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() // 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.
692 if ((lvType == TYP_SIMD12) && !lvIsParam)
694 assert(lvExactSize == 12);
697 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
699 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
702 unsigned lvSlotNum; // original slot # (if remapped)
704 typeInfo lvVerTypeInfo; // type info needed for verification
706 BYTE* lvGcLayout; // GC layout info for structs
709 BlockSet lvRefBlks; // Set of blocks that contain refs
710 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
711 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
713 var_types TypeGet() const
715 return (var_types)lvType;
717 bool lvStackAligned() const
719 assert(lvIsStructField);
720 return ((lvFldOffset % sizeof(void*)) == 0);
722 bool lvNormalizeOnLoad() const
724 return varTypeIsSmall(TypeGet()) &&
725 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
726 (lvIsParam || lvAddrExposed || lvIsStructField);
729 bool lvNormalizeOnStore()
731 return varTypeIsSmall(TypeGet()) &&
732 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
733 !(lvIsParam || lvAddrExposed || lvIsStructField);
736 void lvaResetSortAgainFlag(Compiler* pComp);
737 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
738 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
739 void setPrefReg(regNumber regNum, Compiler* pComp);
740 void addPrefReg(regMaskTP regMask, Compiler* pComp);
741 bool IsFloatRegType() const
743 return isFloatRegType(lvType) || lvIsHfaRegArg();
745 var_types GetHfaType() const
747 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
749 void SetHfaType(var_types type)
751 assert(varTypeIsFloating(type));
752 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
755 #ifndef LEGACY_BACKEND
756 var_types lvaArgType();
759 PerSsaArray lvPerSsaData;
762 // Keep track of the # of SsaNames, for a bounds check.
763 unsigned lvNumSsaNames;
766 // Returns the address of the per-Ssa data for the given ssaNum (which is required
767 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
768 // not an SSA variable).
769 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
771 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
772 assert(SsaConfig::RESERVED_SSA_NUM == 0);
773 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
774 assert(zeroBased < lvNumSsaNames);
775 return &lvPerSsaData.GetRef(zeroBased);
780 void PrintVarReg() const
782 if (isRegPairType(TypeGet()))
784 printf("%s:%s", getRegName(lvOtherReg), // hi32
785 getRegName(lvRegNum)); // lo32
789 printf("%s", getRegName(lvRegNum));
794 }; // class LclVarDsc
797 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
802 XX The temporary lclVars allocated by the compiler for code generation XX
804 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
808 /*****************************************************************************
810 * The following keeps track of temporaries allocated in the stack frame
811 * during code-generation (after register allocation). These spill-temps are
812 * only used if we run out of registers while evaluating a tree.
814 * These are different from the more common temps allocated by lvaGrabTemp().
825 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
833 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
837 0); // temps must have a negative number (so they have a different number from all local variables)
838 tdOffs = BAD_TEMP_OFFSET;
842 IMPL_LIMITATION("too many spill temps");
847 bool tdLegalOffset() const
849 return tdOffs != BAD_TEMP_OFFSET;
853 int tdTempOffs() const
855 assert(tdLegalOffset());
858 void tdSetTempOffs(int offs)
861 assert(tdLegalOffset());
863 void tdAdjustTempOffs(int offs)
866 assert(tdLegalOffset());
869 int tdTempNum() const
874 unsigned tdTempSize() const
878 var_types tdTempType() const
884 // interface to hide linearscan implementation from rest of compiler
885 class LinearScanInterface
888 virtual void doLinearScan() = 0;
889 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
892 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
894 // Information about arrays: their element type and size, and the offset of the first element.
895 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
896 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
897 // for example, in value numbering of array index expressions.
900 var_types m_elemType;
901 CORINFO_CLASS_HANDLE m_elemStructType;
903 unsigned m_elemOffset;
905 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
909 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
910 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
915 // This enumeration names the phases into which we divide compilation. The phases should completely
916 // partition a compilation.
919 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
920 #include "compphases.h"
924 extern const char* PhaseNames[];
925 extern const char* PhaseEnums[];
926 extern const LPCWSTR PhaseShortNames[];
928 // The following enum provides a simple 1:1 mapping to CLR API's
929 enum API_ICorJitInfo_Names
931 #define DEF_CLR_API(name) API_##name,
932 #include "ICorJitInfo_API_names.h"
936 //---------------------------------------------------------------
940 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
941 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
942 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
943 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
944 // by "m_timerFailure" being true.
945 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
948 #ifdef FEATURE_JIT_METHOD_PERF
949 // The string names of the phases.
950 static const char* PhaseNames[];
952 static bool PhaseHasChildren[];
953 static int PhaseParent[];
955 unsigned m_byteCodeBytes;
956 unsigned __int64 m_totalCycles;
957 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
958 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
959 #if MEASURE_CLRAPI_CALLS
960 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
961 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
963 // For better documentation, we call EndPhase on
964 // non-leaf phases. We should also call EndPhase on the
965 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
966 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
967 // We add all such "redundant end phase" intervals to this variable below; we print
968 // it out in a report, so we can verify that it is, indeed, very small. If it ever
969 // isn't, this means that we're doing something significant between the end of the last
970 // declared subphase and the end of its parent.
971 unsigned __int64 m_parentPhaseEndSlop;
974 #if MEASURE_CLRAPI_CALLS
975 // The following measures the time spent inside each individual CLR API call.
976 unsigned m_allClrAPIcalls;
977 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
978 unsigned __int64 m_allClrAPIcycles;
979 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
980 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
981 #endif // MEASURE_CLRAPI_CALLS
983 CompTimeInfo(unsigned byteCodeBytes);
987 #ifdef FEATURE_JIT_METHOD_PERF
989 #if MEASURE_CLRAPI_CALLS
990 struct WrapICorJitInfo;
993 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
994 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
995 // The operation of adding a single method's timing to the summary may be performed concurrently by several
996 // threads, so it is protected by a lock.
997 // This class is intended to be used as a singleton type, with only a single instance.
998 class CompTimeSummaryInfo
1000 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1001 static CritSecObject s_compTimeSummaryLock;
1005 CompTimeInfo m_total;
1006 CompTimeInfo m_maximum;
1008 int m_numFilteredMethods;
1009 CompTimeInfo m_filtered;
1011 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1012 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1013 // If any OS interaction fails, returns 0.0.
1014 double CyclesPerSecond();
1016 // This can use what ever data you want to determine if the value to be added
1017 // belongs in the filtered section (it's always included in the unfiltered section)
1018 bool IncludedInFilteredData(CompTimeInfo& info);
1021 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1022 static CompTimeSummaryInfo s_compTimeSummary;
1024 CompTimeSummaryInfo()
1025 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1029 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1030 // This is thread safe.
1031 void AddInfo(CompTimeInfo& info, bool includePhases);
1033 // Print the summary information to "f".
1034 // This is not thread-safe; assumed to be called by only one thread.
1035 void Print(FILE* f);
1038 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1039 // and when the current phase started. This is intended to be part of a Compilation object. This is
1040 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1044 unsigned __int64 m_start; // Start of the compilation.
1045 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1046 #if MEASURE_CLRAPI_CALLS
1047 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1048 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1049 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1050 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1051 static double s_cyclesPerSec; // Cached for speedier measurements
1054 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1056 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1058 static CritSecObject s_csvLock; // Lock to protect the time log file.
1059 void PrintCsvMethodStats(Compiler* comp);
1062 void* operator new(size_t);
1063 void* operator new[](size_t);
1064 void operator delete(void*);
1065 void operator delete[](void*);
1068 // Initialized the timer instance
1069 JitTimer(unsigned byteCodeSize);
1071 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1073 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1076 static void PrintCsvHeader();
1078 // Ends the current phase (argument is for a redundant check).
1079 void EndPhase(Phases phase);
1081 #if MEASURE_CLRAPI_CALLS
1082 // Start and end a timed CLR API call.
1083 void CLRApiCallEnter(unsigned apix);
1084 void CLRApiCallLeave(unsigned apix);
1085 #endif // MEASURE_CLRAPI_CALLS
1087 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1088 // and adds it to "sum".
1089 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1091 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1092 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1093 // "m_info" to true.
1094 bool GetThreadCycles(unsigned __int64* cycles)
1096 bool res = CycleTimer::GetThreadCyclesS(cycles);
1099 m_info.m_timerFailure = true;
1104 #endif // FEATURE_JIT_METHOD_PERF
1106 //------------------- Function/Funclet info -------------------------------
1107 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1109 FUNC_ROOT, // The main/root function (always id==0)
1110 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1111 FUNC_FILTER, // a funclet associated with an EH filter
1114 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1121 BYTE funFlags; // Currently unused, just here for padding
1122 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1123 // funclet. It is only valid if funKind field indicates this is a
1124 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1126 #if defined(_TARGET_AMD64_)
1128 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1129 emitLocation* startLoc;
1130 emitLocation* endLoc;
1131 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1132 emitLocation* coldEndLoc;
1133 UNWIND_INFO unwindHeader;
1134 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1135 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1136 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1137 unsigned unwindCodeSlot;
1139 #ifdef UNIX_AMD64_ABI
1140 jitstd::vector<CFI_CODE>* cfiCodes;
1141 #endif // UNIX_AMD64_ABI
1143 #elif defined(_TARGET_ARMARCH_)
1145 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1146 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1147 // Note: we only have a pointer here instead of the actual object,
1148 // to save memory in the JIT case (compared to the NGEN case),
1149 // where we don't have any cold section.
1150 // Note 2: we currently don't support hot/cold splitting in functions
1151 // with EH, so uwiCold will be NULL for all funclets.
1153 #endif // _TARGET_ARMARCH_
1155 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1156 // that isn't shared between the main function body and funclets.
1159 struct fgArgTabEntry
1162 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1165 otherRegNum = REG_NA;
1166 isStruct = false; // is this a struct arg
1168 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1170 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1172 // it will point at the actual argument in the gtCallLateArgs list.
1173 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1175 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1177 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1179 unsigned numRegs; // Count of number of registers that this argument uses
1181 // A slot is a pointer sized region in the OutArg area.
1182 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1183 unsigned numSlots; // Count of number of slots that this argument uses
1185 unsigned alignment; // 1 or 2 (slots/registers)
1186 unsigned lateArgInx; // index into gtCallLateArgs list
1187 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1189 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1190 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1191 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1192 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1193 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1194 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1195 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1196 // previous arguments.
1197 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1198 // to be on the stack despite its arg list position.
1200 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1201 bool isStruct : 1; // True if this is a struct arg
1203 regNumber otherRegNum; // The (second) register to use when passing this argument.
1205 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1206 #elif defined(_TARGET_X86_)
1207 __declspec(property(get = getIsStruct)) bool isStruct;
1210 return varTypeIsStruct(node);
1212 #endif // _TARGET_X86_
1215 void SetIsHfaRegArg(bool hfaRegArg)
1217 isHfaRegArg = hfaRegArg;
1220 void SetIsBackFilled(bool backFilled)
1222 isBackFilled = backFilled;
1225 bool IsBackFilled() const
1227 return isBackFilled;
1229 #else // !_TARGET_ARM_
1230 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1232 void SetIsHfaRegArg(bool hfaRegArg)
1236 void SetIsBackFilled(bool backFilled)
1240 bool IsBackFilled() const
1244 #endif // !_TARGET_ARM_
1250 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1252 //-------------------------------------------------------------------------
1254 // The class fgArgInfo is used to handle the arguments
1255 // when morphing a GT_CALL node.
1260 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1261 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1262 unsigned argCount; // Updatable arg count value
1263 unsigned nextSlotNum; // Updatable slot count value
1264 unsigned stkLevel; // Stack depth when we make this call (for x86)
1266 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1267 bool hasRegArgs; // true if we have one or more register arguments
1268 bool hasStackArgs; // true if we have one or more stack arguments
1269 bool argsComplete; // marker for state
1270 bool argsSorted; // marker for state
1271 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1274 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1277 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1278 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1280 fgArgTabEntryPtr AddRegArg(
1281 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1283 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1284 fgArgTabEntryPtr AddRegArg(
1291 const bool isStruct,
1292 const regNumber otherRegNum = REG_NA,
1293 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1294 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1296 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1300 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1302 void RemorphReset();
1303 fgArgTabEntryPtr RemorphRegArg(
1304 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1306 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1308 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1310 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1312 void ArgsComplete();
1316 void EvalArgsToTemps();
1318 void RecordStkLevel(unsigned stkLvl);
1319 unsigned RetrieveStkLevel();
1325 fgArgTabEntryPtr* ArgTable()
1329 unsigned GetNextSlotNum()
1339 return hasStackArgs;
1341 bool AreArgsComplete() const
1343 return argsComplete;
1346 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1347 GenTreePtr GetLateArg(unsigned argIndex);
1351 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1352 // We have the ability to mark source expressions with "Test Labels."
1353 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1354 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1356 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1359 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1360 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1361 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1362 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1363 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1366 struct TestLabelAndNum
1371 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1376 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1378 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1381 // This class implements the "IAllocator" interface, so that we can use
1382 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1384 class CompAllocator : public IAllocator
1387 #if MEASURE_MEM_ALLOC
1391 CompAllocator(Compiler* comp, CompMemKind cmk)
1393 #if MEASURE_MEM_ALLOC
1399 inline void* Alloc(size_t sz);
1401 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1403 // For the compiler's no-release allocator, free operations are no-ops.
1410 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1411 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1413 XX The big guy. The sections are currently organized as : XX
1415 XX o GenTree and BasicBlock XX
1427 XX o PrologScopeInfo XX
1428 XX o CodeGenerator XX
1433 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1434 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1439 friend class emitter;
1440 friend class UnwindInfo;
1441 friend class UnwindFragmentInfo;
1442 friend class UnwindEpilogInfo;
1443 friend class JitTimer;
1444 friend class LinearScan;
1445 friend class fgArgInfo;
1446 friend class Rationalizer;
1448 friend class Lowering;
1449 friend class CSE_DataFlow;
1450 friend class CSE_Heuristic;
1451 friend class CodeGenInterface;
1452 friend class CodeGen;
1453 friend class LclVarDsc;
1454 friend class TempDsc;
1456 friend class ObjectAllocator;
1458 #ifndef _TARGET_64BIT_
1459 friend class DecomposeLongs;
1460 #endif // !_TARGET_64BIT_
1463 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1464 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1466 XX Misc structs definitions XX
1468 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1469 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1473 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1492 bool dumpIRDataflow;
1493 bool dumpIRBlockHeaders;
1495 LPCWSTR dumpIRPhase;
1496 LPCWSTR dumpIRFormat;
1498 bool shouldUseVerboseTrees();
1499 bool asciiTrees; // If true, dump trees using only ASCII characters
1500 bool shouldDumpASCIITrees();
1501 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1502 bool shouldUseVerboseSsa();
1503 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1504 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1506 const char* VarNameToStr(VarName name)
1511 DWORD expensiveDebugCheckLevel;
1514 #if FEATURE_MULTIREG_RET
1515 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1516 #endif // FEATURE_MULTIREG_RET
1519 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1520 #endif // ARM_SOFTFP
1522 //-------------------------------------------------------------------------
1523 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1524 // HFAs are one to four element structs where each element is the same
1525 // type, either all float or all double. They are treated specially
1526 // in the ARM Procedure Call Standard, specifically, they are passed in
1527 // floating-point registers instead of the general purpose registers.
1530 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1531 bool IsHfa(GenTreePtr tree);
1533 var_types GetHfaType(GenTreePtr tree);
1534 unsigned GetHfaCount(GenTreePtr tree);
1536 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1537 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1539 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1540 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1542 //-------------------------------------------------------------------------
1543 // The following is used for validating format of EH table
1547 typedef struct EHNodeDsc* pEHNodeDsc;
1549 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1550 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1563 EHBlockType ehnBlockType; // kind of EH block
1564 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1565 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1566 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1568 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1569 pEHNodeDsc ehnChild; // leftmost nested block
1571 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1572 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1574 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1575 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1577 inline void ehnSetTryNodeType()
1579 ehnBlockType = TryNode;
1581 inline void ehnSetFilterNodeType()
1583 ehnBlockType = FilterNode;
1585 inline void ehnSetHandlerNodeType()
1587 ehnBlockType = HandlerNode;
1589 inline void ehnSetFinallyNodeType()
1591 ehnBlockType = FinallyNode;
1593 inline void ehnSetFaultNodeType()
1595 ehnBlockType = FaultNode;
1598 inline BOOL ehnIsTryBlock()
1600 return ehnBlockType == TryNode;
1602 inline BOOL ehnIsFilterBlock()
1604 return ehnBlockType == FilterNode;
1606 inline BOOL ehnIsHandlerBlock()
1608 return ehnBlockType == HandlerNode;
1610 inline BOOL ehnIsFinallyBlock()
1612 return ehnBlockType == FinallyNode;
1614 inline BOOL ehnIsFaultBlock()
1616 return ehnBlockType == FaultNode;
1619 // returns true if there is any overlap between the two nodes
1620 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1622 if (node1->ehnStartOffset < node2->ehnStartOffset)
1624 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1628 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1632 // fails with BADCODE if inner is not completely nested inside outer
1633 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1635 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1639 //-------------------------------------------------------------------------
1640 // Exception handling functions
1643 #if !FEATURE_EH_FUNCLETS
1645 bool ehNeedsShadowSPslots()
1647 return (info.compXcptnsCount || opts.compDbgEnC);
1650 // 0 for methods with no EH
1651 // 1 for methods with non-nested EH, or where only the try blocks are nested
1652 // 2 for a method with a catch within a catch
1654 unsigned ehMaxHndNestingCount;
1656 #endif // !FEATURE_EH_FUNCLETS
1658 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1659 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1661 bool bbInCatchHandlerILRange(BasicBlock* blk);
1662 bool bbInFilterILRange(BasicBlock* blk);
1663 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1664 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1665 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1666 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1667 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1669 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1670 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1672 // Returns true if "block" is the start of a try region.
1673 bool bbIsTryBeg(BasicBlock* block);
1675 // Returns true if "block" is the start of a handler or filter region.
1676 bool bbIsHandlerBeg(BasicBlock* block);
1678 // Returns true iff "block" is where control flows if an exception is raised in the
1679 // try region, and sets "*regionIndex" to the index of the try for the handler.
1680 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1681 // block of the filter, but not for the filter's handler.
1682 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1684 bool ehHasCallableHandlers();
1686 // Return the EH descriptor for the given region index.
1687 EHblkDsc* ehGetDsc(unsigned regionIndex);
1689 // Return the EH index given a region descriptor.
1690 unsigned ehGetIndex(EHblkDsc* ehDsc);
1692 // Return the EH descriptor index of the enclosing try, for the given region index.
1693 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1695 // Return the EH descriptor index of the enclosing handler, for the given region index.
1696 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1698 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1699 // block is not in a 'try' region).
1700 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1702 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1703 // if this block is not in a filter or handler region).
1704 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1706 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1707 // nullptr if this block's exceptions propagate to caller).
1708 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1710 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1711 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1712 bool ehIsBlockEHLast(BasicBlock* block);
1714 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1716 // Return the region index of the most nested EH region this block is in.
1717 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1719 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1720 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1722 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1723 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1724 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1725 // (It can never be a filter.)
1726 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1728 // A block has been deleted. Update the EH table appropriately.
1729 void ehUpdateForDeletedBlock(BasicBlock* block);
1731 // Determine whether a block can be deleted while preserving the EH normalization rules.
1732 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1734 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1735 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1737 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1738 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1739 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1740 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1741 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1742 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1743 // lives in a filter.)
1744 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1746 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1747 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1748 // (nullptr if the last block is the last block in the program).
1749 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1750 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1753 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1754 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1755 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1758 #if FEATURE_EH_FUNCLETS
1759 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1760 // if there is a filter that protects a region with a nested EH clause (such as a
1761 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1762 // genFuncletProlog() for more details. However, the VM seems to use it for more
1763 // purposes, maybe including debugging. Until we are sure otherwise, always create
1764 // a PSPSym for functions with any EH.
1765 bool ehNeedsPSPSym() const
1769 #else // _TARGET_X86_
1770 return compHndBBtabCount > 0;
1771 #endif // _TARGET_X86_
1774 bool ehAnyFunclets(); // Are there any funclets in this function?
1775 unsigned ehFuncletCount(); // Return the count of funclets in the function
1777 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1778 #else // !FEATURE_EH_FUNCLETS
1779 bool ehAnyFunclets()
1783 unsigned ehFuncletCount()
1788 unsigned bbThrowIndex(BasicBlock* blk)
1790 return blk->bbTryIndex;
1791 } // Get the index to use as the cache key for sharing throw blocks
1792 #endif // !FEATURE_EH_FUNCLETS
1794 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1795 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1796 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1797 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1798 // convenient to also consider it a predecessor.)
1799 flowList* BlockPredsWithEH(BasicBlock* blk);
1801 // This table is useful for memoization of the method above.
1802 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1804 BlockToFlowListMap* m_blockToEHPreds;
1805 BlockToFlowListMap* GetBlockToEHPreds()
1807 if (m_blockToEHPreds == nullptr)
1809 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1811 return m_blockToEHPreds;
1814 void* ehEmitCookie(BasicBlock* block);
1815 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1817 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1819 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1821 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1823 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1825 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1827 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1829 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1831 void fgAllocEHTable();
1833 void fgRemoveEHTableEntry(unsigned XTnum);
1835 #if FEATURE_EH_FUNCLETS
1837 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1839 #endif // FEATURE_EH_FUNCLETS
1843 #endif // !FEATURE_EH
1845 void fgSortEHTable();
1847 // Causes the EH table to obey some well-formedness conditions, by inserting
1848 // empty BB's when necessary:
1849 // * No block is both the first block of a handler and the first block of a try.
