern Histogram genTreeNcntHist;
9243 extern Histogram genTreeNsizHist;
9244 #endif // MEASURE_NODE_SIZE
9246 /*****************************************************************************
9247 * Count fatal errors (including noway_asserts).
9251 extern unsigned fatal_badCode;
9252 extern unsigned fatal_noWay;
9253 extern unsigned fatal_NOMEM;
9254 extern unsigned fatal_noWayAssertBody;
9256 extern unsigned fatal_noWayAssertBodyArgs;
9258 extern unsigned fatal_NYI;
9259 #endif // MEASURE_FATAL
9261 /*****************************************************************************
9265 #ifdef _TARGET_XARCH_
9267 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9268 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9269 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9271 const instruction INS_AND = INS_and;
9272 const instruction INS_OR = INS_or;
9273 const instruction INS_XOR = INS_xor;
9274 const instruction INS_NEG = INS_neg;
9275 const instruction INS_TEST = INS_test;
9276 const instruction INS_MUL = INS_imul;
9277 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9278 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9279 const instruction INS_BREAKPOINT = INS_int3;
9280 const instruction INS_ADDC = INS_adc;
9281 const instruction INS_SUBC = INS_sbb;
9282 const instruction INS_NOT = INS_not;
9288 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9289 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9290 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9292 const instruction INS_AND = INS_and;
9293 const instruction INS_OR = INS_orr;
9294 const instruction INS_XOR = INS_eor;
9295 const instruction INS_NEG = INS_rsb;
9296 const instruction INS_TEST = INS_tst;
9297 const instruction INS_MUL = INS_mul;
9298 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9299 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9300 const instruction INS_BREAKPOINT = INS_bkpt;
9301 const instruction INS_ADDC = INS_adc;
9302 const instruction INS_SUBC = INS_sbc;
9303 const instruction INS_NOT = INS_mvn;
9307 #ifdef _TARGET_ARM64_
9309 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9310 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9311 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9313 const instruction INS_AND = INS_and;
9314 const instruction INS_OR = INS_orr;
9315 const instruction INS_XOR = INS_eor;
9316 const instruction INS_NEG = INS_neg;
9317 const instruction INS_TEST = INS_tst;
9318 const instruction INS_MUL = INS_mul;
9319 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9320 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9321 const instruction INS_BREAKPOINT = INS_bkpt;
9322 const instruction INS_ADDC = INS_adc;
9323 const instruction INS_SUBC = INS_sbc;
9324 const instruction INS_NOT = INS_mvn;
9328 /*****************************************************************************/
9330 extern const BYTE genTypeSizes[];
9331 extern const BYTE genTypeAlignments[];
9332 extern const BYTE genTypeStSzs[];
9333 extern const BYTE genActualTypes[];
9335 /*****************************************************************************/
9337 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9338 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9341 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9342 #elif defined(_TARGET_ARM64_)
9343 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9346 /*****************************************************************************/
9348 #define REG_CORRUPT regNumber(REG_NA + 1)
9349 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9350 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9352 /*****************************************************************************/
9354 extern BasicBlock dummyBB;
9356 /*****************************************************************************/
9357 /*****************************************************************************/
9359 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9360 // nodes of a statement in execution order.
9361 // __stmt: a GT_STMT type GenTree*
9362 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9364 #define foreach_treenode_execution_order(__node, __stmt) \
9365 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9367 // foreach_block: An iterator over all blocks in the function.
9368 // __compiler: the Compiler* object
9369 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9371 #define foreach_block(__compiler, __block) \
9372 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9374 /*****************************************************************************/
9375 /*****************************************************************************/
9379 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9381 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9382 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9384 XX Debugging helpers XX
9386 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9387 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9390 /*****************************************************************************/
9391 /* The following functions are intended to be called from the debugger, to dump
9392 * various data structures. The can be used in the debugger Watch or Quick Watch
9393 * windows. They are designed to be short to type and take as few arguments as
9394 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9395 * See the function definition comment for more details.
9398 void cBlock(Compiler* comp, BasicBlock* block);
9399 void cBlocks(Compiler* comp);
9400 void cBlocksV(Compiler* comp);
9401 void cTree(Compiler* comp, GenTree* tree);
9402 void cTrees(Compiler* comp);
9403 void cEH(Compiler* comp);
9404 void cVar(Compiler* comp, unsigned lclNum);
9405 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9406 void cVars(Compiler* comp);
9407 void cVarsFinal(Compiler* comp);
9408 void cBlockPreds(Compiler* comp, BasicBlock* block);
9409 void cReach(Compiler* comp);
9410 void cDoms(Compiler* comp);
9411 void cLiveness(Compiler* comp);
9412 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9414 void cFuncIR(Compiler* comp);
9415 void cBlockIR(Compiler* comp, BasicBlock* block);
9416 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9417 void cTreeIR(Compiler* comp, GenTree* tree);
9418 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9419 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9420 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9421 int cOperandIR(Compiler* comp, GenTree* operand);
9422 int cLeafIR(Compiler* comp, GenTree* tree);
9423 int cIndirIR(Compiler* comp, GenTree* tree);
9424 int cListIR(Compiler* comp, GenTree* list);
9425 int cSsaNumIR(Compiler* comp, GenTree* tree);
9426 int cValNumIR(Compiler* comp, GenTree* tree);
9427 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9429 void dBlock(BasicBlock* block);
9432 void dTree(GenTree* tree);
9435 void dVar(unsigned lclNum);
9436 void dVarDsc(LclVarDsc* varDsc);
9439 void dBlockPreds(BasicBlock* block);
9443 void dCVarSet(VARSET_VALARG_TP vars);
9445 void dVarSet(VARSET_VALARG_TP vars);
9446 void dRegMask(regMaskTP mask);
9449 void dBlockIR(BasicBlock* block);
9450 void dTreeIR(GenTree* tree);
9451 void dLoopIR(Compiler::LoopDsc* loop);
9452 void dLoopNumIR(unsigned loopNum);
9453 int dTabStopIR(int curr, int tabstop);
9454 int dTreeTypeIR(GenTree* tree);
9455 int dTreeKindsIR(GenTree* tree);
9456 int dTreeFlagsIR(GenTree* tree);
9457 int dOperandIR(GenTree* operand);
9458 int dLeafIR(GenTree* tree);
9459 int dIndirIR(GenTree* tree);
9460 int dListIR(GenTree* list);
9461 int dSsaNumIR(GenTree* tree);
9462 int dValNumIR(GenTree* tree);
9463 int dDependsIR(GenTree* comma);
9466 GenTree* dFindTree(GenTree* tree, unsigned id);
9467 GenTree* dFindTree(unsigned id);
9468 GenTreeStmt* dFindStmt(unsigned id);
9469 BasicBlock* dFindBlock(unsigned bbNum);
9473 #include "compiler.hpp" // All the shared inline functions
9475 /*****************************************************************************/
9476 #endif //_COMPILER_H_
9477 /*****************************************************************************/