1850 // * No block is the first block of multiple 'try' regions.
1851 // * No block is the last block of multiple EH regions.
1852 void fgNormalizeEH();
1853 bool fgNormalizeEHCase1();
1854 bool fgNormalizeEHCase2();
1855 bool fgNormalizeEHCase3();
1858 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1859 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1860 void fgVerifyHandlerTab();
1861 void fgDispHandlerTab();
1864 bool fgNeedToSortEHTable;
1866 void verInitEHTree(unsigned numEHClauses);
1867 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1868 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1869 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1870 void verCheckNestingLevel(EHNodeDsc* initRoot);
1873 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1874 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1876 XX GenTree and BasicBlock XX
1878 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1880 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1881 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1884 // Functions to create nodes
1885 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1888 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1890 // For binary opers.
1891 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1893 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1895 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1896 var_types type = TYP_I_IMPL,
1897 GenTreePtr op1 = nullptr,
1898 GenTreePtr op2 = nullptr);
1900 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1902 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1904 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1906 GenTreePtr gtNewJmpTableNode();
1907 GenTreePtr gtNewIconHandleNode(
1908 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1910 unsigned gtTokenToIconFlags(unsigned token);
1912 GenTreePtr gtNewIconEmbHndNode(void* value,
1915 unsigned handle1 = 0,
1916 void* handle2 = nullptr,
1917 void* compileTimeHandle = nullptr);
1919 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1920 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1921 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1922 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1924 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1926 GenTreePtr gtNewLconNode(__int64 value);
1928 GenTreePtr gtNewDconNode(double value);
1930 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1932 GenTreePtr gtNewZeroConNode(var_types type);
1934 GenTreePtr gtNewOneConNode(var_types type);
1937 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1938 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1941 GenTreeBlk* gtNewBlkOpNode(
1942 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1944 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1947 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1950 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1951 void gtSetObjGcInfo(GenTreeObj* objNode);
1952 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1953 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1955 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1957 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1959 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1960 CORINFO_METHOD_HANDLE handle,
1962 GenTreeArgList* args,
1963 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1965 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1967 GenTreeArgList* args,
1968 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1970 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1973 GenTreeArgList* args = nullptr);
1975 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1978 GenTreeSIMD* gtNewSIMDNode(
1979 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1980 GenTreeSIMD* gtNewSIMDNode(var_types type,
1983 SIMDIntrinsicID simdIntrinsicID,
1988 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1989 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
1990 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
1992 GenTreePtr gtNewCodeRef(BasicBlock* block);
1994 GenTreePtr gtNewFieldRef(
1995 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
1997 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
1999 GenTreeArgList* gtNewArgList(GenTreePtr op);
2000 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2001 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2003 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2004 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2005 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2006 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2008 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2010 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2012 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2013 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2014 CORINFO_ACCESS_FLAGS access,
2015 CORINFO_FIELD_INFO* pFieldInfo,
2017 CORINFO_CLASS_HANDLE structType,
2020 GenTreePtr gtNewNothingNode();
2022 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2024 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2026 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2028 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2030 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2032 //------------------------------------------------------------------------
2033 // Other GenTree functions
2035 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2037 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2038 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2039 // IntCnses with value `deepVarVal`.
2040 GenTreePtr gtCloneExpr(
2041 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2043 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2044 // `varNum` to int constants with value `varVal`.
2045 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2047 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2050 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2052 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2054 // Returns "true" iff the complexity (not formally defined, but first interpretation
2055 // is #of nodes in subtree) of "tree" is greater than "limit".
2056 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2057 // before they have been set.)
2058 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2060 bool gtCompareTree(GenTree* op1, GenTree* op2);
2062 GenTreePtr gtReverseCond(GenTree* tree);
2064 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2066 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2068 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2070 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2073 unsigned gtHashValue(GenTree* tree);
2075 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2078 void gtPrepareCost(GenTree* tree);
2079 bool gtIsLikelyRegVar(GenTree* tree);
2081 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2083 // Returns true iff the secondNode can be swapped with firstNode.
2084 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2086 unsigned gtSetEvalOrder(GenTree* tree);
2088 #if FEATURE_STACK_FP_X87
2090 void gtComputeFPlvls(GenTreePtr tree);
2091 #endif // FEATURE_STACK_FP_X87
2093 void gtSetStmtInfo(GenTree* stmt);
2095 // Returns "true" iff "node" has any of the side effects in "flags".
2096 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2098 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2099 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2101 // Appends 'expr' in front of 'list'
2102 // 'list' will typically start off as 'nullptr'
2103 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2104 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2106 void gtExtractSideEffList(GenTreePtr expr,
2108 unsigned flags = GTF_SIDE_EFFECT,
2109 bool ignoreRoot = false);
2111 GenTreePtr gtGetThisArg(GenTreePtr call);
2113 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2114 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2115 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2116 // the given "fldHnd", is such an object pointer.
2117 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2119 // Return true if call is a recursive call; return false otherwise.
2120 bool gtIsRecursiveCall(GenTreeCall* call)
2122 return (call->gtCallMethHnd == info.compMethodHnd);
2125 //-------------------------------------------------------------------------
2127 GenTreePtr gtFoldExpr(GenTreePtr tree);
2130 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2131 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2132 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2133 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2134 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2135 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2136 // optimizations for now.
2137 __attribute__((optnone))
2139 gtFoldExprConst(GenTreePtr tree);
2140 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2141 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2143 //-------------------------------------------------------------------------
2144 // Get the handle, if any.
2145 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2146 // Get the handle, and assert if not found.
2147 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2149 //-------------------------------------------------------------------------
2150 // Functions to display the trees
2153 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2155 void gtDispVN(GenTreePtr tree);
2156 void gtDispConst(GenTreePtr tree);
2157 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2158 void gtDispNodeName(GenTreePtr tree);
2159 void gtDispRegVal(GenTreePtr tree);
2171 void gtDispChild(GenTreePtr child,
2172 IndentStack* indentStack,
2174 __in_opt const char* msg = nullptr,
2175 bool topOnly = false);
2176 void gtDispTree(GenTreePtr tree,
2177 IndentStack* indentStack = nullptr,
2178 __in_opt const char* msg = nullptr,
2179 bool topOnly = false,
2180 bool isLIR = false);
2181 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2182 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2183 char* gtGetLclVarName(unsigned lclNum);
2184 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2185 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2186 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2187 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2188 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2189 void gtDispFieldSeq(FieldSeqNode* pfsn);
2191 void gtDispRange(LIR::ReadOnlyRange const& range);
2193 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2195 void gtDispLIRNode(GenTree* node);
2207 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2208 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2211 static fgWalkPreFn gtAssertColonCond;
2213 static fgWalkPreFn gtMarkColonCond;
2214 static fgWalkPreFn gtClearColonCond;
2216 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2217 bool gtHasCatchArg(GenTreePtr tree);
2218 bool gtHasUnmanagedCall(GenTreePtr tree);
2220 typedef ArrayStack<GenTree*> GenTreeStack;
2222 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2223 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2225 //=========================================================================
2226 // BasicBlock functions
2228 // This is a debug flag we will use to assert when creating block during codegen
2229 // as this interferes with procedure splitting. If you know what you're doing, set
2230 // it to true before creating the block. (DEBUG only)
2231 bool fgSafeBasicBlockCreation;
2234 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2237 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2238 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2242 XX The variables to be used by the code generator. XX
2244 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2245 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2249 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2250 // be placed in the stack frame and it's fields must be laid out sequentially.
2252 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2253 // a local variable that can be enregistered or placed in the stack frame.
2254 // The fields do not need to be laid out sequentially
2256 enum lvaPromotionType
2258 PROMOTION_TYPE_NONE, // The struct local is not promoted
2259 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2260 // and its field locals are independent of its parent struct local.
2261 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2262 // but its field locals depend on its parent struct local.
2265 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2266 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2268 /*****************************************************************************/
2270 enum FrameLayoutState
2273 INITIAL_FRAME_LAYOUT,
2274 PRE_REGALLOC_FRAME_LAYOUT,
2275 REGALLOC_FRAME_LAYOUT,
2276 TENTATIVE_FRAME_LAYOUT,
2281 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2282 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2283 bool lvaSortAgain; // true: We need to sort the lvaTable
2284 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2285 unsigned lvaCount; // total number of locals
2287 unsigned lvaRefCount; // total number of references to locals
2288 LclVarDsc* lvaTable; // variable descriptor table
2289 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2291 LclVarDsc** lvaRefSorted; // table sorted by refcount
2293 unsigned short lvaTrackedCount; // actual # of locals being tracked
2294 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2296 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2297 // Only for AMD64 System V cache the first caller stack homed argument.
2298 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2299 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2302 VARSET_TP lvaTrackedVars; // set of tracked variables
2304 #ifndef _TARGET_64BIT_
2305 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2307 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2309 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2310 // It that changes, this changes. VarSets from different epochs
2311 // cannot be meaningfully combined.
2313 unsigned GetCurLVEpoch()
2318 // reverse map of tracked number to var number
2319 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2321 #ifdef LEGACY_BACKEND
2322 // variable interference graph
2323 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2326 // variable preference graph
2327 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2331 // # of procs compiled a with double-aligned stack
2332 static unsigned s_lvaDoubleAlignedProcsCount;
2336 // Getters and setters for address-exposed and do-not-enregister local var properties.
2337 bool lvaVarAddrExposed(unsigned varNum);
2338 void lvaSetVarAddrExposed(unsigned varNum);
2339 bool lvaVarDoNotEnregister(unsigned varNum);
2341 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2342 enum DoNotEnregisterReason
2347 DNER_VMNeedsStackAddr,
2348 DNER_LiveInOutOfHandler,
2349 DNER_LiveAcrossUnmanagedCall,
2350 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2351 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2352 #ifdef JIT32_GCENCODER
2357 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2359 unsigned lvaVarargsHandleArg;
2361 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2363 #endif // _TARGET_X86_
2365 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2366 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2367 #if FEATURE_FIXED_OUT_ARGS
2368 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2370 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2371 // that tracks whether the lock has been taken
2373 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2374 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2375 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2377 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2378 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2380 #if FEATURE_FIXED_OUT_ARGS
2381 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2382 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2383 #endif // FEATURE_FIXED_OUT_ARGS
2386 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2387 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2388 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2389 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2390 // this variable to be this scratch word whenever struct promotion occurs.
2391 unsigned lvaPromotedStructAssemblyScratchVar;
2392 #endif // _TARGET_ARM_
2395 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2396 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2399 bool lvaGenericsContextUsed;
2401 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2402 // CORINFO_GENERICS_CTXT_FROM_THIS?
2403 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2405 //-------------------------------------------------------------------------
2406 // All these frame offsets are inter-related and must be kept in sync
2408 #if !FEATURE_EH_FUNCLETS
2409 // This is used for the callable handlers
2410 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2411 #endif // FEATURE_EH_FUNCLETS
2413 unsigned lvaCachedGenericContextArgOffs;
2414 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2417 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2419 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2421 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2422 // after the reg predict we will use a computed maxTmpSize
2423 // which is based upon the number of spill temps predicted by reg predict
2424 // All this is necessary because if we under-estimate the size of the spill
2425 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2427 // Pre codegen max spill temp size.
2428 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2430 //-------------------------------------------------------------------------
2432 unsigned lvaGetMaxSpillTempSize();
2434 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2435 #endif // _TARGET_ARM_
2436 void lvaAssignFrameOffsets(FrameLayoutState curState);
2437 void lvaFixVirtualFrameOffsets();
2439 #ifndef LEGACY_BACKEND
2440 void lvaUpdateArgsWithInitialReg();
2441 #endif // !LEGACY_BACKEND
2443 void lvaAssignVirtualFrameOffsetsToArgs();
2444 #ifdef UNIX_AMD64_ABI
2445 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2446 #else // !UNIX_AMD64_ABI
2447 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2448 #endif // !UNIX_AMD64_ABI
2449 void lvaAssignVirtualFrameOffsetsToLocals();
2450 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2451 #ifdef _TARGET_AMD64_
2452 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2453 bool lvaIsCalleeSavedIntRegCountEven();
2455 void lvaAlignFrame();
2456 void lvaAssignFrameOffsetsToPromotedStructs();
2457 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2460 void lvaDumpRegLocation(unsigned lclNum);
2461 void lvaDumpFrameLocation(unsigned lclNum);
2462 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2463 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2464 // layout state defined by lvaDoneFrameLayout
2467 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2468 // to avoid bugs from borderline cases.
2469 #define MAX_FrameSize 0x3FFFFFFF
2470 void lvaIncrementFrameSize(unsigned size);
2472 unsigned lvaFrameSize(FrameLayoutState curState);
2474 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2475 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2477 // Returns the caller-SP-relative offset for the local variable "varNum."
2478 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2480 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2481 int lvaGetSPRelativeOffset(unsigned varNum);
2483 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2484 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2486 //------------------------ For splitting types ----------------------------
2488 void lvaInitTypeRef();
2490 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2491 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2492 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2493 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2494 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2495 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2497 void lvaInitVarDsc(LclVarDsc* varDsc,
2499 CorInfoType corInfoType,
2500 CORINFO_CLASS_HANDLE typeHnd,
2501 CORINFO_ARG_LIST_HANDLE varList,
2502 CORINFO_SIG_INFO* varSig);
2504 static unsigned lvaTypeRefMask(var_types type);
2506 var_types lvaGetActualType(unsigned lclNum);
2507 var_types lvaGetRealType(unsigned lclNum);
2509 //-------------------------------------------------------------------------
2513 unsigned lvaLclSize(unsigned varNum);
2514 unsigned lvaLclExactSize(unsigned varNum);
2516 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2518 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2519 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2520 // the return result.
2521 bool lvaLclVarRefsAccum(
2522 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2524 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2525 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2526 // and (destructively) unions "trkedVars" into "*result".
2527 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2529 ALLVARSET_VALARG_TP allVars,
2530 VARSET_VALARG_TP trkdVars);
2532 bool lvaHaveManyLocals() const;
2534 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2535 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2536 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2539 void lvaSortByRefCount();
2540 void lvaDumpRefCounts();
2542 void lvaMarkLocalVars(BasicBlock* block);
2544 void lvaMarkLocalVars(); // Local variable ref-counting
2546 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2548 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2550 static fgWalkPreFn lvaIncRefCntsCB;
2551 void lvaIncRefCnts(GenTreePtr tree);
2553 static fgWalkPreFn lvaDecRefCntsCB;
2554 void lvaDecRefCnts(GenTreePtr tree);
2555 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2556 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2557 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2560 struct lvaStressLclFldArgs
2562 Compiler* m_pCompiler;
2566 static fgWalkPreFn lvaStressLclFldCB;
2567 void lvaStressLclFld();
2569 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2570 void lvaDispVarSet(VARSET_VALARG_TP set);
2575 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2577 int lvaFrameAddress(int varNum, bool* pFPbased);
2580 bool lvaIsParameter(unsigned varNum);
2581 bool lvaIsRegArgument(unsigned varNum);
2582 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2583 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2584 // that writes to arg0
2586 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2587 // (this is an overload of lvIsTemp because there are no temp parameters).
2588 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2589 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2590 bool lvaIsImplicitByRefLocal(unsigned varNum)
2592 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2593 LclVarDsc* varDsc = &(lvaTable[varNum]);
2594 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2596 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2599 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2603 // Returns true if this local var is a multireg struct
2604 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2606 // If the class is a TYP_STRUCT, get/set a class handle describing it
2608 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2609 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2611 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2613 // Info about struct fields
2614 struct lvaStructFieldInfo
2616 CORINFO_FIELD_HANDLE fldHnd;
2617 unsigned char fldOffset;
2618 unsigned char fldOrdinal;
2621 CORINFO_CLASS_HANDLE fldTypeHnd;
2624 // Info about struct to be promoted.
2625 struct lvaStructPromotionInfo
2627 CORINFO_CLASS_HANDLE typeHnd;
2629 bool requiresScratchVar;
2632 unsigned char fieldCnt;
2633 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2635 lvaStructPromotionInfo()
2636 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2641 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2642 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2643 lvaStructPromotionInfo* StructPromotionInfo,
2645 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2646 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2647 #if !defined(_TARGET_64BIT_)
2648 void lvaPromoteLongVars();
2649 #endif // !defined(_TARGET_64BIT_)
2650 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2651 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2652 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2653 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2654 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2655 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2656 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2658 BYTE* lvaGetGcLayout(unsigned varNum);
2659 bool lvaTypeIsGC(unsigned varNum);
2660 unsigned lvaGSSecurityCookie; // LclVar number
2661 bool lvaTempsHaveLargerOffsetThanVars();
2663 unsigned lvaSecurityObject; // variable representing the security object on the stack
2664 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2666 #if FEATURE_EH_FUNCLETS
2667 unsigned lvaPSPSym; // variable representing the PSPSym
2670 InlineInfo* impInlineInfo;
2671 InlineStrategy* m_inlineStrategy;
2673 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2674 Compiler* impInlineRoot();
2676 #if defined(DEBUG) || defined(INLINE_DATA)
2677 unsigned __int64 getInlineCycleCount()
2679 return m_compCycles;
2681 #endif // defined(DEBUG) || defined(INLINE_DATA)
2683 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2684 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2686 //=========================================================================
2688 //=========================================================================
2691 //---------------- Local variable ref-counting ----------------------------
2694 BasicBlock* lvaMarkRefsCurBlock;
2695 GenTreePtr lvaMarkRefsCurStmt;
2697 BasicBlock::weight_t lvaMarkRefsWeight;
2699 static fgWalkPreFn lvaMarkLclRefsCallback;
2700 void lvaMarkLclRefs(GenTreePtr tree);
2702 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2703 PerSsaArray lvHeapPerSsaData;
2704 unsigned lvHeapNumSsaNames;
2707 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2708 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2709 // not an SSA variable).
2710 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2712 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2713 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2715 assert(ssaNum < lvHeapNumSsaNames);
2716 return &lvHeapPerSsaData.GetRef(ssaNum);
2720 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2721 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2725 XX Imports the given method and converts it to semantic trees XX
2727 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2728 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2734 void impImport(BasicBlock* method);
2736 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2737 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2738 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2739 CORINFO_CLASS_HANDLE impGetStringClass();
2740 CORINFO_CLASS_HANDLE impGetObjectClass();
2742 //=========================================================================
2744 //=========================================================================
2747 //-------------------- Stack manipulation ---------------------------------
2749 unsigned impStkSize; // Size of the full stack
2751 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2753 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2755 struct SavedStack // used to save/restore stack contents.
2757 unsigned ssDepth; // number of values on stack
2758 StackEntry* ssTrees; // saved tree values
2761 bool impIsPrimitive(CorInfoType type);
2762 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2764 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2765 void impPushOnStackNoType(GenTreePtr tree);
2767 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2768 void impPushNullObjRefOnStack();
2769 StackEntry impPopStack();
2770 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2771 GenTreePtr impPopStack(typeInfo& ti);
2772 StackEntry& impStackTop(unsigned n = 0);
2774 void impSaveStackState(SavedStack* savePtr, bool copy);
2775 void impRestoreStackState(SavedStack* savePtr);
2777 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2778 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2779 CORINFO_CALL_INFO* pCallInfo);
2781 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2783 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2785 bool impCanPInvokeInline();
2786 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2787 void impCheckForPInvokeCall(
2788 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2789 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2790 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2792 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2793 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2794 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2796 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2797 CORINFO_METHOD_HANDLE calleeMethodHnd,
2798 CORINFO_CLASS_HANDLE delegateTypeHnd);
2800 var_types impImportCall(OPCODE opcode,
2801 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2802 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2804 GenTreePtr newobjThis,
2806 CORINFO_CALL_INFO* callInfo,
2807 IL_OFFSET rawILOffset);
2809 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2811 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2813 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2816 var_types impImportJitTestLabelMark(int numArgs);
2819 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2821 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2823 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2824 CORINFO_ACCESS_FLAGS access,
2825 CORINFO_FIELD_INFO* pFieldInfo,
2828 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2830 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2832 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2834 void impImportLeave(BasicBlock* block);
2835 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2836 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2837 CORINFO_CLASS_HANDLE clsHnd,
2838 CORINFO_METHOD_HANDLE method,
2839 CORINFO_SIG_INFO* sig,
2843 CorInfoIntrinsics* pIntrinsicID);
2844 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2845 CORINFO_SIG_INFO* sig,
2848 CorInfoIntrinsics intrinsicID);
2849 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2851 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2853 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2854 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2855 CORINFO_THIS_TRANSFORM transform);
2857 //----------------- Manipulating the trees and stmts ----------------------
2859 GenTreePtr impTreeList; // Trees for the BB being imported
2860 GenTreePtr impTreeLast; // The last tree for the current BB
2864 CHECK_SPILL_ALL = -1,
2865 CHECK_SPILL_NONE = -2
2869 void impBeginTreeList();
2870 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2871 void impEndTreeList(BasicBlock* block);
2872 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2873 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2874 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2875 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2876 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2877 void impAssignTempGen(unsigned tmp,
2880 GenTreePtr* pAfterStmt = nullptr,
2881 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2882 BasicBlock* block = nullptr);
2883 void impAssignTempGen(unsigned tmpNum,
2885 CORINFO_CLASS_HANDLE structHnd,
2887 GenTreePtr* pAfterStmt = nullptr,
2888 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2889 BasicBlock* block = nullptr);
2890 GenTreePtr impCloneExpr(GenTreePtr tree,
2892 CORINFO_CLASS_HANDLE structHnd,
2894 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2895 GenTreePtr impAssignStruct(GenTreePtr dest,
2897 CORINFO_CLASS_HANDLE structHnd,
2899 GenTreePtr* pAfterStmt = nullptr,
2900 BasicBlock* block = nullptr);
2901 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2903 CORINFO_CLASS_HANDLE structHnd,
2905 GenTreePtr* pAfterStmt = nullptr,
2906 BasicBlock* block = nullptr);
2908 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2909 CORINFO_CLASS_HANDLE structHnd,
2913 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2914 BYTE* gcLayout = nullptr,
2915 unsigned* numGCVars = nullptr,
2916 var_types* simdBaseType = nullptr);
2918 GenTreePtr impNormStructVal(GenTreePtr structVal,
2919 CORINFO_CLASS_HANDLE structHnd,
2921 bool forceNormalization = false);
2923 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2924 BOOL* pRuntimeLookup = nullptr,
2925 BOOL mustRestoreHandle = FALSE,
2926 BOOL importParent = FALSE);
2928 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2929 BOOL* pRuntimeLookup = nullptr,
2930 BOOL mustRestoreHandle = FALSE)
2932 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2935 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2936 CORINFO_LOOKUP* pLookup,
2938 void* compileTimeHandle);
2940 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2942 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2943 CORINFO_LOOKUP* pLookup,
2944 void* compileTimeHandle);
2946 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2948 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2949 CorInfoHelpFunc helper,
2951 GenTreeArgList* arg = nullptr,
2952 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
2954 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
2956 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2959 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
2960 CORINFO_CLASS_HANDLE typeClass,
2964 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
2965 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
2966 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
2967 static bool IsMathIntrinsic(GenTreePtr tree);
2970 //----------------- Importing the method ----------------------------------
2972 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
2975 unsigned impCurOpcOffs;
2976 const char* impCurOpcName;
2977 bool impNestedStackSpill;
2979 // For displaying instrs with generated native code (-n:B)
2980 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
2981 void impNoteLastILoffs();
2984 /* IL offset of the stmt currently being imported. It gets set to
2985 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
2986 updated at IL offsets for which we have to report mapping info.
2987 It also includes flag bits, so use jitGetILoffs()
2988 to get the actual IL offset value.
2991 IL_OFFSETX impCurStmtOffs;
2992 void impCurStmtOffsSet(IL_OFFSET offs);
2994 void impNoteBranchOffs();
2996 unsigned impInitBlockLineInfo();
2998 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
2999 bool impIsThis(GenTreePtr obj);
3000 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3001 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3002 bool impIsAnySTLOC(OPCODE opcode)
3004 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3005 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3008 GenTreeArgList* impPopList(unsigned count,
3010 CORINFO_SIG_INFO* sig,
3011 GenTreeArgList* prefixTree = nullptr);
3013 GenTreeArgList* impPopRevList(unsigned count,
3015 CORINFO_SIG_INFO* sig,
3016 unsigned skipReverseCount = 0);
3019 * Get current IL offset with stack-empty info incoporated
3021 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3023 //---------------- Spilling the importer stack ----------------------------
3029 SavedStack pdSavedStack;
3030 ThisInitState pdThisPtrInit;
3033 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3034 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3036 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3037 ExpandArray<BYTE> impPendingBlockMembers;
3039 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3040 // Operates on the map in the top-level ancestor.
3041 BYTE impGetPendingBlockMember(BasicBlock* blk)
3043 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3046 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3047 // Operates on the map in the top-level ancestor.
3048 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3050 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3053 bool impCanReimport;
3055 bool impSpillStackEntry(unsigned level,
3059 bool bAssertOnRecursion,
3064 void impSpillStackEnsure(bool spillLeaves = false);
3065 void impEvalSideEffects();
3066 void impSpillSpecialSideEff();
3067 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3068 void impSpillValueClasses();
3069 void impSpillEvalStack();
3070 static fgWalkPreFn impFindValueClasses;
3071 void impSpillLclRefs(ssize_t lclNum);
3073 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3075 void impImportBlockCode(BasicBlock* block);
3077 void impReimportMarkBlock(BasicBlock* block);
3078 void impReimportMarkSuccessors(BasicBlock* block);
3080 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3082 void impImportBlockPending(BasicBlock* block);
3084 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3085 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3086 // for the block, but instead, just re-uses the block's existing EntryState.
3087 void impReimportBlockPending(BasicBlock* block);
3089 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3091 void impImportBlock(BasicBlock* block);
3093 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3094 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3095 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3096 // the variables that will be used -- and for all the predecessors of those successors, and the
3097 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3098 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3099 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3100 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3101 // of local variable numbers, so we represent them with the base local variable number), returns that.
3102 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3103 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3104 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3105 // on which kind of member of the clique the block is).
3106 unsigned impGetSpillTmpBase(BasicBlock* block);
3108 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3109 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3110 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3111 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3112 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3113 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3114 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3115 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3116 // successors receive a native int. Similarly float and double are unified to double.
3117 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3118 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3119 // predecessors, so they insert an upcast if needed).
3120 void impReimportSpillClique(BasicBlock* block);
3122 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3123 // block, and represent the predecessor and successor members of the clique currently being computed.
3124 // *** Access to these will need to be locked in a parallel compiler.
3125 ExpandArray<BYTE> impSpillCliquePredMembers;
3126 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3134 // Abstract class for receiving a callback while walking a spill clique
3135 class SpillCliqueWalker
3138 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3141 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3142 class SetSpillTempsBase : public SpillCliqueWalker
3147 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3150 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3153 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3154 class ReimportSpillClique : public SpillCliqueWalker
3159 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3162 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3165 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3166 // predecessor or successor within the spill clique
3167 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3169 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3170 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3171 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3172 void impRetypeEntryStateTemps(BasicBlock* blk);
3174 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3175 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3177 void impPushVar(GenTree* op, typeInfo tiRetVal);
3178 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3179 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3181 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3183 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3184 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3185 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3188 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3191 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3192 struct BlockListNode
3195 BlockListNode* m_next;
3196 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3199 void* operator new(size_t sz, Compiler* comp);
3201 BlockListNode* impBlockListNodeFreeList;
3203 BlockListNode* AllocBlockListNode();
3204 void FreeBlockListNode(BlockListNode* node);
3206 bool impIsValueType(typeInfo* pTypeInfo);
3207 var_types mangleVarArgsType(var_types type);
3210 regNumber getCallArgIntRegister(regNumber floatReg);
3211 regNumber getCallArgFloatRegister(regNumber intReg);
3212 #endif // FEATURE_VARARG
3215 static unsigned jitTotalMethodCompiled;
3219 static LONG jitNestingLevel;
3222 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3224 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3226 // STATIC inlining decision based on the IL code.
3227 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3228 CORINFO_METHOD_INFO* methInfo,
3230 InlineResult* inlineResult);
3232 void impCheckCanInline(GenTreePtr call,
3233 CORINFO_METHOD_HANDLE fncHandle,
3235 CORINFO_CONTEXT_HANDLE exactContextHnd,
3236 InlineCandidateInfo** ppInlineCandidateInfo,
3237 InlineResult* inlineResult);
3239 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3240 GenTreePtr curArgVal,
3242 InlineResult* inlineResult);
3244 void impInlineInitVars(InlineInfo* pInlineInfo);
3246 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3248 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3250 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3252 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3253 GenTreePtr variableBeingDereferenced,
3254 InlArgInfo* inlArgInfo);
3256 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3258 bool impTailCallRetTypeCompatible(var_types callerRetType,
3259 CORINFO_CLASS_HANDLE callerRetTypeClass,
3260 var_types calleeRetType,
3261 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3263 bool impIsTailCallILPattern(bool tailPrefixed,
3265 const BYTE* codeAddrOfNextOpcode,
3266 const BYTE* codeEnd,
3268 bool* IsCallPopRet = nullptr);
3270 bool impIsImplicitTailCallCandidate(
3271 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3274 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3275 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3279 XX Info about the basic-blocks, their contents and the flow analysis XX
3281 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3282 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3286 BasicBlock* fgFirstBB; // Beginning of the basic block list
3287 BasicBlock* fgLastBB; // End of the basic block list
3288 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3289 #if FEATURE_EH_FUNCLETS
3290 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3292 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3294 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3295 unsigned fgEdgeCount; // # of control flow edges between the BBs
3296 unsigned fgBBcount; // # of BBs in the method
3298 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3300 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3301 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3302 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3303 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3305 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3306 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3307 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3308 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3309 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3310 // index). The arrays are of size fgBBNumMax + 1.
3311 unsigned* fgDomTreePreOrder;
3312 unsigned* fgDomTreePostOrder;
3314 bool fgBBVarSetsInited;
3316 // Allocate array like T* a = new T[fgBBNumMax + 1];
3317 // Using helper so we don't keep forgetting +1.
3318 template <typename T>
3319 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3321 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3324 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3325 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3326 // cannot be meaningfully combined. Note that new blocks can be created with higher
3327 // block numbers without changing the basic block epoch. These blocks *cannot*
3328 // participate in a block set until the blocks are all renumbered, causing the epoch
3329 // to change. This is useful if continuing to use previous block sets is valuable.
3330 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3331 unsigned fgCurBBEpoch;
3333 unsigned GetCurBasicBlockEpoch()
3335 return fgCurBBEpoch;
3338 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3339 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3340 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3341 unsigned fgCurBBEpochSize;
3343 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3344 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3345 unsigned fgBBSetCountInSizeTUnits;
3347 void NewBasicBlockEpoch()
3349 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3351 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3353 fgCurBBEpochSize = fgBBNumMax + 1;
3354 fgBBSetCountInSizeTUnits =
3355 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3358 // All BlockSet objects are now invalid!
3359 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3360 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3364 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3365 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3366 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3367 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3369 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3370 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3371 // array of size_t bitsets), then print that out.
3372 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3379 void EnsureBasicBlockEpoch()
3381 if (fgCurBBEpochSize != fgBBNumMax + 1)
3383 NewBasicBlockEpoch();
3387 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3388 void fgEnsureFirstBBisScratch();
3389 bool fgFirstBBisScratch();
3390 bool fgBBisScratch(BasicBlock* block);
3392 void fgExtendEHRegionBefore(BasicBlock* block);
3393 void fgExtendEHRegionAfter(BasicBlock* block);
3395 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3397 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3399 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3402 BasicBlock* nearBlk,
3403 bool putInFilter = false,
3404 bool runRarely = false,
3405 bool insertAtEnd = false);
3407 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3409 bool runRarely = false,
3410 bool insertAtEnd = false);
3412 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3414 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3415 BasicBlock* afterBlk,
3416 unsigned xcptnIndex,
3417 bool putInTryRegion);
3419 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3420 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3421 void fgUnlinkBlock(BasicBlock* block);
3423 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3424 bool fgMultipleNots;
3427 bool fgModified; // True if the flow graph has been modified recently
3428 bool fgComputePredsDone; // Have we computed the bbPreds list
3429 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3430 bool fgDomsComputed; // Have we computed the dominator sets?
3431 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3433 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3434 bool fgHasPostfix; // any postfix ++/-- found?
3435 unsigned fgIncrCount; // number of increment nodes found
3437 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3441 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3442 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3445 bool fgRemoveRestOfBlock; // true if we know that we will throw
3446 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3448 // There are two modes for ordering of the trees.
3449 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3450 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3451 // by traversing the tree according to the order of the operands.
3452 // - In FGOrderLinear, the dominant ordering is the linear order.
3459 FlowGraphOrder fgOrder;
3461 // The following are boolean flags that keep track of the state of internal data structures
3463 bool fgStmtListThreaded;
3464 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3465 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3466 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3467 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3468 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3469 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3470 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3471 // This is derived from the profile data
3472 // or is BB_UNITY_WEIGHT when we don't have profile data
3474 #if FEATURE_EH_FUNCLETS
3475 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3476 #endif // FEATURE_EH_FUNCLETS
3478 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3479 // since fgMorphTree can be called from several places
3480 bool fgExpandInline; // indicates that we are creating tree for the inliner
3482 bool impBoxTempInUse; // the temp below is valid and available
3483 unsigned impBoxTemp; // a temporary that is used for boxing
3486 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3487 // and we are trying to compile again in a "safer", minopts mode?
3491 unsigned impInlinedCodeSize;
3494 //-------------------------------------------------------------------------
3502 void fgRemoveEmptyTry();
3504 void fgRemoveEmptyFinally();
3506 void fgCloneFinally();
3508 void fgCleanupContinuation(BasicBlock* continuation);
3510 void fgUpdateFinallyTargetFlags();
3512 GenTreePtr fgGetCritSectOfStaticMethod();
3514 #if !defined(_TARGET_X86_)
3516 void fgAddSyncMethodEnterExit();
3518 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3520 void fgConvertSyncReturnToLeave(BasicBlock* block);
3522 #endif // !_TARGET_X86_
3524 void fgAddReversePInvokeEnterExit();
3526 bool fgMoreThanOneReturnBlock();
3528 // The number of separate return points in the method.
3529 unsigned fgReturnCount;
3531 void fgAddInternal();
3533 bool fgFoldConditional(BasicBlock* block);
3535 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3536 void fgMorphBlocks();
3538 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3540 void fgCheckArgCnt();
3541 void fgSetOptions();
3544 static fgWalkPreFn fgAssertNoQmark;
3545 void fgPreExpandQmarkChecks(GenTreePtr expr);
3546 void fgPostExpandQmarkChecks();
3547 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3550 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3552 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3553 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3554 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3555 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3556 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3558 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3559 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3560 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3561 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3563 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3564 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3565 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3566 void fgExpandQmarkNodes();
3570 // Do "simple lowering." This functionality is (conceptually) part of "general"
3571 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3572 void fgSimpleLowering();
3574 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3576 GenTreePtr fgInitThisClass();
3578 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3580 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3582 void fgLocalVarLiveness();
3584 void fgLocalVarLivenessInit();
3586 #ifdef LEGACY_BACKEND
3587 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3589 void fgPerNodeLocalVarLiveness(GenTree* node);
3591 void fgPerBlockLocalVarLiveness();
3593 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3595 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3597 // This is used in the liveness computation, as a temporary. When we use the
3598 // arbitrary-length VarSet representation, it is better not to allocate a new one
3600 VARSET_TP fgMarkIntfUnionVS;
3602 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3604 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3606 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3608 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3610 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3612 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3614 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3615 GenTreePtr startNode,
3617 VARSET_VALARG_TP volatileVars,
3618 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3620 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3622 bool fgRemoveDeadStore(GenTree** pTree,
3626 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3628 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3630 // For updating liveset during traversal AFTER fgComputeLife has completed
3631 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3632 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3634 // Returns the set of live variables after endTree,
3635 // assuming that liveSet is the set of live variables BEFORE tree.
3636 // Requires that fgComputeLife has completed, and that tree is in the same
3637 // statement as endTree, and that it comes before endTree in execution order
3639 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3641 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3642 while (tree != nullptr && tree != endTree->gtNext)
3644 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3645 tree = tree->gtNext;
3647 assert(tree == endTree->gtNext);
3651 void fgInterBlockLocalVarLiveness();
3653 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3654 // "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
3655 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3656 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3657 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3658 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3659 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3661 if (m_opAsgnVarDefSsaNums == nullptr)
3663 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3665 return m_opAsgnVarDefSsaNums;
3668 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3669 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3670 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3672 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3674 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3675 // Except: assumes that lcl is a def, and if it is
3676 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3677 // rather than the "use" SSA number recorded in the tree "lcl".
3678 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3680 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3681 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3682 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3683 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3684 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3686 // (byref addrS1 = &s1,
3687 // *(addrS1 * offsetof(f0)) = s2f0,
3689 // *(addrS1 * offsetof(fn)) = s2fn)
3691 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3692 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3693 // give it SSA names and value numbers?
3695 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3696 // end with an instance of the structure below, whose fields are described in the declaration.
3697 struct IndirectAssignmentAnnotation
3699 unsigned m_lclNum; // The local num that is being indirectly assigned.
3700 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3701 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3702 // be the singleton field sequence "g". The individual assignments would
3703 // further append the fields of "s.g" to that.
3704 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3705 // structure has a single field).
3706 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3707 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3710 IndirectAssignmentAnnotation(unsigned lclNum,
3711 FieldSeqNode* fldSeq,
3713 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3714 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3715 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3719 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3720 NodeToIndirAssignMap;
3721 NodeToIndirAssignMap* m_indirAssignMap;
3722 NodeToIndirAssignMap* GetIndirAssignMap()
3724 if (m_indirAssignMap == nullptr)
3726 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3727 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3728 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3730 return m_indirAssignMap;
3733 // Performs SSA conversion.
3736 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3737 void fgResetForSsa();
3739 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3741 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3742 inline bool fgExcludeFromSsa(unsigned lclNum);
3744 // The value numbers for this compilation.
3745 ValueNumStore* vnStore;
3748 ValueNumStore* GetValueNumStore()
3753 // Do value numbering (assign a value number to each
3755 void fgValueNumber();
3757 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3758 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3759 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3760 // match the element type of the array or fldSeq. When this type doesn't match
3761 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3763 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3766 FieldSeqNode* fldSeq,
3770 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3771 // has been parsed to yield the other input arguments. If evaluation of the address
3772 // can raise exceptions, those should be captured in the exception set "excVN."
3773 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3774 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3775 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3776 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3777 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3779 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3780 CORINFO_CLASS_HANDLE elemTypeEq,
3784 FieldSeqNode* fldSeq);
3786 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3787 // by evaluating the array index expression "tree". Returns the value number resulting from
3788 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3789 // "GT_IND" that does the dereference, and it is given the returned value number.
3790 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3792 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3794 // Utility functions for fgValueNumber.
3796 // Perform value-numbering for the trees in "blk".
3797 void fgValueNumberBlock(BasicBlock* blk);
3799 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3800 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3801 // assumed for the heap at the start "entryBlk".
3802 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3804 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3805 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3807 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3808 // value in that SSA #.
3809 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3811 // The input 'tree' is a leaf node that is a constant
3812 // Assign the proper value number to the tree
3813 void fgValueNumberTreeConst(GenTreePtr tree);
3815 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3816 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3818 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3820 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3822 // Does value-numbering for a block assignment.
3823 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3825 // Does value-numbering for a cast tree.
3826 void fgValueNumberCastTree(GenTreePtr tree);
3828 // Does value-numbering for an intrinsic tree.
3829 void fgValueNumberIntrinsic(GenTreePtr tree);
3831 // Does value-numbering for a call. We interpret some helper calls.
3832 void fgValueNumberCall(GenTreeCall* call);
3834 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3835 void fgUpdateArgListVNs(GenTreeArgList* args);
3837 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3838 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3840 // Requires "helpCall" to be a helper call. Assigns it a value number;
3841 // we understand the semantics of some of the calls. Returns "true" if
3842 // the call may modify the heap (we assume arbitrary memory side effects if so).
3843 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3845 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3846 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3848 // This is the current value number for the "Heap" implicit variable while
3849 // doing value numbering. This is the value number under the "liberal" interpretation
3850 // of heap values; the "conservative" interpretation needs no VN, since every access of
3851 // the heap yields an unknown value.
3852 ValueNum fgCurHeapVN;
3854 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3855 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3856 // is 1, and the rest is an encoding of "elemTyp".
3857 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3859 if (elemStructType != nullptr)
3861 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3862 varTypeIsIntegral(elemTyp));
3863 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3864 return elemStructType;
3868 elemTyp = varTypeUnsignedToSigned(elemTyp);
3869 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3872 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3873 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3874 // the struct type of the element).
3875 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3877 size_t clsHndVal = size_t(clsHnd);
3878 if (clsHndVal & 0x1)
3880 return var_types(clsHndVal >> 1);
3888 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3889 var_types getJitGCType(BYTE gcType);
3891 enum structPassingKind
3893 SPK_Unknown, // Invalid value, never returned
3894 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3895 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3896 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3897 // parameters registers are used, then the stack will be used)
3898 // for X86 passed on the stack, for ARM32 passed in registers
3899 // or the stack or split between registers and the stack.
3900 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3902 }; // The struct is passed/returned by reference to a copy/buffer.
3904 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3905 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3906 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3907 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3909 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3911 // Get the type that is used to pass values of the given struct type.
3912 // If you have already retrieved the struct size then pass it as the optional third argument
3914 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3915 structPassingKind* wbPassStruct,
3916 unsigned structSize = 0);
3918 // Get the type that is used to return values of the given struct type.
3919 // If you have already retrieved the struct size then pass it as the optional third argument
3921 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3922 structPassingKind* wbPassStruct = nullptr,
3923 unsigned structSize = 0);
3926 // Print a representation of "vnp" or "vn" on standard output.
3927 // If "level" is non-zero, we also print out a partial expansion of the value.
3928 void vnpPrint(ValueNumPair vnp, unsigned level);
3929 void vnPrint(ValueNum vn, unsigned level);
3932 // Dominator computation member functions
3933 // Not exposed outside Compiler
3935 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3937 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3939 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3940 // flow graph. We first assume the fields bbIDom on each
3941 // basic block are invalid. This computation is needed later
3942 // by fgBuildDomTree to build the dominance tree structure.
3943 // Based on: A Simple, Fast Dominance Algorithm
3944 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3946 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3947 // Note: this is relatively slow compared to calling fgDominate(),
3948 // especially if dealing with a single block versus block check.
3950 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3952 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3954 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3956 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3958 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3960 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3961 // processed in topological sort, this function takes care of that.
3963 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3965 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3966 // Returns this as a set.
3968 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3969 // root nodes. Returns this as a set.
3972 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
3975 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
3976 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
3979 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
3980 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
3981 // && postOrder(A) >= postOrder(B) making the computation O(1).
3982 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
3984 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
3986 void fgUpdateChangedFlowGraph();
3989 // Compute the predecessors of the blocks in the control flow graph.
3990 void fgComputePreds();
3992 // Remove all predecessor information.
3993 void fgRemovePreds();
3995 // Compute the cheap flow graph predecessors lists. This is used in some early phases
3996 // before the full predecessors lists are computed.
3997 void fgComputeCheapPreds();
4000 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4002 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4012 // Initialize the per-block variable sets (used for liveness analysis).
4013 void fgInitBlockVarSets();
4015 // true if we've gone through and created GC Poll calls.
4016 bool fgGCPollsCreated;
4017 void fgMarkGCPollBlocks();
4018 void fgCreateGCPolls();
4019 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4021 // Requires that "block" is a block that returns from
4022 // a finally. Returns the number of successors (jump targets of
4023 // of blocks in the covered "try" that did a "LEAVE".)
4024 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4026 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4027 // a finally. Returns its "i"th successor (jump targets of
4028 // of blocks in the covered "try" that did a "LEAVE".)
4029 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4030 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4033 // Factor out common portions of the impls of the methods above.
4034 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4037 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4038 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4039 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4040 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4041 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4042 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4043 // we leave the entry associated with the block, but it will no longer be accessed.)
4044 struct SwitchUniqueSuccSet
4046 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4047 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4050 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4051 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4052 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4053 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4056 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4057 BlockToSwitchDescMap;
4060 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4061 // iteration over only the distinct successors.
4062 BlockToSwitchDescMap* m_switchDescMap;
4065 BlockToSwitchDescMap* GetSwitchDescMap()
4067 if (m_switchDescMap == nullptr)
4069 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4071 return m_switchDescMap;
4074 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4075 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4076 // we don't accidentally look up and return the wrong switch data.
4077 void InvalidateUniqueSwitchSuccMap()
4079 m_switchDescMap = nullptr;
4082 // Requires "switchBlock" to be a block that ends in a switch. Returns
4083 // the corresponding SwitchUniqueSuccSet.
4084 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4086 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4087 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4088 // remove it from "this", and ensure that "to" is a member.
4089 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4091 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4092 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4094 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4096 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4098 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4100 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4102 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4104 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4106 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4108 void fgRemoveBlockAsPred(BasicBlock* block);
4110 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4112 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4114 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4116 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4118 flowList* fgAddRefPred(BasicBlock* block,
4119 BasicBlock* blockPred,
4120 flowList* oldEdge = nullptr,
4121 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4124 void fgFindBasicBlocks();
4126 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4128 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4130 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4131 bool putInTryRegion,
4132 BasicBlock* startBlk,
4134 BasicBlock* nearBlk,
4135 BasicBlock* jumpBlk,
4138 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4140 void fgRemoveEmptyBlocks();
4142 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4144 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4146 void fgCreateLoopPreHeader(unsigned lnum);
4148 void fgUnreachableBlock(BasicBlock* block);
4150 void fgRemoveConditionalJump(BasicBlock* block);
4152 BasicBlock* fgLastBBInMainFunction();
4154 BasicBlock* fgEndBBAfterMainFunction();
4156 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4158 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4160 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4162 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4164 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4166 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4168 bool fgRenumberBlocks();
4170 bool fgExpandRarelyRunBlocks();
4172 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4174 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4176 enum FG_RELOCATE_TYPE
4178 FG_RELOCATE_TRY, // relocate the 'try' region
4179 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4181 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4183 #if FEATURE_EH_FUNCLETS
4184 #if defined(_TARGET_ARM_)
4185 void fgClearFinallyTargetBit(BasicBlock* block);
4186 #endif // defined(_TARGET_ARM_)
4187 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4188 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4189 void fgInsertFuncletPrologBlock(BasicBlock* block);
4190 void fgCreateFuncletPrologBlocks();
4191 void fgCreateFunclets();
4192 #else // !FEATURE_EH_FUNCLETS
4193 bool fgRelocateEHRegions();
4194 #endif // !FEATURE_EH_FUNCLETS
4196 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4198 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4200 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4202 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4204 bool fgOptimizeEmptyBlock(BasicBlock* block);
4206 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4208 bool fgOptimizeBranch(BasicBlock* bJump);
4210 bool fgOptimizeSwitchBranches(BasicBlock* block);
4212 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4214 bool fgOptimizeSwitchJumps();
4216 void fgPrintEdgeWeights();
4218 void fgComputeEdgeWeights();
4220 void fgReorderBlocks();
4222 void fgDetermineFirstColdBlock();
4224 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4226 bool fgUpdateFlowGraph(bool doTailDup = false);
4228 void fgFindOperOrder();
4230 // method that returns if you should split here
4231 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4233 void fgSetBlockOrder();
4235 void fgRemoveReturnBlock(BasicBlock* block);
4237 /* Helper code that has been factored out */
4238 inline void fgConvertBBToThrowBB(BasicBlock* block);
4240 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4241 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4242 GenTreePtr fgMakeTmpArgNode(
4243 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4245 // The following check for loops that don't execute calls
4246 bool fgLoopCallMarked;
4248 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4249 void fgLoopCallMark();
4251 void fgMarkLoopHead(BasicBlock* block);
4253 unsigned fgGetCodeEstimate(BasicBlock* block);
4256 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4257 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4258 bool fgDumpFlowGraph(Phases phase);
4260 #endif // DUMP_FLOWGRAPHS
4265 void fgDispBBLiveness(BasicBlock* block);
4266 void fgDispBBLiveness();
4267 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4268 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4269 void fgDispBasicBlocks(bool dumpTrees = false);
4270 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4271 void fgDumpBlock(BasicBlock* block);
4272 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4274 static fgWalkPreFn fgStress64RsltMulCB;
4275 void fgStress64RsltMul();
4276 void fgDebugCheckUpdate();
4277 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4278 void fgDebugCheckBlockLinks();
4279 void fgDebugCheckLinks(bool morphTrees = false);
4280 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4281 void fgDebugCheckFlags(GenTreePtr tree);
4282 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4283 void fgDebugCheckTryFinallyExits();
4286 #ifdef LEGACY_BACKEND
4287 static void fgOrderBlockOps(GenTreePtr tree,
4291 GenTreePtr* opsPtr, // OUT
4292 regMaskTP* regsPtr); // OUT
4293 #endif // LEGACY_BACKEND
4295 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4296 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4298 inline bool fgIsInlining()
4300 return fgExpandInline;
4303 void fgTraverseRPO();
4305 //--------------------- Walking the trees in the IR -----------------------
4310 fgWalkPreFn* wtprVisitorFn;
4311 fgWalkPostFn* wtpoVisitorFn;
4312 void* pCallbackData; // user-provided data
4313 bool wtprLclsOnly; // whether to only visit lclvar nodes
4314 GenTreePtr parent; // parent of current node, provided to callback
4315 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4317 bool printModified; // callback can use this
4321 template <bool computeStack>
4322 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4324 // general purpose tree-walker that is capable of doing pre- and post- order
4325 // callbacks at the same time
4326 template <bool doPreOrder, bool doPostOrder>
4327 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4329 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4330 fgWalkPreFn* visitor,
4331 void* pCallBackData = nullptr,
4332 bool lclVarsOnly = false,
4333 bool computeStack = false);
4335 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4336 fgWalkPreFn* preVisitor,
4337 fgWalkPostFn* postVisitor,
4338 void* pCallBackData = nullptr);
4340 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4344 template <bool computeStack>
4345 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4347 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4348 fgWalkPostFn* visitor,
4349 void* pCallBackData = nullptr,
4350 bool computeStack = false);
4352 // An fgWalkPreFn that looks for expressions that have inline throws in
4353 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4354 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4355 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4356 // properly propagated to parent trees). It returns WALK_CONTINUE
4358 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4359 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4360 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4362 /**************************************************************************
4364 *************************************************************************/
4367 friend class SsaBuilder;
4368 friend struct ValueNumberState;
4370 //--------------------- Detect the basic blocks ---------------------------
4372 BasicBlock** fgBBs; // Table of pointers to the BBs
4374 void fgInitBBLookup();
4375 BasicBlock* fgLookupBB(unsigned addr);
4377 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4379 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4381 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4383 void fgLinkBasicBlocks();
4385 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4387 void fgCheckBasicBlockControlFlow();
4389 void fgControlFlowPermitted(BasicBlock* blkSrc,
4390 BasicBlock* blkDest,
4391 BOOL IsLeave = false /* is the src a leave block */);
4393 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4395 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4397 void fgAdjustForAddressExposedOrWrittenThis();
4399 bool fgProfileData_ILSizeMismatch;
4400 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4401 ULONG fgProfileBufferCount;
4402 ULONG fgNumProfileRuns;
4404 unsigned fgStressBBProf()
4407 unsigned result = JitConfig.JitStressBBProf();
4410 if (compStressCompile(STRESS_BB_PROFILE, 15))
4421 bool fgHaveProfileData();
4422 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4424 bool fgIsUsingProfileWeights()
4426 return (fgHaveProfileData() || fgStressBBProf());
4428 void fgInstrumentMethod();
4430 //-------- Insert a statement at the start or end of a basic block --------
4434 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4438 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4440 public: // Used by linear scan register allocation
4441 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4444 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4445 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4447 public: // Used by linear scan register allocation
4448 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4451 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4453 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4455 // Create a new temporary variable to hold the result of *ppTree,
4456 // and transform the graph accordingly.
4457 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4458 GenTree* fgMakeMultiUse(GenTree** ppTree);
4461 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4462 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4463 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4465 //-------- Determine the order in which the trees will be evaluated -------
4467 unsigned fgTreeSeqNum;
4468 GenTree* fgTreeSeqLst;
4469 GenTree* fgTreeSeqBeg;
4471 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4472 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4473 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4474 void fgSetStmtSeq(GenTree* tree);
4475 void fgSetBlockOrder(BasicBlock* block);
4477 //------------------------- Morphing --------------------------------------
4479 unsigned fgPtrArgCntCur;
4480 unsigned fgPtrArgCntMax;
4481 hashBv* fgOutgoingArgTemps;
4482 hashBv* fgCurrentlyInUseArgTemps;
4484 bool compCanEncodePtrArgCntMax();
4486 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4489 void fgMoveOpsLeft(GenTreePtr tree);
4492 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4494 bool fgIsThrow(GenTreePtr tree);
4496 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4497 bool fgIsBlockCold(BasicBlock* block);
4499 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4501 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4503 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4505 bool fgMorphRelopToQmark(GenTreePtr tree);
4507 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4508 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4509 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4510 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4511 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4512 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4513 // small; hence the other fields of MorphAddrContext.
4514 enum MorphAddrContextKind
4519 struct MorphAddrContext
4521 MorphAddrContextKind m_kind;
4522 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4523 // top-level indirection and here have been constants.
4524 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4525 // In that case, is the sum of those constant offsets.
4527 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4532 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4533 static MorphAddrContext s_CopyBlockMAC;
4536 GenTreePtr fgCopySIMDNode(GenTreeSIMD* simdNode);
4537 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4538 var_types* baseTypeOut,
4540 unsigned* simdSizeOut,
4541 bool ignoreUsedInSIMDIntrinsic = false);
4542 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4543 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4544 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4545 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4547 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4548 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4549 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4551 #endif // FEATURE_SIMD
4552 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4553 GenTreePtr fgMorphCast(GenTreePtr tree);
4554 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4555 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4557 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4560 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4561 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4563 void fgFixupStructReturn(GenTreePtr call);
4564 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4565 bool fgAddrCouldBeNull(GenTreePtr addr);
4566 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4567 bool fgCanFastTailCall(GenTreeCall* call);
4568 void fgMorphTailCall(GenTreeCall* call);
4569 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4570 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4571 fgArgTabEntryPtr argTabEntry,
4573 IL_OFFSETX callILOffset,
4574 GenTreePtr tmpAssignmentInsertionPoint,
4575 GenTreePtr paramAssignmentInsertionPoint);
4576 static int fgEstimateCallStackSize(GenTreeCall* call);
4577 GenTreePtr fgMorphCall(GenTreeCall* call);
4578 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4579 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4581 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4582 static fgWalkPreFn fgFindNonInlineCandidate;
4584 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4585 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4586 void fgAssignSetVarDef(GenTreePtr tree);
4587 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4588 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4589 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4590 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4591 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4592 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4593 void fgMorphUnsafeBlk(GenTreeObj* obj);
4594 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4595 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4596 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4597 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4598 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4599 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4600 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4602 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4603 GenTreePtr fgMorphConst(GenTreePtr tree);
4606 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4609 #if LOCAL_ASSERTION_PROP
4610 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4611 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4613 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4615 GenTreeStmt* fgMorphStmt;
4617 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4618 // used when morphing big offset.
4620 //----------------------- Liveness analysis -------------------------------
4622 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4623 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4625 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4626 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4627 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4629 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4631 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4632 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4634 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4635 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4637 void fgExtendDbgScopes();
4638 void fgExtendDbgLifetimes();
4641 void fgDispDebugScopes();
4644 //-------------------------------------------------------------------------
4646 // The following keeps track of any code we've added for things like array
4647 // range checking or explicit calls to enable GC, and so on.
4652 AddCodeDsc* acdNext;
4653 BasicBlock* acdDstBlk; // block to which we jump
4655 SpecialCodeKind acdKind; // what kind of a special block is this?
4656 unsigned short acdStkLvl;
4660 static unsigned acdHelper(SpecialCodeKind codeKind);
4662 AddCodeDsc* fgAddCodeList;
4664 bool fgRngChkThrowAdded;
4665 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4667 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4669 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4672 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4675 bool fgIsCodeAdded();
4677 bool fgIsThrowHlpBlk(BasicBlock* block);
4678 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4680 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4682 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4683 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4684 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4685 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4686 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4688 #if FEATURE_MULTIREG_RET
4689 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4690 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4691 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4692 #endif // FEATURE_MULTIREG_RET
4694 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4697 static fgWalkPreFn fgDebugCheckInlineCandidates;
4700 void fgPromoteStructs();
4701 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4702 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4703 void fgMarkImplicitByRefArgs();
4704 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4705 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4706 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4707 void fgMarkAddressExposedLocals();
4708 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4710 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4712 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4714 // The given local variable, required to be a struct variable, is being assigned via
4715 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4716 // the variable is not enregistered, and is therefore not promoted independently.
4717 void fgLclFldAssign(unsigned lclNum);
4719 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4720 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4721 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4722 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4725 bool fgPrintInlinedMethods;
4728 bool fgIsBigOffset(size_t offset);
4730 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4731 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4732 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4733 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4734 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4737 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4742 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4743 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4750 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4753 void optRemoveRangeCheck(
4754 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4755 bool optIsRangeCheckRemovable(GenTreePtr tree);
4758 static fgWalkPreFn optValidRangeCheckIndex;
4759 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4762 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4764 /**************************************************************************
4766 *************************************************************************/
4769 // Do hoisting for all loops.
4770 void optHoistLoopCode();
4772 // To represent sets of VN's that have already been hoisted in outer loops.
4773 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4774 typedef VNToBoolMap VNSet;
4776 struct LoopHoistContext
4779 // The set of variables hoisted in the current loop (or nullptr if there are none).
4780 VNSet* m_pHoistedInCurLoop;
4783 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4784 VNSet m_hoistedInParentLoops;
4785 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4786 // Previous decisions on loop-invariance of value numbers in the current loop.
4787 VNToBoolMap m_curLoopVnInvariantCache;
4789 VNSet* GetHoistedInCurLoop(Compiler* comp)
4791 if (m_pHoistedInCurLoop == nullptr)
4793 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4795 return m_pHoistedInCurLoop;
4798 VNSet* ExtractHoistedInCurLoop()
4800 VNSet* res = m_pHoistedInCurLoop;
4801 m_pHoistedInCurLoop = nullptr;
4805 LoopHoistContext(Compiler* comp)
4806 : m_pHoistedInCurLoop(nullptr)
4807 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4808 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4813 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4814 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4815 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4816 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4818 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4819 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4820 // "m_hoistedInParentLoops".
4822 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4824 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4825 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4826 // expressions to "hoistInLoop".
4827 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4829 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4830 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4832 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4833 // that are invariant in loop "lnum" (an index into the optLoopTable)
4834 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4835 // expressions to "hoistInLoop".
4836 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4837 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4838 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4839 bool optHoistLoopExprsForTree(GenTreePtr tree,
4841 LoopHoistContext* hoistCtxt,
4842 bool* firstBlockAndBeforeSideEffect,
4845 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4846 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4848 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4849 // Constants and init values are always loop invariant.
4850 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4851 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4853 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4854 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4855 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4856 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4857 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4859 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4860 // in the loop table.
4861 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4863 // Records the set of "side effects" of all loops: fields (object instance and static)
4864 // written to, and SZ-array element type equivalence classes updated.
4865 void optComputeLoopSideEffects();
4868 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4869 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4870 // static) written to, and SZ-array element type equivalence classes updated.
4871 void optComputeLoopNestSideEffects(unsigned lnum);
4873 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4874 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4876 // Hoist the expression "expr" out of loop "lnum".
4877 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4880 void optOptimizeBools();
4883 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4885 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4888 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4890 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4891 // the loop into a "do-while" loop
4892 // Also finds all natural loops and records them in the loop table
4894 // Optionally clone loops in the loop table.
4895 void optCloneLoops();
4897 // Clone loop "loopInd" in the loop table.
4898 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4900 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4901 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4902 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4904 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4906 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4909 // This enumeration describes what is killed by a call.
4913 CALLINT_NONE, // no interference (most helpers)
4914 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4915 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4916 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4917 CALLINT_ALL, // kills everything (normal method call)
4921 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4922 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4923 // in bbNext order; we use comparisons on the bbNum to decide order.)
4924 // The blocks that define the body are
4925 // first <= top <= entry <= bottom .
4926 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4927 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4928 // Compiler::optFindNaturalLoops().
4931 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4932 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4933 // loop, but not the outer loop.)
4934 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4936 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4937 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4938 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4940 callInterf lpAsgCall; // "callInterf" for calls in the loop
4941 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4942 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4944 unsigned short lpFlags; // Mask of the LPFLG_* constants
4946 unsigned char lpExitCnt; // number of exits from the loop
4948 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4949 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4950 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4951 // (Actually, an "immediately" nested loop --
4952 // no other child of this loop is a parent of lpChild.)
4953 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4954 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4955 // by following "lpChild" then "lpSibling" links.
4957 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4958 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4960 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4961 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4962 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4964 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4965 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4967 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4968 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4969 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4970 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4972 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
4973 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
4974 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
4976 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
4977 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
4978 // type are assigned to.
4980 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
4981 // If this is set, the fields below may not be accurate (since they become irrelevant.)
4982 bool lpContainsCall; // True if executing the loop body *may* execute a call
4984 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
4985 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
4987 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
4989 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
4990 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
4992 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
4994 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
4995 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
4997 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
4998 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5000 JitSimplerHashBehavior>
5002 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5003 // instance fields modified
5006 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5007 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5009 JitSimplerHashBehavior>
5011 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5012 // arrays of that type are modified
5015 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5016 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5018 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5019 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5020 // (shifted left, with a low-order bit set to distinguish.)
5021 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5022 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5024 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5026 GenTreePtr lpIterTree; // The "i <op>= const" tree
5027 unsigned lpIterVar(); // iterator variable #
5028 int lpIterConst(); // the constant with which the iterator is incremented
5029 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5030 void VERIFY_lpIterTree();
5032 var_types lpIterOperType(); // For overflow instructions
5035 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5036 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5040 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5042 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5043 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5044 void VERIFY_lpTestTree();
5046 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5047 GenTreePtr lpIterator(); // the iterator node in the loop test
5048 GenTreePtr lpLimit(); // the limit node in the loop test
5050 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5051 // LPFLG_CONST_LIMIT
5052 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5054 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5055 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5056 // LPFLG_ARRLEN_LIMIT
5058 // Returns "true" iff "*this" contains the blk.
5059 bool lpContains(BasicBlock* blk)
5061 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5063 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5064 // to be equal, but requiring bottoms to be different.)
5065 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5067 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5070 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5071 // bottoms to be different.)
5072 bool lpContains(const LoopDsc& lp2)
5074 return lpContains(lp2.lpFirst, lp2.lpBottom);
5077 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5078 // (allowing firsts to be equal, but requiring bottoms to be different.)
5079 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5081 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5084 // Returns "true" iff "*this" is (properly) contained by "lp2"
5085 // (allowing firsts to be equal, but requiring bottoms to be different.)
5086 bool lpContainedBy(const LoopDsc& lp2)
5088 return lpContains(lp2.lpFirst, lp2.lpBottom);
5091 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5092 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5094 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5096 // Returns "true" iff "*this" is disjoint from "lp2".
5097 bool lpDisjoint(const LoopDsc& lp2)
5099 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5101 // Returns "true" iff the loop is well-formed (see code for defn).
5104 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5105 lpEntry->bbNum <= lpBottom->bbNum &&
5106 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5111 bool fgMightHaveLoop(); // returns true if there are any backedges
5112 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5115 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5116 unsigned char optLoopCount; // number of tracked loops
5119 unsigned optCallCount; // number of calls made in the method
5120 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5121 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5122 unsigned optLoopsCloned; // number of loops cloned in the current method.
5125 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5126 void optPrintLoopInfo(unsigned loopNum,
5128 BasicBlock* lpFirst,
5130 BasicBlock* lpEntry,
5131 BasicBlock* lpBottom,
5132 unsigned char lpExitCnt,
5134 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5135 void optPrintLoopInfo(unsigned lnum);
5136 void optPrintLoopRecording(unsigned lnum);
5138 void optCheckPreds();
5141 void optSetBlockWeights();
5143 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5145 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5147 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5149 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5150 unsigned optIsLoopIncrTree(GenTreePtr incr);
5151 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5152 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5153 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5154 bool optExtractInitTestIncr(BasicBlock* head,
5159 GenTreePtr* ppIncr);
5161 void optRecordLoop(BasicBlock* head,
5167 unsigned char exitCnt);
5169 void optFindNaturalLoops();
5171 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5172 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5173 bool optCanonicalizeLoopNest(unsigned char loopInd);
5175 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5176 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5177 bool optCanonicalizeLoop(unsigned char loopInd);
5179 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5180 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5181 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5182 bool optLoopContains(unsigned l1, unsigned l2);
5184 // Requires "loopInd" to be a valid index into the loop table.
5185 // Updates the loop table by changing loop "loopInd", whose head is required
5186 // to be "from", to be "to". Also performs this transformation for any
5187 // loop nested in "loopInd" that shares the same head as "loopInd".
5188 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5190 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5191 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5192 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5194 // Marks the containsCall information to "lnum" and any parent loops.
5195 void AddContainsCallAllContainingLoops(unsigned lnum);
5196 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5197 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5198 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5199 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5200 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5201 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5203 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5204 // of "from".) Copies the jump destination from "from" to "to".
5205 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5207 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5208 unsigned optLoopDepth(unsigned lnum)
5210 unsigned par = optLoopTable[lnum].lpParent;
5211 if (par == BasicBlock::NOT_IN_LOOP)
5217 return 1 + optLoopDepth(par);
5221 void fgOptWhileLoop(BasicBlock* block);
5223 bool optComputeLoopRep(int constInit,
5226 genTreeOps iterOper,
5228 genTreeOps testOper,
5231 unsigned* iterCount);
5232 #if FEATURE_STACK_FP_X87
5235 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5236 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5237 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5238 #endif // FEATURE_STACK_FP_X87
5241 static fgWalkPreFn optIsVarAssgCB;
5244 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5246 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5248 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5250 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5252 /**************************************************************************
5253 * Optimization conditions
5254 *************************************************************************/
5256 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5257 bool optPentium4(void);
5258 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5259 bool optAvoidIntMult(void);
5264 // The following is the upper limit on how many expressions we'll keep track
5265 // of for the CSE analysis.
5267 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5269 static const int MIN_CSE_COST = 2;
5271 // Keeps tracked cse indices
5272 BitVecTraits* cseTraits;
5276 /* Generic list of nodes - used by the CSE logic */
5284 typedef struct treeLst* treeLstPtr;
5288 treeStmtLst* tslNext;
5289 GenTreePtr tslTree; // tree node
5290 GenTreePtr tslStmt; // statement containing the tree
5291 BasicBlock* tslBlock; // block containing the statement
5294 typedef struct treeStmtLst* treeStmtLstPtr;
5296 // The following logic keeps track of expressions via a simple hash table.
5300 CSEdsc* csdNextInBucket; // used by the hash table
5302 unsigned csdHashValue; // the orginal hashkey
5304 unsigned csdIndex; // 1..optCSECandidateCount
5305 char csdLiveAcrossCall; // 0 or 1
5307 unsigned short csdDefCount; // definition count
5308 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5310 unsigned csdDefWtCnt; // weighted def count
5311 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5313 GenTreePtr csdTree; // treenode containing the 1st occurance
5314 GenTreePtr csdStmt; // stmt containing the 1st occurance
5315 BasicBlock* csdBlock; // block containing the 1st occurance
5317 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5318 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5321 static const size_t s_optCSEhashSize;
5322 CSEdsc** optCSEhash;
5327 CSEdsc* optCSEfindDsc(unsigned index);
5328 void optUnmarkCSE(GenTreePtr tree);
5330 // user defined callback data for the tree walk function optCSE_MaskHelper()
5331 struct optCSE_MaskData
5333 EXPSET_TP CSE_defMask;
5334 EXPSET_TP CSE_useMask;
5337 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5338 static fgWalkPreFn optCSE_MaskHelper;
5340 // This function walks all the node for an given tree
5341 // and return the mask of CSE definitions and uses for the tree
5343 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5345 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5346 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5347 bool optCSE_canSwap(GenTree* tree);
5349 static fgWalkPostFn optPropagateNonCSE;
5350 static fgWalkPreFn optHasNonCSEChild;
5352 static fgWalkPreFn optUnmarkCSEs;
5354 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5355 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5357 void optCleanupCSEs();
5360 void optEnsureClearCSEInfo();
5363 #endif // FEATURE_ANYCSE
5365 #if FEATURE_VALNUM_CSE
5366 /**************************************************************************
5367 * Value Number based CSEs
5368 *************************************************************************/
5371 void optOptimizeValnumCSEs();
5374 void optValnumCSE_Init();
5375 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5376 unsigned optValnumCSE_Locate();
5377 void optValnumCSE_InitDataFlow();
5378 void optValnumCSE_DataFlow();
5379 void optValnumCSE_Availablity();
5380 void optValnumCSE_Heuristic();
5381 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5383 #endif // FEATURE_VALNUM_CSE
5386 bool optDoCSE; // True when we have found a duplicate CSE tree
5387 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5388 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5389 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5390 unsigned optCSEstart; // The first local variable number that is a CSE
5391 unsigned optCSEcount; // The total count of CSE's introduced.
5392 unsigned optCSEweight; // The weight of the current block when we are
5393 // scanning for CSE expressions
5395 bool optIsCSEcandidate(GenTreePtr tree);
5397 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5399 bool lclNumIsTrueCSE(unsigned lclNum) const
5401 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5404 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5406 bool lclNumIsCSE(unsigned lclNum) const
5408 return lvaTable[lclNum].lvIsCSE;
5412 bool optConfigDisableCSE();
5413 bool optConfigDisableCSE2();
5415 void optOptimizeCSEs();
5417 #endif // FEATURE_ANYCSE
5425 unsigned ivaVar; // Variable we are interested in, or -1
5426 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5427 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5428 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5429 callInterf ivaMaskCall; // What kind of calls are there?
5432 static callInterf optCallInterf(GenTreePtr call);
5435 // VN based copy propagation.
5436 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5437 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5438 LclNumToGenTreePtrStack;
5440 // Kill set to track variables with intervening definitions.
5441 VARSET_TP optCopyPropKillSet;
5443 // Copy propagation functions.
5444 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5445 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5446 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5447 bool optIsSsaLocal(GenTreePtr tree);
5448 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5449 void optVnCopyProp();
5451 /**************************************************************************
5452 * Early value propagation
5453 *************************************************************************/
5459 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5463 static unsigned GetHashCode(SSAName ssaNm)
5465 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5468 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5470 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5474 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5475 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5476 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5477 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5478 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5480 unsigned optMethodFlags;
5482 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5483 // No throughput diff was found with backward walk bound between 3-8.
5484 static const int optEarlyPropRecurBound = 5;
5486 enum class optPropKind
5494 bool gtIsVtableRef(GenTreePtr tree);
5495 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5496 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5497 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5498 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5499 bool optEarlyPropRewriteTree(GenTreePtr tree);
5500 bool optDoEarlyPropForBlock(BasicBlock* block);
5501 bool optDoEarlyPropForFunc();
5502 void optEarlyProp();
5503 void optFoldNullCheck(GenTreePtr tree);
5504 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5507 /**************************************************************************
5508 * Value/Assertion propagation
5509 *************************************************************************/
5511 // Data structures for assertion prop
5512 BitVecTraits* apTraits;
5516 enum optAssertionKind
5531 O1K_ARRLEN_OPER_BND,
5532 O1K_ARRLEN_LOOP_BND,
5533 O1K_CONSTANT_LOOP_BND,
5554 optAssertionKind assertionKind;
5557 unsigned lclNum; // assigned to or property of this local var number
5565 struct AssertionDscOp1
5567 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5574 struct AssertionDscOp2
5576 optOp2Kind kind; // a const or copy assignment
5580 ssize_t iconVal; // integer
5581 unsigned iconFlags; // gtFlags
5583 struct Range // integer subrange
5597 bool IsArrLenArithBound()
5599 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5601 bool IsArrLenBound()
5603 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5605 bool IsConstantBound()
5607 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5608 op1.kind == O1K_CONSTANT_LOOP_BND);
5610 bool IsBoundsCheckNoThrow()
5612 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5615 bool IsCopyAssertion()
5617 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5620 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5622 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5623 a1->op2.kind == a2->op2.kind;
5626 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5628 if (kind == OAK_EQUAL)
5630 return kind2 == OAK_NOT_EQUAL;
5632 else if (kind == OAK_NOT_EQUAL)
5634 return kind2 == OAK_EQUAL;
5639 static ssize_t GetLowerBoundForIntegralType(var_types type)
5659 static ssize_t GetUpperBoundForIntegralType(var_types type)
5683 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5685 return (op1.kind == that->op1.kind) &&
5686 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5689 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5691 if (op2.kind != that->op2.kind)
5697 case O2K_IND_CNS_INT:
5699 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5701 case O2K_CONST_LONG:
5702 return (op2.lconVal == that->op2.lconVal);
5704 case O2K_CONST_DOUBLE:
5705 // exact match because of positive and negative zero.
5706 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5708 case O2K_LCLVAR_COPY:
5710 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5711 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5714 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5717 // we will return false
5721 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5727 bool Complementary(AssertionDsc* that, bool vnBased)
5729 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5730 HasSameOp2(that, vnBased);
5733 bool Equals(AssertionDsc* that, bool vnBased)
5735 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5739 typedef unsigned short AssertionIndex;
5742 static fgWalkPreFn optAddCopiesCallback;
5743 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5744 unsigned optAddCopyLclNum;
5745 GenTreePtr optAddCopyAsgnNode;
5747 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5748 bool optAssertionPropagated; // set to true if we modified the trees
5749 bool optAssertionPropagatedCurrentStmt;
5751 GenTreePtr optAssertionPropCurrentTree;
5753 AssertionIndex* optComplementaryAssertionMap;
5754 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5755 // using the value of a local var) for each local var
5756 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5757 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5758 AssertionIndex optMaxAssertionCount;
5761 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5762 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5763 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5764 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5765 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5766 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5768 AssertionIndex GetAssertionCount()
5770 return optAssertionCount;
5772 ASSERT_TP* bbJtrueAssertionOut;
5773 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5774 ValueNumToAssertsMap;
5775 ValueNumToAssertsMap* optValueNumToAsserts;
5777 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5779 // Assertion prop helpers.
5780 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5781 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5782 void optAssertionInit(bool isLocalProp);
5783 void optAssertionTraitsInit(AssertionIndex assertionCount);
5784 #if LOCAL_ASSERTION_PROP
5785 void optAssertionReset(AssertionIndex limit);
5786 void optAssertionRemove(AssertionIndex index);
5789 // Assertion prop data flow functions.
5790 void optAssertionPropMain();
5791 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5792 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5793 ASSERT_TP* optInitAssertionDataflowFlags();
5794 ASSERT_TP* optComputeAssertionGen();
5796 // Assertion Gen functions.
5797 void optAssertionGen(GenTreePtr tree);
5798 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5799 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5800 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5801 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5802 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5803 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5805 // Assertion creation functions.
5806 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5807 AssertionIndex optCreateAssertion(GenTreePtr op1,
5809 optAssertionKind assertionKind,
5810 AssertionDsc* assertion);
5811 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5813 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5814 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5815 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5817 void optPrintVnAssertionMapping();
5819 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5821 // Used for respective assertion propagations.
5822 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5823 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5824 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5825 bool optAssertionIsNonNull(GenTreePtr op,
5826 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5828 // Used for Relop propagation.
5829 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5830 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5831 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5833 // Assertion prop for lcl var functions.
5834 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5835 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5837 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5838 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5839 const GenTreePtr tree,
5840 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5841 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5843 // Assertion propagation functions.
5844 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5845 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5846 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5847 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5848 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5849 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5850 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5851 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5852 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5853 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5854 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5855 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5857 // Implied assertion functions.
5858 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5859 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5860 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5861 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5863 ASSERT_VALRET_TP optNewFullAssertSet();
5864 ASSERT_VALRET_TP optNewEmptyAssertSet();
5867 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5868 void optDebugCheckAssertion(AssertionDsc* assertion);
5869 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5871 void optAddCopies();
5872 #endif // ASSERTION_PROP
5874 /**************************************************************************
5876 *************************************************************************/
5879 struct LoopCloneVisitorInfo
5881 LoopCloneContext* context;
5884 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5885 : context(context), loopNum(loopNum), stmt(nullptr)
5890 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5891 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5892 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5893 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5894 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5895 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5896 void optObtainLoopCloningOpts(LoopCloneContext* context);
5897 bool optIsLoopClonable(unsigned loopInd);
5899 bool optCanCloneLoops();
5902 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5904 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5905 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5906 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5907 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5911 void optInsertLoopCloningStress(BasicBlock* head);
5913 #if COUNT_RANGECHECKS
5914 static unsigned optRangeChkRmv;
5915 static unsigned optRangeChkAll;
5924 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5929 RngChkDsc* rcdNextInBucket; // used by the hash table
5931 unsigned short rcdHashValue; // to make matching faster
5932 unsigned short rcdIndex; // 0..optRngChkCount-1
5934 GenTreePtr rcdTree; // the array index tree
5937 unsigned optRngChkCount;
5938 static const size_t optRngChkHashSize;
5940 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5941 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5943 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5946 bool optLoopsMarked;
5949 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5950 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5954 XX Does the register allocation and puts the remaining lclVars on the stack XX
5956 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5957 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5961 #ifndef LEGACY_BACKEND
5966 #else // LEGACY_BACKEND
5971 #endif // LEGACY_BACKEND
5973 #ifdef LEGACY_BACKEND
5975 void raAssignVars(); // register allocation
5976 #endif // LEGACY_BACKEND
5978 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
5980 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
5982 void raMarkStkVars();
5985 // Some things are used by both LSRA and regpredict allocators.
5987 FrameType rpFrameType;
5988 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
5990 #ifdef LEGACY_BACKEND
5991 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
5993 #endif // LEGACY_BACKEND
5995 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
5997 #if FEATURE_FP_REGALLOC
5998 enum enumConfigRegisterFP
6000 CONFIG_REGISTER_FP_NONE = 0x0,
6001 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6002 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6003 CONFIG_REGISTER_FP_FULL = 0x3,
6005 enumConfigRegisterFP raConfigRegisterFP();
6006 #endif // FEATURE_FP_REGALLOC
6009 regMaskTP raConfigRestrictMaskFP();
6012 #ifndef LEGACY_BACKEND
6013 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6014 #else // LEGACY_BACKEND
6015 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6016 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6017 bool raNewBlocks; // True is we added killing blocks for FPU registers
6018 unsigned rpPasses; // Number of passes made by the register predicter
6019 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6020 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6021 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6022 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6023 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6024 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6025 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6026 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6027 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6028 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6029 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6030 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6032 bool rpRegAllocDone; // Set to true after we have completed register allocation
6034 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6036 void raSetupArgMasks(RegState* r);
6038 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6040 void raDumpVarIntf(); // Dump the variable to variable interference graph
6041 void raDumpRegIntf(); // Dump the variable to register interference graph
6043 void raAdjustVarIntf();
6045 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6047 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6049 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6050 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6052 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6054 static fgWalkPreFn rpMarkRegIntf;
6056 regMaskTP rpPredictAddressMode(
6057 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6059 void rpPredictRefAssign(unsigned lclNum);
6061 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6063 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6065 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6067 void rpPredictRegUse(); // Entry point
6069 unsigned raPredictTreeRegUse(GenTreePtr tree);
6070 unsigned raPredictListRegUse(GenTreePtr list);
6072 void raSetRegVarOrder(var_types regType,
6073 regNumber* customVarOrder,
6074 unsigned* customVarOrderSize,
6076 regMaskTP avoidReg);
6078 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6079 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6080 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6081 void raAddToStkPredict(unsigned val)
6083 unsigned newStkPredict = rpStkPredict + val;
6084 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6085 rpStkPredict = UINT_MAX - 1;
6087 rpStkPredict = newStkPredict;
6091 #if !FEATURE_FP_REGALLOC
6092 void raDispFPlifeInfo();
6096 regMaskTP genReturnRegForTree(GenTreePtr tree);
6097 #endif // LEGACY_BACKEND
6099 /* raIsVarargsStackArg is called by raMaskStkVars and by
6100 lvaSortByRefCount. It identifies the special case
6101 where a varargs function has a parameter passed on the
6102 stack, other than the special varargs handle. Such parameters
6103 require special treatment, because they cannot be tracked
6104 by the GC (their offsets in the stack are not known
6108 bool raIsVarargsStackArg(unsigned lclNum)
6112 LclVarDsc* varDsc = &lvaTable[lclNum];
6114 assert(varDsc->lvIsParam);
6116 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6118 #else // _TARGET_X86_
6122 #endif // _TARGET_X86_
6125 #ifdef LEGACY_BACKEND
6126 // Records the current prediction, if it's better than any previous recorded prediction.
6127 void rpRecordPrediction();
6128 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6129 void rpUseRecordedPredictionIfBetter();
6131 // Data members used in the methods above.
6132 unsigned rpBestRecordedStkPredict;
6133 struct VarRegPrediction
6135 bool m_isEnregistered;
6136 regNumberSmall m_regNum;
6137 regNumberSmall m_otherReg;
6139 VarRegPrediction* rpBestRecordedPrediction;
6140 #endif // LEGACY_BACKEND
6143 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6144 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6148 XX Get to the class and method info from the Execution Engine given XX
6149 XX tokens for the class and method XX
6151 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6152 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6156 /* These are the different addressing modes used to access a local var.
6157 * The JIT has to report the location of the locals back to the EE
6158 * for debugging purposes.
6164 VLT_REG_BYREF, // this type is currently only used for value types on X64
6167 VLT_STK_BYREF, // this type is currently only used for value types on X64
6181 siVarLocType vlType;
6184 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6186 // VLT_REG_BYREF -- the specified register contains the address of the variable
6194 // VLT_STK -- Any 32 bit value which is on the stack
6195 // eg. [ESP+0x20], or [EBP-0x28]
6196 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6197 // eg. mov EAX, [ESP+0x20]; [EAX]
6201 regNumber vlsBaseReg;
6202 NATIVE_OFFSET vlsOffset;
6205 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6214 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6215 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6223 regNumber vlrssBaseReg;
6224 NATIVE_OFFSET vlrssOffset;
6228 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6229 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6235 regNumber vlsrsBaseReg;
6236 NATIVE_OFFSET vlsrsOffset;
6242 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6243 // eg 2 DWords at [ESP+0x10]
6247 regNumber vls2BaseReg;
6248 NATIVE_OFFSET vls2Offset;
6251 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6252 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6259 // VLT_FIXED_VA -- fixed argument of a varargs function.
6260 // The argument location depends on the size of the variable
6261 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6262 // location of the first arg. This argument can then be accessed
6263 // relative to the position of the first arg
6267 unsigned vlfvOffset;
6274 void* rpValue; // pointer to the in-process
6275 // location of the value.
6281 bool vlIsInReg(regNumber reg);
6282 bool vlIsOnStk(regNumber reg, signed offset);
6285 /*************************************************************************/
6290 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6291 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6292 CORINFO_CALLINFO_FLAGS flags,
6293 CORINFO_CALL_INFO* pResult);
6294 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6296 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6297 CORINFO_ACCESS_FLAGS flags,
6298 CORINFO_FIELD_INFO* pResult);
6302 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6304 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6306 bool IsSuperPMIException(unsigned code)
6308 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6310 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6311 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6312 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6313 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6314 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6315 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6316 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6317 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6321 case EXCEPTIONCODE_DebugBreakorAV:
6322 case EXCEPTIONCODE_MC:
6323 case EXCEPTIONCODE_LWM:
6324 case EXCEPTIONCODE_SASM:
6325 case EXCEPTIONCODE_SSYM:
6326 case EXCEPTIONCODE_CALLUTILS:
6327 case EXCEPTIONCODE_TYPEUTILS:
6328 case EXCEPTIONCODE_ASSERT:
6335 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6336 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6338 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6339 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6342 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6343 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6344 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6346 // VOM info, method sigs
6348 void eeGetSig(unsigned sigTok,
6349 CORINFO_MODULE_HANDLE scope,
6350 CORINFO_CONTEXT_HANDLE context,
6351 CORINFO_SIG_INFO* retSig);
6353 void eeGetCallSiteSig(unsigned sigTok,
6354 CORINFO_MODULE_HANDLE scope,
6355 CORINFO_CONTEXT_HANDLE context,
6356 CORINFO_SIG_INFO* retSig);
6358 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6360 // Method entry-points, instrs
6362 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6364 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6366 CORINFO_EE_INFO eeInfo;
6367 bool eeInfoInitialized;
6369 CORINFO_EE_INFO* eeGetEEInfo();
6371 // Gets the offset of a SDArray's first element
6372 unsigned eeGetArrayDataOffset(var_types type);
6373 // Gets the offset of a MDArray's first element
6374 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6376 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6378 // Returns the page size for the target machine as reported by the EE.
6379 inline size_t eeGetPageSize()
6381 #if COR_JIT_EE_VERSION > 460
6382 return eeGetEEInfo()->osPageSize;
6383 #else // COR_JIT_EE_VERSION <= 460
6384 return CORINFO_PAGE_SIZE;
6385 #endif // COR_JIT_EE_VERSION > 460
6388 // Returns the frame size at which we will generate a loop to probe the stack.
6389 inline size_t getVeryLargeFrameSize()
6392 // The looping probe code is 40 bytes, whereas the straight-line probing for
6393 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6394 // or greater, to generate smaller code.
6395 return 2 * eeGetPageSize();
6397 return 3 * eeGetPageSize();
6401 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6403 #if COR_JIT_EE_VERSION > 460
6404 return eeGetEEInfo()->targetAbi == abi;
6406 return CORINFO_DESKTOP_ABI == abi;
6410 inline bool generateCFIUnwindCodes()
6412 #ifdef UNIX_AMD64_ABI
6413 return IsTargetAbi(CORINFO_CORERT_ABI);
6421 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6423 // Debugging support - Line number info
6425 void eeGetStmtOffsets();
6427 unsigned eeBoundariesCount;
6429 struct boundariesDsc
6431 UNATIVE_OFFSET nativeIP;
6433 unsigned sourceReason;
6434 } * eeBoundaries; // Boundaries to report to EE
6435 void eeSetLIcount(unsigned count);
6436 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6440 static void eeDispILOffs(IL_OFFSET offs);
6441 static void eeDispLineInfo(const boundariesDsc* line);
6442 void eeDispLineInfos();
6445 // Debugging support - Local var info
6449 unsigned eeVarsCount;
6451 struct VarResultInfo
6453 UNATIVE_OFFSET startOffset;
6454 UNATIVE_OFFSET endOffset;
6458 void eeSetLVcount(unsigned count);
6459 void eeSetLVinfo(unsigned which,
6460 UNATIVE_OFFSET startOffs,
6461 UNATIVE_OFFSET length,
6466 const siVarLoc& loc);
6470 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6471 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6474 // ICorJitInfo wrappers
6476 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6478 void eeAllocUnwindInfo(BYTE* pHotCode,
6484 CorJitFuncKind funcKind);
6486 void eeSetEHcount(unsigned cEH);
6488 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6490 WORD eeGetRelocTypeHint(void* target);
6492 // ICorStaticInfo wrapper functions
6494 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6496 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6498 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6501 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6502 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6503 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6504 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6506 template <typename ParamType>
6507 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6509 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6512 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6514 // Utility functions
6516 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6519 const wchar_t* eeGetCPString(size_t stringHandle);
6522 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6524 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6525 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6527 static fgWalkPreFn CountSharedStaticHelper;
6528 static bool IsSharedStaticHelper(GenTreePtr tree);
6529 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6531 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6532 // returns true/false if 'field' is a Jit Data offset
6533 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6534 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6535 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6537 /*****************************************************************************/
6542 enum TEMP_USAGE_TYPE
6548 static var_types tmpNormalizeType(var_types type);
6549 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6550 void tmpRlsTemp(TempDsc* temp);
6551 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6554 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6555 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6559 bool tmpAllFree() const;
6562 #ifndef LEGACY_BACKEND
6563 void tmpPreAllocateTemps(var_types type, unsigned count);
6564 #endif // !LEGACY_BACKEND
6567 #ifdef LEGACY_BACKEND
6568 unsigned tmpIntSpillMax; // number of int-sized spill temps
6569 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6570 #endif // LEGACY_BACKEND
6572 unsigned tmpCount; // Number of temps
6573 unsigned tmpSize; // Size of all the temps
6576 // Used by RegSet::rsSpillChk()
6577 unsigned tmpGetCount; // Temps which haven't been released yet
6580 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6582 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6583 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6586 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6587 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6591 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6592 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6596 CodeGenInterface* codeGen;
6598 // The following holds information about instr offsets in terms of generated code.
6602 IPmappingDsc* ipmdNext; // next line# record
6603 IL_OFFSETX ipmdILoffsx; // the instr offset
6604 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6605 bool ipmdIsLabel; // Can this code be a branch label?
6608 // Record the instr offset mapping to the generated code
6610 IPmappingDsc* genIPmappingList;
6611 IPmappingDsc* genIPmappingLast;
6613 // Managed RetVal - A side hash table meant to record the mapping from a
6614 // GT_CALL node to its IL offset. This info is used to emit sequence points
6615 // that can be used by debugger to determine the native offset at which the
6616 // managed RetVal will be available.
6618 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6619 // favor of a side table for two reasons: 1) We need IL offset for only those
6620 // GT_CALL nodes (created during importation) that correspond to an IL call and
6621 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6622 // structure and IL offset is needed only when generating debuggable code. Therefore
6623 // it is desirable to avoid memory size penalty in retail scenarios.
6624 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6625 CallSiteILOffsetTable;
6626 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6628 unsigned genReturnLocal; // Local number for the return value when applicable.
6629 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6631 // The following properties are part of CodeGenContext. Getters are provided here for
6632 // convenience and backward compatibility, but the properties can only be set by invoking
6633 // the setter on CodeGenContext directly.
6635 __declspec(property(get = getEmitter)) emitter* genEmitter;
6636 emitter* getEmitter()
6638 return codeGen->getEmitter();
6641 const bool isFramePointerUsed()
6643 return codeGen->isFramePointerUsed();
6646 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6647 bool getInterruptible()
6649 return codeGen->genInterruptible;
6651 void setInterruptible(bool value)
6653 codeGen->setInterruptible(value);
6657 const bool genDoubleAlign()
6659 return codeGen->doDoubleAlign();
6661 DWORD getCanDoubleAlign();
6662 bool shouldDoubleAlign(unsigned refCntStk,
6664 unsigned refCntWtdReg,
6665 unsigned refCntStkParam,
6666 unsigned refCntWtdStkDbl);
6667 #endif // DOUBLE_ALIGN
6669 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6670 bool getFullPtrRegMap()
6672 return codeGen->genFullPtrRegMap;
6674 void setFullPtrRegMap(bool value)
6676 codeGen->setFullPtrRegMap(value);
6679 // Things that MAY belong either in CodeGen or CodeGenContext
6681 #if FEATURE_EH_FUNCLETS
6682 FuncInfoDsc* compFuncInfos;
6683 unsigned short compCurrFuncIdx;
6684 unsigned short compFuncInfoCount;
6686 unsigned short compFuncCount()
6688 assert(fgFuncletsCreated);
6689 return compFuncInfoCount;
6692 #else // !FEATURE_EH_FUNCLETS
6694 // This is a no-op when there are no funclets!
6695 void genUpdateCurrentFunclet(BasicBlock* block)
6700 FuncInfoDsc compFuncInfoRoot;
6702 static const unsigned compCurrFuncIdx = 0;
6704 unsigned short compFuncCount()
6709 #endif // !FEATURE_EH_FUNCLETS
6711 FuncInfoDsc* funCurrentFunc();
6712 void funSetCurrentFunc(unsigned funcIdx);
6713 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6714 unsigned int funGetFuncIdx(BasicBlock* block);
6718 VARSET_TP compCurLife; // current live variables
6719 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6721 template <bool ForCodeGen>
6722 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6724 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6726 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6729 template <bool ForCodeGen>
6730 void compUpdateLife(GenTreePtr tree);
6732 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6733 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6734 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6735 template <bool ForCodeGen>
6736 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6738 template <bool ForCodeGen>
6739 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6741 // Gets a register mask that represent the kill set for a helper call since
6742 // not all JIT Helper calls follow the standard ABI on the target architecture.
6743 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6745 // Gets a register mask that represent the kill set for a NoGC helper call.
6746 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6749 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6750 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6751 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6752 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6753 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6754 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6755 #endif // _TARGET_ARM_
6757 // 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
6759 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6761 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6762 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6763 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6764 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6765 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6766 // for the tracked var indices of the field vars, as in a live var set).
6767 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6769 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6771 if (m_promotedStructDeathVars == nullptr)
6773 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6775 return m_promotedStructDeathVars;
6779 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6780 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6784 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6785 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6788 #if !defined(__GNUC__)
6789 #pragma region Unwind information
6794 // Infrastructure functions: start/stop/reserve/emit.
6797 void unwindBegProlog();
6798 void unwindEndProlog();
6799 void unwindBegEpilog();
6800 void unwindEndEpilog();
6801 void unwindReserve();
6802 void unwindEmit(void* pHotCode, void* pColdCode);
6805 // Specific unwind information functions: called by code generation to indicate a particular
6806 // prolog or epilog unwindable instruction has been generated.
6809 void unwindPush(regNumber reg);
6810 void unwindAllocStack(unsigned size);
6811 void unwindSetFrameReg(regNumber reg, unsigned offset);
6812 void unwindSaveReg(regNumber reg, unsigned offset);
6814 #if defined(_TARGET_ARM_)
6815 void unwindPushMaskInt(regMaskTP mask);
6816 void unwindPushMaskFloat(regMaskTP mask);
6817 void unwindPopMaskInt(regMaskTP mask);
6818 void unwindPopMaskFloat(regMaskTP mask);
6819 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6820 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6821 // called via unwindPadding().
6822 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6823 // instruction and the current location.
6824 #endif // _TARGET_ARM_
6826 #if defined(_TARGET_ARM64_)
6828 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6829 // instruction and the current location.
6830 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6831 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6832 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6833 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6834 void unwindSaveNext(); // unwind code: save_next
6835 void unwindReturn(regNumber reg); // ret lr
6836 #endif // defined(_TARGET_ARM64_)
6839 // Private "helper" functions for the unwind implementation.
6843 #if FEATURE_EH_FUNCLETS
6844 void unwindGetFuncLocations(FuncInfoDsc* func,
6845 bool getHotSectionData,
6846 /* OUT */ emitLocation** ppStartLoc,
6847 /* OUT */ emitLocation** ppEndLoc);
6848 #endif // FEATURE_EH_FUNCLETS
6850 void unwindReserveFunc(FuncInfoDsc* func);
6851 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6853 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6855 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6856 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6858 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6860 #if defined(_TARGET_AMD64_)
6862 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6864 void unwindBegPrologWindows();
6865 void unwindPushWindows(regNumber reg);
6866 void unwindAllocStackWindows(unsigned size);
6867 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6868 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6870 #ifdef UNIX_AMD64_ABI
6871 void unwindBegPrologCFI();
6872 void unwindPushCFI(regNumber reg);
6873 void unwindAllocStackCFI(unsigned size);
6874 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6875 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6876 int mapRegNumToDwarfReg(regNumber reg);
6877 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6878 #endif // UNIX_AMD64_ABI
6879 #elif defined(_TARGET_ARM_)
6881 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6882 void unwindPushPopMaskFloat(regMaskTP mask);
6883 void unwindSplit(FuncInfoDsc* func);
6885 #endif // _TARGET_ARM_
6887 #if !defined(__GNUC__)
6888 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6892 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6893 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6897 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6898 XX that contains the distinguished, well-known SIMD type definitions). XX
6900 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6901 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6904 // Get highest available instruction set for floating point codegen
6905 InstructionSet getFloatingPointInstructionSet()
6907 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6910 return InstructionSet_AVX;
6915 return InstructionSet_SSE3_4;
6919 assert(canUseSSE2());
6920 return InstructionSet_SSE2;
6922 assert(!"getFPInstructionSet() is not implemented for target arch");
6924 return InstructionSet_NONE;
6928 // Get highest available instruction set for SIMD codegen
6929 InstructionSet getSIMDInstructionSet()
6931 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6932 return getFloatingPointInstructionSet();
6934 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6936 return InstructionSet_NONE;
6942 // Should we support SIMD intrinsics?
6945 // Have we identified any SIMD types?
6946 // This is currently used by struct promotion to avoid getting type information for a struct
6947 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
6949 bool _usesSIMDTypes;
6950 bool usesSIMDTypes()
6952 return _usesSIMDTypes;
6954 void setUsesSIMDTypes(bool value)
6956 _usesSIMDTypes = value;
6959 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6960 // that require indexed access to the individual fields of the vector, which is not well supported
6961 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6962 unsigned lvaSIMDInitTempVarNum;
6965 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6966 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6967 CORINFO_CLASS_HANDLE SIMDIntHandle;
6968 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6969 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6970 CORINFO_CLASS_HANDLE SIMDShortHandle;
6971 CORINFO_CLASS_HANDLE SIMDByteHandle;
6972 CORINFO_CLASS_HANDLE SIMDLongHandle;
6973 CORINFO_CLASS_HANDLE SIMDUIntHandle;
6974 CORINFO_CLASS_HANDLE SIMDULongHandle;
6975 CORINFO_CLASS_HANDLE SIMDVector2Handle;
6976 CORINFO_CLASS_HANDLE SIMDVector3Handle;
6977 CORINFO_CLASS_HANDLE SIMDVector4Handle;
6978 CORINFO_CLASS_HANDLE SIMDVectorHandle;
6980 // Get the handle for a SIMD type.
6981 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
6983 if (simdBaseType == TYP_FLOAT)
6988 return SIMDVector2Handle;
6990 return SIMDVector3Handle;
6992 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
6994 return SIMDVector4Handle;
7003 assert(simdType == getSIMDVectorType());
7004 switch (simdBaseType)
7007 return SIMDFloatHandle;
7009 return SIMDDoubleHandle;
7011 return SIMDIntHandle;
7013 return SIMDUShortHandle;
7015 return SIMDUShortHandle;
7017 return SIMDUByteHandle;
7019 return SIMDShortHandle;
7021 return SIMDByteHandle;
7023 return SIMDLongHandle;
7025 return SIMDUIntHandle;
7027 return SIMDULongHandle;
7029 assert(!"Didn't find a class handle for simdType");
7031 return NO_CLASS_HANDLE;
7035 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7036 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7037 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7039 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7040 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7041 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7042 bool isSIMDTypeLocal(GenTree* tree)
7044 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7047 // Returns true if the type of the tree is a byref of TYP_SIMD
7048 bool isAddrOfSIMDType(GenTree* tree)
7050 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7052 switch (tree->OperGet())
7055 return varTypeIsSIMD(tree->gtGetOp1());
7057 case GT_LCL_VAR_ADDR:
7058 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7061 return isSIMDTypeLocal(tree);
7068 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7070 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7071 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7072 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7075 // Returns base type of a TYP_SIMD local.
7076 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7077 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7079 if (isSIMDTypeLocal(tree))
7081 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7087 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7089 return info.compCompHnd->isInSIMDModule(clsHnd);
7092 bool isSIMDClass(typeInfo* pTypeInfo)
7094 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7097 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7098 // if it is not a SIMD type or is an unsupported base type.
7099 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7101 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7103 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7106 // Get SIMD Intrinsic info given the method handle.
7107 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7108 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7109 CORINFO_METHOD_HANDLE methodHnd,
7110 CORINFO_SIG_INFO* sig,
7113 var_types* baseType,
7114 unsigned* sizeBytes);
7116 // Pops and returns GenTree node from importers type stack.
7117 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7118 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7120 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7121 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7123 // Creates a GT_SIMD tree for Select operation
7124 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7126 unsigned simdVectorSize,
7131 // Creates a GT_SIMD tree for Min/Max operation
7132 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7133 CORINFO_CLASS_HANDLE typeHnd,
7135 unsigned simdVectorSize,
7139 // Transforms operands and returns the SIMD intrinsic to be applied on
7140 // transformed operands to obtain given relop result.
7141 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7142 CORINFO_CLASS_HANDLE typeHnd,
7143 unsigned simdVectorSize,
7144 var_types* baseType,
7148 // Creates a GT_SIMD tree for Abs intrinsic.
7149 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7151 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7152 // Transforms operands and returns the SIMD intrinsic to be applied on
7153 // transformed operands to obtain == comparison result.
7154 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7155 unsigned simdVectorSize,
7159 // Transforms operands and returns the SIMD intrinsic to be applied on
7160 // transformed operands to obtain > comparison result.
7161 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7162 unsigned simdVectorSize,
7166 // Transforms operands and returns the SIMD intrinsic to be applied on
7167 // transformed operands to obtain >= comparison result.
7168 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7169 unsigned simdVectorSize,
7173 // Transforms operands and returns the SIMD intrinsic to be applied on
7174 // transformed operands to obtain >= comparison result in case of int32
7175 // and small int base type vectors.
7176 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7177 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7178 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7180 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7181 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7182 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7183 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7184 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7186 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7187 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7188 GenTreePtr newobjThis,
7189 CORINFO_CLASS_HANDLE clsHnd,
7190 CORINFO_METHOD_HANDLE method,
7191 CORINFO_SIG_INFO* sig,
7194 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7196 // Whether SIMD vector occupies part of SIMD register.
7197 // SSE2: vector2f/3f are considered sub register SIMD types.
7198 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7199 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7201 unsigned sizeBytes = 0;
7202 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7203 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7206 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7208 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7211 // Get the type for the hardware SIMD vector.
7212 // This is the maximum SIMD type supported for this target.
7213 var_types getSIMDVectorType()
7215 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7222 assert(canUseSSE2());
7226 assert(!"getSIMDVectorType() unimplemented on target arch");
7231 // Get the size of the SIMD type in bytes
7232 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7234 unsigned sizeBytes = 0;
7235 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7239 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7240 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7242 // Get the the number of elements of basetype of SIMD vector given by its type handle
7243 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7245 // Get preferred alignment of SIMD type.
7246 int getSIMDTypeAlignment(var_types simdType);
7248 // Get the number of bytes in a SIMD Vector for the current compilation.
7249 unsigned getSIMDVectorRegisterByteLength()
7251 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7254 return YMM_REGSIZE_BYTES;
7258 assert(canUseSSE2());
7259 return XMM_REGSIZE_BYTES;
7262 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7267 // The minimum and maximum possible number of bytes in a SIMD vector.
7268 unsigned int maxSIMDStructBytes()
7270 return getSIMDVectorRegisterByteLength();
7272 unsigned int minSIMDStructBytes()
7274 return emitTypeSize(TYP_SIMD8);
7277 #ifdef FEATURE_AVX_SUPPORT
7278 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7279 static const unsigned maxPossibleSIMDStructBytes = 32;
7280 #else // !FEATURE_AVX_SUPPORT
7281 static const unsigned maxPossibleSIMDStructBytes = 16;
7282 #endif // !FEATURE_AVX_SUPPORT
7284 // Returns the codegen type for a given SIMD size.
7285 var_types getSIMDTypeForSize(unsigned size)
7287 var_types simdType = TYP_UNDEF;
7290 simdType = TYP_SIMD8;
7292 else if (size == 12)
7294 simdType = TYP_SIMD12;
7296 else if (size == 16)
7298 simdType = TYP_SIMD16;
7300 #ifdef FEATURE_AVX_SUPPORT
7301 else if (size == 32)
7303 simdType = TYP_SIMD32;
7305 #endif // FEATURE_AVX_SUPPORT
7308 noway_assert(!"Unexpected size for SIMD type");
7313 unsigned getSIMDInitTempVarNum()
7315 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7317 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7318 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7320 return lvaSIMDInitTempVarNum;
7323 #endif // FEATURE_SIMD
7326 //------------------------------------------------------------------------
7327 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7329 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7330 // candidate for enregistration.
7332 unsigned largestEnregisterableStructSize()
7335 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7336 if (vectorRegSize > TARGET_POINTER_SIZE)
7338 return vectorRegSize;
7341 #endif // FEATURE_SIMD
7343 return TARGET_POINTER_SIZE;
7348 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7349 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7350 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7352 // Is this var is of type simd struct?
7353 bool lclVarIsSIMDType(unsigned varNum)
7355 LclVarDsc* varDsc = lvaTable + varNum;
7356 return varDsc->lvIsSIMDType();
7359 // Is this Local node a SIMD local?
7360 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7362 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7365 // Returns true if the TYP_SIMD locals on stack are aligned at their
7366 // preferred byte boundary specified by getSIMDTypeAlignment().
7368 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7369 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7370 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7371 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7372 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7373 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7374 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7377 bool isSIMDTypeLocalAligned(unsigned varNum)
7379 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7380 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7383 int off = lvaFrameAddress(varNum, &ebpBased);
7384 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7385 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7386 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7389 #endif // FEATURE_SIMD
7394 // Whether SSE2 is available
7395 bool canUseSSE2() const
7397 #ifdef _TARGET_XARCH_
7398 return opts.compCanUseSSE2;
7404 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7405 bool CanUseSSE3_4() const
7407 #ifdef _TARGET_XARCH_
7408 return opts.compCanUseSSE3_4;
7414 bool canUseAVX() const
7416 #ifdef FEATURE_AVX_SUPPORT
7417 return opts.compCanUseAVX;
7424 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7425 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7429 XX Generic info about the compilation and the method being compiled. XX
7430 XX It is responsible for driving the other phases. XX
7431 XX It is also responsible for all the memory management. XX
7433 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7434 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7438 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7440 InlineResult* compInlineResult; // The result of importing the inlinee method.
7442 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7443 bool compJmpOpUsed; // Does the method do a JMP
7444 bool compLongUsed; // Does the method use TYP_LONG
7445 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7446 bool compTailCallUsed; // Does the method do a tailcall
7447 bool compLocallocUsed; // Does the method use localloc.
7448 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7449 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7450 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7452 // NOTE: These values are only reliable after
7453 // the importing is completely finished.
7455 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7456 // we can iterate over these efficiently.
7458 #if CPU_USES_BLOCK_MOVE
7459 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7463 // State information - which phases have completed?
7464 // These are kept together for easy discoverability
7466 bool bRangeAllowStress;
7467 bool compCodeGenDone;
7468 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7469 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7470 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7471 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7474 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7475 bool fgLocalVarLivenessChanged;
7477 bool compStackProbePrologDone;
7479 #ifndef LEGACY_BACKEND
7481 #endif // !LEGACY_BACKEND
7482 bool compRationalIRForm;
7484 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7486 bool compGeneratingProlog;
7487 bool compGeneratingEpilog;
7488 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7489 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7490 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7491 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7492 bool getNeedsGSSecurityCookie() const
7494 return compNeedsGSSecurityCookie;
7496 void setNeedsGSSecurityCookie()
7498 compNeedsGSSecurityCookie = true;
7501 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7502 // frame layout calculations, this is the level we are currently
7505 //---------------------------- JITing options -----------------------------
7518 JitFlags* jitFlags; // all flags passed from the EE
7519 unsigned compFlags; // method attributes
7521 codeOptimize compCodeOpt; // what type of code optimizations
7525 #ifdef _TARGET_XARCH_
7526 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7527 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7529 #ifdef FEATURE_AVX_SUPPORT
7530 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7531 #endif // FEATURE_AVX_SUPPORT
7532 #endif // _TARGET_XARCH_
7534 // optimize maximally and/or favor speed over size?
7536 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7537 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7538 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7539 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7540 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7542 // Maximun number of locals before turning off the inlining
7543 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7546 unsigned instrCount;
7547 unsigned lvRefCount;
7548 bool compMinOptsIsSet;
7550 bool compMinOptsIsUsed;
7552 inline bool MinOpts()
7554 assert(compMinOptsIsSet);
7555 compMinOptsIsUsed = true;
7558 inline bool IsMinOptsSet()
7560 return compMinOptsIsSet;
7563 inline bool MinOpts()
7567 inline bool IsMinOptsSet()
7569 return compMinOptsIsSet;
7572 inline void SetMinOpts(bool val)
7574 assert(!compMinOptsIsUsed);
7575 assert(!compMinOptsIsSet || (compMinOpts == val));
7577 compMinOptsIsSet = true;
7580 // true if the CLFLG_* for an optimization is set.
7581 inline bool OptEnabled(unsigned optFlag)
7583 return !!(compFlags & optFlag);
7586 #ifdef FEATURE_READYTORUN_COMPILER
7587 inline bool IsReadyToRun()
7589 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7592 inline bool IsReadyToRun()
7598 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7599 // PInvoke transitions inline (e.g. when targeting CoreRT).
7600 inline bool ShouldUsePInvokeHelpers()
7602 #if COR_JIT_EE_VERSION > 460
7603 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7609 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7611 inline bool IsReversePInvoke()
7613 #if COR_JIT_EE_VERSION > 460
7614 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7620 // true if we must generate code compatible with JIT32 quirks
7621 inline bool IsJit32Compat()
7623 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7624 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7630 // true if we must generate code compatible with Jit64 quirks
7631 inline bool IsJit64Compat()
7633 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7634 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7635 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7642 bool compScopeInfo; // Generate the LocalVar info ?
7643 bool compDbgCode; // Generate debugger-friendly code?
7644 bool compDbgInfo; // Gather debugging info?
7647 #ifdef PROFILING_SUPPORTED
7648 bool compNoPInvokeInlineCB;
7650 static const bool compNoPInvokeInlineCB;
7654 bool compGcChecks; // Check arguments and return values to ensure they are sane
7655 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7656 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7660 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7661 // to be allocated on the stack.
7662 // It will be set to true in the following cases:
7663 // 1. When the method being compiled has a declarative security
7664 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7665 // This is also the case when we inject a prolog and epilog in the method.
7667 // 2. When the method being compiled has imperative security (i.e. the method
7668 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7670 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7672 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7673 // which gets reported as a GC root to stackwalker.
7674 // (See also ICodeManager::GetAddrOfSecurityObject.)
7681 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7682 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7686 #ifdef UNIX_AMD64_ABI
7687 // This flag is indicating if there is a need to align the frame.
7688 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7689 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7690 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7691 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7692 // there are calls and making sure the frame alignment logic is executed.
7693 bool compNeedToAlignFrame;
7694 #endif // UNIX_AMD64_ABI
7696 bool compProcedureSplitting; // Separate cold code from hot code
7698 bool genFPorder; // Preserve FP order (operations are non-commutative)
7699 bool genFPopt; // Can we do frame-pointer-omission optimization?
7700 bool altJit; // True if we are an altjit and are compiling this method
7703 bool optRepeat; // Repeat optimizer phases k times
7704 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7705 bool dspCode; // Display native code generated
7706 bool dspEHTable; // Display the EH table reported to the VM
7707 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7708 bool dspEmit; // Display emitter output
7709 bool dspLines; // Display source-code lines intermixed with native code output
7710 bool dmpHex; // Display raw bytes in hex of native code output
7711 bool varNames; // Display variables names in native code output
7712 bool disAsm; // Display native code as it is generated
7713 bool disAsmSpilled; // Display native code when any register spilling occurs
7714 bool disDiffable; // Makes the Disassembly code 'diff-able'
7715 bool disAsm2; // Display native code after it is generated using external disassembler
7716 bool dspOrder; // Display names of each of the methods that we ngen/jit
7717 bool dspUnwind; // Display the unwind info output
7718 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7719 bool compLongAddress; // Force using large pseudo instructions for long address
7720 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7721 bool dspGCtbls; // Display the GC tables
7725 bool doLateDisasm; // Run the late disassembler
7726 #endif // LATE_DISASM
7728 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7729 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7730 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7731 static const bool dspGCtbls = true;
7734 // We need stack probes to guarantee that we won't trigger a stack overflow
7735 // when calling unmanaged code until they get a chance to set up a frame, because
7736 // the EE will have no idea where it is.
7738 // We will only be doing this currently for hosted environments. Unfortunately
7739 // we need to take care of stubs, so potentially, we will have to do the probes
7740 // for any call. We have a plan for not needing for stubs though
7741 bool compNeedStackProbes;
7743 #ifdef PROFILING_SUPPORTED
7744 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7745 // This option helps make the JIT behave as if it is running under a profiler.
7746 bool compJitELTHookEnabled;
7747 #endif // PROFILING_SUPPORTED
7749 #if FEATURE_TAILCALL_OPT
7750 // Whether opportunistic or implicit tail call optimization is enabled.
7751 bool compTailCallOpt;
7752 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7753 bool compTailCallLoopOpt;
7757 static const bool compUseSoftFP = true;
7758 #else // !ARM_SOFTFP
7759 static const bool compUseSoftFP = false;
7762 GCPollType compGCPollType;
7766 static bool s_pAltJitExcludeAssembliesListInitialized;
7767 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7772 template <typename T>
7775 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7778 template <typename T>
7781 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7784 static int dspTreeID(GenTree* tree)
7786 return tree->gtTreeID;
7788 static void printTreeID(GenTree* tree)
7790 if (tree == nullptr)
7796 printf("[%06d]", dspTreeID(tree));
7803 #define STRESS_MODES \
7807 /* "Variations" stress areas which we try to mix up with each other. */ \
7808 /* These should not be exhaustively used as they might */ \
7809 /* hide/trivialize other areas */ \
7811 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7812 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7813 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7814 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7815 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7816 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7817 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7818 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7819 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7820 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7821 STRESS_MODE(NULL_OBJECT_CHECK) \
7822 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7823 STRESS_MODE(RANDOM_INLINE) \
7825 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7827 /* "Check" stress areas that can be exhaustively used if we */ \
7828 /* dont care about performance at all */ \
7830 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7831 STRESS_MODE(CHK_FLOW_UPDATE) \
7832 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7834 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7838 #define STRESS_MODE(mode) STRESS_##mode,
7845 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7846 BYTE compActiveStressModes[STRESS_COUNT];
7849 #define MAX_STRESS_WEIGHT 100
7851 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7855 bool compInlineStress()
7857 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7860 bool compRandomInlineStress()
7862 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7867 bool compTailCallStress()
7870 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7876 codeOptimize compCodeOpt()
7879 // Switching between size & speed has measurable throughput impact
7880 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7881 // DEBUG, but should generate identical code between CHK & RET builds,
7882 // so that's not acceptable.
7883 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7884 // Investigate the cause of the throughput regression.
7886 return opts.compCodeOpt;
7888 return BLENDED_CODE;
7892 //--------------------- Info about the procedure --------------------------
7896 COMP_HANDLE compCompHnd;
7897 CORINFO_MODULE_HANDLE compScopeHnd;
7898 CORINFO_CLASS_HANDLE compClassHnd;
7899 CORINFO_METHOD_HANDLE compMethodHnd;
7900 CORINFO_METHOD_INFO* compMethodInfo;
7902 BOOL hasCircularClassConstraints;
7903 BOOL hasCircularMethodConstraints;
7905 #if defined(DEBUG) || defined(LATE_DISASM)
7906 const char* compMethodName;
7907 const char* compClassName;
7908 const char* compFullName;
7909 #endif // defined(DEBUG) || defined(LATE_DISASM)
7911 #if defined(DEBUG) || defined(INLINE_DATA)
7912 // Method hash is logcally const, but computed
7914 mutable unsigned compMethodHashPrivate;
7915 unsigned compMethodHash() const;
7916 #endif // defined(DEBUG) || defined(INLINE_DATA)
7918 #ifdef PSEUDORANDOM_NOP_INSERTION
7919 // things for pseudorandom nop insertion
7920 unsigned compChecksum;
7924 // The following holds the FLG_xxxx flags for the method we're compiling.
7927 // The following holds the class attributes for the method we're compiling.
7928 unsigned compClassAttr;
7930 const BYTE* compCode;
7931 IL_OFFSET compILCodeSize; // The IL code size
7932 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7933 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7934 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7935 // (2) the code is hot/cold split, and we issued less code than we expected
7936 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7938 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7939 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7940 bool compIsContextful : 1; // contextful method
7941 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7942 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7943 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7944 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7945 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7947 var_types compRetType; // Return type of the method as declared in IL
7948 var_types compRetNativeType; // Normalized return type as per target arch ABI
7949 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7950 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7951 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7952 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7953 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7954 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7955 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7956 unsigned compMaxStack;
7957 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7958 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7960 unsigned compCallUnmanaged; // count of unmanaged calls
7961 unsigned compLvFrameListRoot; // lclNum for the Frame root
7962 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7963 // You should generally use compHndBBtabCount instead: it is the
7964 // current number of EH clauses (after additions like synchronized
7965 // methods and funclets, and removals like unreachable code deletion).
7967 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7968 // and the VM expects that, or the JIT is a "self-host" compiler
7969 // (e.g., x86 hosted targeting x86) and the VM expects that.
7971 /* The following holds IL scope information about local variables.
7974 unsigned compVarScopesCount;
7975 VarScopeDsc* compVarScopes;
7977 /* The following holds information about instr offsets for
7978 * which we need to report IP-mappings
7981 IL_OFFSET* compStmtOffsets; // sorted
7982 unsigned compStmtOffsetsCount;
7983 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
7985 #define CPU_X86 0x0100 // The generic X86 CPU
7986 #define CPU_X86_PENTIUM_4 0x0110
7988 #define CPU_X64 0x0200 // The generic x64 CPU
7989 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
7990 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
7992 #define CPU_ARM 0x0300 // The generic ARM CPU
7994 unsigned genCPU; // What CPU are we running on
7997 // Returns true if the method being compiled returns a non-void and non-struct value.
7998 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
7999 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8000 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8001 // Methods returning such structs are considered to return non-struct return value and
8002 // this method returns true in that case.
8003 bool compMethodReturnsNativeScalarType()
8005 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8008 // Returns true if the method being compiled returns RetBuf addr as its return value
8009 bool compMethodReturnsRetBufAddr()
8011 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8012 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8014 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8015 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8016 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8017 // methods with hidden RetBufArg.
8019 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8020 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8021 // returning the address of RetBuf.
8023 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8024 // to be returned in RAX.
8025 CLANG_FORMAT_COMMENT_ANCHOR;
8027 #ifdef _TARGET_AMD64_
8028 return (info.compRetBuffArg != BAD_VAR_NUM);
8029 #else // !_TARGET_AMD64_
8030 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8031 #endif // !_TARGET_AMD64_
8034 // Returns true if the method returns a value in more than one return register
8035 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8036 // TODO-ARM64: Does this apply for ARM64 too?
8037 bool compMethodReturnsMultiRegRetType()
8039 #if FEATURE_MULTIREG_RET
8040 #if defined(_TARGET_X86_)
8041 // On x86 only 64-bit longs are returned in multiple registers
8042 return varTypeIsLong(info.compRetNativeType);
8043 #else // targets: X64-UNIX, ARM64 or ARM32
8044 // On all other targets that support multireg return values:
8045 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8046 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8047 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8048 #endif // TARGET_XXX
8050 #else // not FEATURE_MULTIREG_RET
8052 // For this architecture there are no multireg returns
8055 #endif // FEATURE_MULTIREG_RET
8058 #if FEATURE_MULTIREG_ARGS
8059 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8060 // return the gcPtr layout for the pointers sized fields
8061 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8062 #endif // FEATURE_MULTIREG_ARGS
8064 // Returns true if the method being compiled returns a value
8065 bool compMethodHasRetVal()
8067 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8068 compMethodReturnsMultiRegRetType();
8073 void compDispLocalVars();
8077 //-------------------------- Global Compiler Data ------------------------------------
8080 static unsigned s_compMethodsCount; // to produce unique label names
8081 unsigned compGenTreeID;
8084 BasicBlock* compCurBB; // the current basic block in process
8085 GenTreePtr compCurStmt; // the current statement in process
8087 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8090 // The following is used to create the 'method JIT info' block.
8091 size_t compInfoBlkSize;
8092 BYTE* compInfoBlkAddr;
8094 EHblkDsc* compHndBBtab; // array of EH data
8095 unsigned compHndBBtabCount; // element count of used elements in EH data array
8096 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8098 #if defined(_TARGET_X86_)
8100 //-------------------------------------------------------------------------
8101 // Tracking of region covered by the monitor in synchronized methods
8102 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8103 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8105 #endif // !_TARGET_X86_
8107 Phases previousCompletedPhase; // the most recently completed phase
8109 //-------------------------------------------------------------------------
8110 // The following keeps track of how many bytes of local frame space we've
8111 // grabbed so far in the current function, and how many argument bytes we
8112 // need to pop when we return.
8115 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8117 // Count of callee-saved regs we pushed in the prolog.
8118 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8119 // In case of Amd64 this doesn't include float regs saved on stack.
8120 unsigned compCalleeRegsPushed;
8122 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8123 // Mask of callee saved float regs on stack.
8124 regMaskTP compCalleeFPRegsSavedMask;
8126 #ifdef _TARGET_AMD64_
8127 // Quirk for VS debug-launch scenario to work:
8128 // Bytes of padding between save-reg area and locals.
8129 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8130 unsigned compVSQuirkStackPaddingNeeded;
8131 bool compQuirkForPPPflag;
8134 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8136 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8137 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8138 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8140 //-------------------------------------------------------------------------
8142 static void compStartup(); // One-time initialization
8143 static void compShutdown(); // One-time finalization
8145 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8148 static void compDisplayStaticSizes(FILE* fout);
8150 //------------ Some utility functions --------------
8152 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8153 void** ppIndirection); /* OUT */
8155 // Several JIT/EE interface functions return a CorInfoType, and also return a
8156 // class handle as an out parameter if the type is a value class. Returns the
8157 // size of the type these describe.
8158 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8161 // Components used by the compiler may write unit test suites, and
8162 // have them run within this method. They will be run only once per process, and only
8163 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8164 // These should fail by asserting.
8165 void compDoComponentUnitTestsOnce();
8168 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8169 CORINFO_MODULE_HANDLE classPtr,
8170 COMP_HANDLE compHnd,
8171 CORINFO_METHOD_INFO* methodInfo,
8172 void** methodCodePtr,
8173 ULONG* methodCodeSize,
8174 JitFlags* compileFlags);
8175 void compCompileFinish();
8176 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8177 COMP_HANDLE compHnd,
8178 CORINFO_METHOD_INFO* methodInfo,
8179 void** methodCodePtr,
8180 ULONG* methodCodeSize,
8181 JitFlags* compileFlags,
8182 CorInfoInstantiationVerification instVerInfo);
8184 ArenaAllocator* compGetAllocator();
8186 #if MEASURE_MEM_ALLOC
8188 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8192 unsigned allocCnt; // # of allocs
8193 UINT64 allocSz; // total size of those alloc.
8194 UINT64 allocSzMax; // Maximum single allocation.
8195 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8196 UINT64 nraTotalSizeAlloc;
8197 UINT64 nraTotalSizeUsed;
8199 static const char* s_CompMemKindNames[]; // Names of the kinds.
8201 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8203 for (int i = 0; i < CMK_Count; i++)
8205 allocSzByKind[i] = 0;
8208 MemStats(const MemStats& ms)
8209 : allocCnt(ms.allocCnt)
8210 , allocSz(ms.allocSz)
8211 , allocSzMax(ms.allocSzMax)
8212 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8213 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8215 for (int i = 0; i < CMK_Count; i++)
8217 allocSzByKind[i] = ms.allocSzByKind[i];
8221 // Until we have ubiquitous constructors.
8224 this->MemStats::MemStats();
8227 void AddAlloc(size_t sz, CompMemKind cmk)
8231 if (sz > allocSzMax)
8235 allocSzByKind[cmk] += sz;
8238 void Print(FILE* f); // Print these stats to f.
8239 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8241 MemStats genMemStats;
8243 struct AggregateMemStats : public MemStats
8247 AggregateMemStats() : MemStats(), nMethods(0)
8251 void Add(const MemStats& ms)
8254 allocCnt += ms.allocCnt;
8255 allocSz += ms.allocSz;
8256 allocSzMax = max(allocSzMax, ms.allocSzMax);
8257 for (int i = 0; i < CMK_Count; i++)
8259 allocSzByKind[i] += ms.allocSzByKind[i];
8261 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8262 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8265 void Print(FILE* f); // Print these stats to jitstdout.
8268 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8269 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8270 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8272 #endif // MEASURE_MEM_ALLOC
8274 #if LOOP_HOIST_STATS
8275 unsigned m_loopsConsidered;
8276 bool m_curLoopHasHoistedExpression;
8277 unsigned m_loopsWithHoistedExpressions;
8278 unsigned m_totalHoistedExpressions;
8280 void AddLoopHoistStats();
8281 void PrintPerMethodLoopHoistStats();
8283 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8284 static unsigned s_loopsConsidered;
8285 static unsigned s_loopsWithHoistedExpressions;
8286 static unsigned s_totalHoistedExpressions;
8288 static void PrintAggregateLoopHoistStats(FILE* f);
8289 #endif // LOOP_HOIST_STATS
8291 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8292 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8293 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8294 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8295 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8296 void compFreeMem(void*);
8298 bool compIsForImportOnly();
8299 bool compIsForInlining();
8300 bool compDonotInline();
8303 const char* compLocalVarName(unsigned varNum, unsigned offs);
8304 VarName compVarName(regNumber reg, bool isFloatReg = false);
8305 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8306 const char* compRegPairName(regPairNo regPair);
8307 const char* compRegNameForSize(regNumber reg, size_t size);
8308 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8309 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8310 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8313 //-------------------------------------------------------------------------
8315 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8317 struct VarScopeMapInfo
8319 VarScopeListNode* head;
8320 VarScopeListNode* tail;
8321 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8323 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8330 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8331 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8333 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8334 VarNumToScopeDscMap;
8336 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8337 VarNumToScopeDscMap* compVarScopeMap;
8339 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8341 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8343 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8345 void compInitVarScopeMap();
8347 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8348 // enter scope, sorted by instr offset
8349 unsigned compNextEnterScope;
8351 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8352 // go out of scope, sorted by instr offset
8353 unsigned compNextExitScope;
8355 void compInitScopeLists();
8357 void compResetScopeLists();
8359 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8361 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8363 void compProcessScopesUntil(unsigned offset,
8365 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8366 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8369 void compDispScopeLists();
8372 bool compIsProfilerHookNeeded();
8374 //-------------------------------------------------------------------------
8375 /* Statistical Data Gathering */
8377 void compJitStats(); // call this function and enable
8378 // various ifdef's below for statistical data
8381 void compCallArgStats();
8382 static void compDispCallArgStats(FILE* fout);
8385 //-------------------------------------------------------------------------
8392 ArenaAllocator* compAllocator;
8395 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8396 // suitable for use by utilcode collection types.
8397 IAllocator* compAsIAllocator;
8399 #if MEASURE_MEM_ALLOC
8400 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8401 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8402 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8404 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8406 #endif // MEASURE_MEM_ALLOC
8408 void compFunctionTraceStart();
8409 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8412 size_t compMaxUncheckedOffsetForNullObject;
8414 void compInitOptions(JitFlags* compileFlags);
8416 void compSetProcessor();
8417 void compInitDebuggingInfo();
8418 void compSetOptimizationLevel();
8419 #ifdef _TARGET_ARMARCH_
8420 bool compRsvdRegCheck(FrameLayoutState curState);
8422 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8424 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8425 void ResetOptAnnotations();
8427 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8428 void RecomputeLoopInfo();
8430 #ifdef PROFILING_SUPPORTED
8431 // Data required for generating profiler Enter/Leave/TailCall hooks
8433 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8434 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8435 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8438 #ifdef _TARGET_AMD64_
8439 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8442 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8443 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8445 IAllocator* getAllocator()
8447 return compAsIAllocator;
8450 #if MEASURE_MEM_ALLOC
8451 IAllocator* getAllocatorBitset()
8453 return compAsIAllocatorBitset;
8455 IAllocator* getAllocatorGC()
8457 return compAsIAllocatorGC;
8459 IAllocator* getAllocatorLoopHoist()
8461 return compAsIAllocatorLoopHoist;
8463 #else // !MEASURE_MEM_ALLOC
8464 IAllocator* getAllocatorBitset()
8466 return compAsIAllocator;
8468 IAllocator* getAllocatorGC()
8470 return compAsIAllocator;
8472 IAllocator* getAllocatorLoopHoist()
8474 return compAsIAllocator;
8476 #endif // !MEASURE_MEM_ALLOC
8479 IAllocator* getAllocatorDebugOnly()
8481 #if MEASURE_MEM_ALLOC
8482 return compAsIAllocatorDebugOnly;
8483 #else // !MEASURE_MEM_ALLOC
8484 return compAsIAllocator;
8485 #endif // !MEASURE_MEM_ALLOC
8490 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8491 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8495 XX Checks for type compatibility and merges types XX
8497 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8498 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8502 // Set to TRUE if verification cannot be skipped for this method
8503 // If we detect unverifiable code, we will lazily check
8504 // canSkipMethodVerification() to see if verification is REALLY needed.
8505 BOOL tiVerificationNeeded;
8507 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8508 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8509 BOOL tiIsVerifiableCode;
8511 // Set to TRUE if runtime callout is needed for this method
8512 BOOL tiRuntimeCalloutNeeded;
8514 // Set to TRUE if security prolog/epilog callout is needed for this method
8515 // Note: This flag is different than compNeedSecurityCheck.
8516 // compNeedSecurityCheck means whether or not a security object needs
8517 // to be allocated on the stack, which is currently true for EnC as well.
8518 // tiSecurityCalloutNeeded means whether or not security callouts need
8519 // to be inserted in the jitted code.
8520 BOOL tiSecurityCalloutNeeded;
8522 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8523 // This support is necessary to suport attributes that are not described in
8524 // for example, signatures. For example, the permanent home byref (byref that
8525 // points to the gc heap), isn't a property of method signatures, therefore,
8526 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8527 // but when deciding if we need to reimport a block, we need to take these
8529 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8531 // Returns TRUE if child is equal to or a subtype of parent.
8532 // normalisedForStack indicates that both types are normalised for the stack
8533 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8535 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8536 // *pDest is modified to represent the merged type. Sets "*changed" to true
8537 // if this changes "*pDest".
8538 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8540 // Set pDest from the primitive value type.
8541 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8543 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8546 // <BUGNUM> VSW 471305
8547 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8548 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8549 // We use a "short" as we need to push/pop this scope.
8551 short compRegSetCheckLevel;
8555 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8556 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8558 XX IL verification stuff XX
8561 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8562 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8566 // The following is used to track liveness of local variables, initialization
8567 // of valueclass constructors, and type safe use of IL instructions.
8569 // dynamic state info needed for verification
8570 EntryState verCurrentState;
8572 // this ptr of object type .ctors are considered intited only after
8573 // the base class ctor is called, or an alternate ctor is called.
8574 // An uninited this ptr can be used to access fields, but cannot
8575 // be used to call a member function.
8576 BOOL verTrackObjCtorInitState;
8578 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8580 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8581 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8582 void verInitCurrentState();
8583 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8585 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8586 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8587 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8589 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8590 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8591 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8592 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8593 typeInfo verMakeTypeInfo(CorInfoType ciType,
8594 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8595 BOOL verIsSDArray(typeInfo ti);
8596 typeInfo verGetArrayElemType(typeInfo ti);
8598 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8599 BOOL verNeedsVerification();
8600 BOOL verIsByRefLike(const typeInfo& ti);
8601 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8603 // generic type variables range over types that satisfy IsBoxable
8604 BOOL verIsBoxable(const typeInfo& ti);
8606 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8607 DEBUGARG(unsigned line));
8608 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8609 DEBUGARG(unsigned line));
8610 bool verCheckTailCallConstraint(OPCODE opcode,
8611 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8612 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8613 // on a type parameter?
8614 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8615 // return false to the caller.
8616 // If false, it will throw.
8618 bool verIsBoxedValueType(typeInfo ti);
8620 void verVerifyCall(OPCODE opcode,
8621 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8622 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8624 bool readonlyCall, // is this a "readonly." call?
8625 const BYTE* delegateCreateStart,
8626 const BYTE* codeAddr,
8627 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8629 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8631 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8632 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8633 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8634 const CORINFO_FIELD_INFO& fieldInfo,
8635 const typeInfo* tiThis,
8637 BOOL allowPlainStructAsThis = FALSE);
8638 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8639 void verVerifyThisPtrInitialised();
8640 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8642 // Register allocator
8643 void raInitStackFP();
8644 void raEnregisterVarsPrePassStackFP();
8645 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8646 void raEnregisterVarsPostPassStackFP();
8647 void raGenerateFPRefCounts();
8648 void raEnregisterVarsStackFP();
8649 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8651 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8652 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8654 // returns true if enregistering v1 would save more mem accesses than v2
8655 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8658 void raDumpHeightsStackFP();
8659 void raDumpVariableRegIntfFloat();
8662 #if FEATURE_STACK_FP_X87
8664 // Currently, we use FP transition blocks in only 2 situations:
8666 // -conditional jump on longs where FP stack differs with target: it's not strictly
8667 // necessary, but its low frequency and the code would get complicated if we try to
8668 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8669 // minimize the way we generate the jump code.
8670 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8671 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8673 // However, transition blocks have 2 problems
8675 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8676 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8677 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8678 // in the right place without preordering them), this causes us to have to generate the transition
8679 // blocks in the cold area if we want procedure splitting.
8682 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8683 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8684 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8685 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8686 // a big change in the exception.
8688 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8689 // optimizations. For these 2 cases:
8691 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8692 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8693 // a switch statement.
8695 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8696 // current procedure splitting and exception code have.
8697 bool compMayHaveTransitionBlocks;
8699 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8701 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8703 unsigned raCntStkStackFP;
8704 unsigned raCntWtdStkDblStackFP;
8705 unsigned raCntStkParamDblStackFP;
8707 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8708 // TODO: Do we want to put this in LclVarDsc?
8709 unsigned raPayloadStackFP[lclMAX_TRACKED];
8710 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8712 // Useful for debugging
8713 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8715 #endif // FEATURE_STACK_FP_X87
8718 // One line log function. Default level is 0. Increasing it gives you
8719 // more log information
8721 // levels are currently unused: #define JITDUMP(level,...) ();
8722 void JitLogEE(unsigned level, const char* fmt, ...);
8724 bool compDebugBreak;
8726 bool compJitHaltMethod();
8731 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8732 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8734 XX GS Security checks for unsafe buffers XX
8736 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8737 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8740 struct ShadowParamVarInfo
8742 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8743 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8745 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8747 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8748 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8749 // slots and update all trees to refer to shadow slots is done immediately after
8750 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8751 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8752 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8753 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8754 // creating a shadow slot even though this routine returns true.
8756 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8757 // required. There are two cases under which a reg arg could potentially be used from its
8759 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8760 // b) LSRA spills it
8762 // Possible solution to address case (a)
8763 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8764 // in this routine. Note that live out of exception handler is something we may not be
8765 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8766 // Therefore, for methods with exception handling and need GS cookie check we might have
8767 // to take conservative approach.
8769 // Possible solution to address case (b)
8770 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8771 // create a new spill temp if the method needs GS cookie check.
8772 return varDsc->lvIsParam;
8773 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8774 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8781 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8786 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8787 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8788 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8790 void gsGSChecksInitCookie(); // Grabs cookie variable
8791 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8792 bool gsFindVulnerableParams(); // Shadow param analysis code
8793 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8795 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8796 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8798 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8799 // This can be overwritten by setting complus_JITInlineSize env variable.
8801 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8804 #ifdef FEATURE_JIT_METHOD_PERF
8805 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8806 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8808 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8809 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8811 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8813 #if MEASURE_CLRAPI_CALLS
8814 // Thin wrappers that call into JitTimer (if present).
8815 inline void CLRApiCallEnter(unsigned apix);
8816 inline void CLRApiCallLeave(unsigned apix);
8819 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8820 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8825 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8826 // These variables are associated with maintaining SQM data about compile time.
8827 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8828 // in the current compilation.
8829 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8830 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8831 // the inlining phase in the current compilation.
8832 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8834 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8835 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8836 // type-loading and class initialization).
8837 void RecordStateAtEndOfInlining();
8838 // Assumes being called at the end of compilation. Update the SQM state.
8839 void RecordStateAtEndOfCompilation();
8841 #ifdef FEATURE_CLRSQM
8842 // Does anything SQM related necessary at process shutdown time.
8843 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8844 #endif // FEATURE_CLRSQM
8847 #if FUNC_INFO_LOGGING
8848 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8849 // filename to write it to.
8850 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8851 #endif // FUNC_INFO_LOGGING
8853 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8855 // Is the compilation in a full trust context?
8856 bool compIsFullTrust();
8858 #ifndef FEATURE_TRACELOGGING
8859 // Should we actually fire the noway assert body and the exception handler?
8860 bool compShouldThrowOnNoway();
8861 #else // FEATURE_TRACELOGGING
8862 // Should we actually fire the noway assert body and the exception handler?
8863 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8865 // Telemetry instance to use per method compilation.
8866 JitTelemetry compJitTelemetry;
8868 // Get common parameters that have to be logged with most telemetry data.
8869 void compGetTelemetryDefaults(const char** assemblyName,
8870 const char** scopeName,
8871 const char** methodName,
8872 unsigned* methodHash);
8873 #endif // !FEATURE_TRACELOGGING
8877 NodeToTestDataMap* m_nodeTestData;
8879 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8880 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8881 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8882 // Current kept in this.
8884 NodeToTestDataMap* GetNodeTestData()
8886 Compiler* compRoot = impInlineRoot();
8887 if (compRoot->m_nodeTestData == nullptr)
8889 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8891 return compRoot->m_nodeTestData;
8894 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8896 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8897 // currently occur in the AST graph.
8898 NodeToIntMap* FindReachableNodesInNodeTestData();
8900 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8901 // test data, associate that data with "to".
8902 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8904 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8905 // have annotations, attach similar annotations to the corresponding nodes in "to".
8906 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8908 // These are the methods that test that the various conditions implied by the
8909 // test attributes are satisfied.
8910 void JitTestCheckSSA(); // SSA builder tests.
8911 void JitTestCheckVN(); // Value numbering tests.
8914 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8916 FieldSeqStore* m_fieldSeqStore;
8918 FieldSeqStore* GetFieldSeqStore()
8920 Compiler* compRoot = impInlineRoot();
8921 if (compRoot->m_fieldSeqStore == nullptr)
8923 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8924 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8925 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8927 return compRoot->m_fieldSeqStore;
8930 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8932 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8933 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8934 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8935 // attach the field sequence directly to the address node.
8936 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8938 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8940 // Don't need to worry about inlining here
8941 if (m_zeroOffsetFieldMap == nullptr)
8943 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8945 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8946 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8948 return m_zeroOffsetFieldMap;
8951 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8952 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8953 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8954 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8955 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8956 // record the the field sequence using the ZeroOffsetFieldMap described above.
8958 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8959 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8960 // CoreRT. Such case is handled same as the default case.
8961 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8963 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8965 NodeToArrayInfoMap* m_arrayInfoMap;
8967 NodeToArrayInfoMap* GetArrayInfoMap()
8969 Compiler* compRoot = impInlineRoot();
8970 if (compRoot->m_arrayInfoMap == nullptr)
8972 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8973 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8974 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
8976 return compRoot->m_arrayInfoMap;
8979 NodeToUnsignedMap* m_heapSsaMap;
8981 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
8982 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
8983 // all the possible heap states are possible initial states of the corresponding catch block(s).)
8984 NodeToUnsignedMap* GetHeapSsaMap()
8986 Compiler* compRoot = impInlineRoot();
8987 if (compRoot->m_heapSsaMap == nullptr)
8989 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8990 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8991 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
8993 return compRoot->m_heapSsaMap;
8996 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
8997 CORINFO_CLASS_HANDLE m_refAnyClass;
8998 CORINFO_FIELD_HANDLE GetRefanyDataField()
9000 if (m_refAnyClass == nullptr)
9002 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9004 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9006 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9008 if (m_refAnyClass == nullptr)
9010 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9012 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9016 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9018 #if ALLVARSET_COUNTOPS
9019 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9022 static HelperCallProperties s_helperCallProperties;
9024 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9025 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9026 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9028 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9031 unsigned __int8* offset0,
9032 unsigned __int8* offset1);
9033 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9034 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9036 void fgMorphMultiregStructArgs(GenTreeCall* call);
9037 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9039 }; // end of class Compiler
9041 // Inline methods of CompAllocator.
9042 void* CompAllocator::Alloc(size_t sz)
9044 #if MEASURE_MEM_ALLOC
9045 return m_comp->compGetMem(sz, m_cmk);
9047 return m_comp->compGetMem(sz);
9051 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9053 #if MEASURE_MEM_ALLOC
9054 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9056 return m_comp->compGetMemArray(elems, elemSize);
9060 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9061 inline LclVarDsc::LclVarDsc(Compiler* comp)
9062 : // Initialize the ArgRegs to REG_STK.
9063 // The morph will do the right thing to change
9064 // to the right register if passed in register.
9067 #if FEATURE_MULTIREG_ARGS
9068 _lvOtherArgReg(REG_STK)
9070 #endif // FEATURE_MULTIREG_ARGS
9072 lvRefBlks(BlockSetOps::UninitVal())
9074 #endif // ASSERTION_PROP
9075 lvPerSsaData(comp->getAllocator())
9080 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9081 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9083 XX Miscellaneous Compiler stuff XX
9085 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9086 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9089 // Values used to mark the types a stack slot is used for
9091 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9092 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9093 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9094 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9095 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9096 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9097 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9098 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9100 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9102 /*****************************************************************************
9104 * Variables to keep track of total code amounts.
9109 extern size_t grossVMsize;
9110 extern size_t grossNCsize;
9111 extern size_t totalNCsize;
9113 extern unsigned genMethodICnt;
9114 extern unsigned genMethodNCnt;
9115 extern size_t gcHeaderISize;
9116 extern size_t gcPtrMapISize;
9117 extern size_t gcHeaderNSize;
9118 extern size_t gcPtrMapNSize;
9120 #endif // DISPLAY_SIZES
9122 /*****************************************************************************
9124 * Variables to keep track of basic block counts (more data on 1 BB methods)
9127 #if COUNT_BASIC_BLOCKS
9128 extern Histogram bbCntTable;
9129 extern Histogram bbOneBBSizeTable;
9132 /*****************************************************************************
9134 * Used by optFindNaturalLoops to gather statistical information such as
9135 * - total number of natural loops
9136 * - number of loops with 1, 2, ... exit conditions
9137 * - number of loops that have an iterator (for like)
9138 * - number of loops that have a constant iterator
9143 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9144 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9145 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9146 extern unsigned totalLoopCount; // counts the total number of natural loops
9147 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9148 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9149 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9150 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9152 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9153 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9154 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9155 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9156 extern Histogram loopCountTable; // Histogram of loop counts
9157 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9159 #endif // COUNT_LOOPS
9161 /*****************************************************************************
9162 * variables to keep track of how many iterations we go in a dataflow pass
9167 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9168 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9170 #endif // DATAFLOW_ITER
9172 #if MEASURE_BLOCK_SIZE
9173 extern size_t genFlowNodeSize;
9174 extern size_t genFlowNodeCnt;
9175 #endif // MEASURE_BLOCK_SIZE
9177 #if MEASURE_NODE_SIZE
9178 struct NodeSizeStats
9183 genTreeNodeSize = 0;
9184 genTreeNodeActualSize = 0;
9187 size_t genTreeNodeCnt;
9188 size_t genTreeNodeSize; // The size we allocate
9189 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9190 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9191 // a smaller node to a larger one. TODO-Cleanup: add stats on
9192 // SetOper()/ChangeOper() usage to quanitfy this.
9194 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9195 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9196 extern Histogram genTreeNcntHist;
9197 extern Histogram genTreeNsizHist;
9198 #endif // MEASURE_NODE_SIZE
9200 /*****************************************************************************
9201 * Count fatal errors (including noway_asserts).
9205 extern unsigned fatal_badCode;
9206 extern unsigned fatal_noWay;
9207 extern unsigned fatal_NOMEM;
9208 extern unsigned fatal_noWayAssertBody;
9210 extern unsigned fatal_noWayAssertBodyArgs;
9212 extern unsigned fatal_NYI;
9213 #endif // MEASURE_FATAL
9215 /*****************************************************************************
9219 #ifdef _TARGET_XARCH_
9221 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9222 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9223 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9225 const instruction INS_AND = INS_and;
9226 const instruction INS_OR = INS_or;
9227 const instruction INS_XOR = INS_xor;
9228 const instruction INS_NEG = INS_neg;
9229 const instruction INS_TEST = INS_test;
9230 const instruction INS_MUL = INS_imul;
9231 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9232 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9233 const instruction INS_BREAKPOINT = INS_int3;
9234 const instruction INS_ADDC = INS_adc;
9235 const instruction INS_SUBC = INS_sbb;
9236 const instruction INS_NOT = INS_not;
9242 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9243 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9244 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9246 const instruction INS_AND = INS_and;
9247 const instruction INS_OR = INS_orr;
9248 const instruction INS_XOR = INS_eor;
9249 const instruction INS_NEG = INS_rsb;
9250 const instruction INS_TEST = INS_tst;
9251 const instruction INS_MUL = INS_mul;
9252 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9253 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9254 const instruction INS_BREAKPOINT = INS_bkpt;
9255 const instruction INS_ADDC = INS_adc;
9256 const instruction INS_SUBC = INS_sbc;
9257 const instruction INS_NOT = INS_mvn;
9261 #ifdef _TARGET_ARM64_
9263 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9264 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9265 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9267 const instruction INS_AND = INS_and;
9268 const instruction INS_OR = INS_orr;
9269 const instruction INS_XOR = INS_eor;
9270 const instruction INS_NEG = INS_neg;
9271 const instruction INS_TEST = INS_tst;
9272 const instruction INS_MUL = INS_mul;
9273 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9274 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9275 const instruction INS_BREAKPOINT = INS_bkpt;
9276 const instruction INS_ADDC = INS_adc;
9277 const instruction INS_SUBC = INS_sbc;
9278 const instruction INS_NOT = INS_mvn;
9282 /*****************************************************************************/
9284 extern const BYTE genTypeSizes[];
9285 extern const BYTE genTypeAlignments[];
9286 extern const BYTE genTypeStSzs[];
9287 extern const BYTE genActualTypes[];
9289 /*****************************************************************************/
9291 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9292 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9295 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9296 #elif defined(_TARGET_ARM64_)
9297 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9300 /*****************************************************************************/
9302 #define REG_CORRUPT regNumber(REG_NA + 1)
9303 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9304 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9306 /*****************************************************************************/
9308 extern BasicBlock dummyBB;
9310 /*****************************************************************************/
9311 /*****************************************************************************/
9313 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9314 // nodes of a statement in execution order.
9315 // __stmt: a GT_STMT type GenTree*
9316 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9318 #define foreach_treenode_execution_order(__node, __stmt) \
9319 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9321 // foreach_block: An iterator over all blocks in the function.
9322 // __compiler: the Compiler* object
9323 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9325 #define foreach_block(__compiler, __block) \
9326 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9328 /*****************************************************************************/
9329 /*****************************************************************************/
9333 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9335 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9336 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9338 XX Debugging helpers XX
9340 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9341 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9344 /*****************************************************************************/
9345 /* The following functions are intended to be called from the debugger, to dump
9346 * various data structures. The can be used in the debugger Watch or Quick Watch
9347 * windows. They are designed to be short to type and take as few arguments as
9348 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9349 * See the function definition comment for more details.
9352 void cBlock(Compiler* comp, BasicBlock* block);
9353 void cBlocks(Compiler* comp);
9354 void cBlocksV(Compiler* comp);
9355 void cTree(Compiler* comp, GenTree* tree);
9356 void cTrees(Compiler* comp);
9357 void cEH(Compiler* comp);
9358 void cVar(Compiler* comp, unsigned lclNum);
9359 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9360 void cVars(Compiler* comp);
9361 void cVarsFinal(Compiler* comp);
9362 void cBlockPreds(Compiler* comp, BasicBlock* block);
9363 void cReach(Compiler* comp);
9364 void cDoms(Compiler* comp);
9365 void cLiveness(Compiler* comp);
9366 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9368 void cFuncIR(Compiler* comp);
9369 void cBlockIR(Compiler* comp, BasicBlock* block);
9370 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9371 void cTreeIR(Compiler* comp, GenTree* tree);
9372 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9373 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9374 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9375 int cOperandIR(Compiler* comp, GenTree* operand);
9376 int cLeafIR(Compiler* comp, GenTree* tree);
9377 int cIndirIR(Compiler* comp, GenTree* tree);
9378 int cListIR(Compiler* comp, GenTree* list);
9379 int cSsaNumIR(Compiler* comp, GenTree* tree);
9380 int cValNumIR(Compiler* comp, GenTree* tree);
9381 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9383 void dBlock(BasicBlock* block);
9386 void dTree(GenTree* tree);
9389 void dVar(unsigned lclNum);
9390 void dVarDsc(LclVarDsc* varDsc);
9393 void dBlockPreds(BasicBlock* block);
9397 void dCVarSet(VARSET_VALARG_TP vars);
9399 void dVarSet(VARSET_VALARG_TP vars);
9400 void dRegMask(regMaskTP mask);
9403 void dBlockIR(BasicBlock* block);
9404 void dTreeIR(GenTree* tree);
9405 void dLoopIR(Compiler::LoopDsc* loop);
9406 void dLoopNumIR(unsigned loopNum);
9407 int dTabStopIR(int curr, int tabstop);
9408 int dTreeTypeIR(GenTree* tree);
9409 int dTreeKindsIR(GenTree* tree);
9410 int dTreeFlagsIR(GenTree* tree);
9411 int dOperandIR(GenTree* operand);
9412 int dLeafIR(GenTree* tree);
9413 int dIndirIR(GenTree* tree);
9414 int dListIR(GenTree* list);
9415 int dSsaNumIR(GenTree* tree);
9416 int dValNumIR(GenTree* tree);
9417 int dDependsIR(GenTree* comma);
9420 GenTree* dFindTree(GenTree* tree, unsigned id);
9421 GenTree* dFindTree(unsigned id);
9422 GenTreeStmt* dFindStmt(unsigned id);
9423 BasicBlock* dFindBlock(unsigned bbNum);
9427 #include "compiler.hpp" // All the shared inline functions
9429 /*****************************************************************************/
9430 #endif //_COMPILER_H_
9431 /*****************************************************************************/