1 // Licensed to the .NET Foundation under one or more agreements.
2 // The .NET Foundation licenses this file to you under the MIT license.
3 // See the LICENSE file in the project root for more information.
5 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10 XX Represents the method data we are currently JIT-compiling. XX
11 XX An instance of this class is created for every method we JIT. XX
12 XX This contains all the info needed for the method. So allocating a XX
13 XX a new instance per method makes it thread-safe. XX
14 XX It should be used to do all the memory management for the compiler run. XX
16 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
17 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
20 /*****************************************************************************/
23 /*****************************************************************************/
36 #include "simplerhash.h"
37 #include "cycletimer.h"
40 #include "arraystack.h"
43 #include "expandarray.h"
44 #include "tinyarray.h"
47 #include "jittelemetry.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
64 // This is only used locally in the JIT to indicate that
65 // a verification block should be inserted
66 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
68 /*****************************************************************************
69 * Forward declarations
72 struct InfoHdr; // defined in GCInfo.h
73 struct escapeMapping_t; // defined in flowgraph.cpp
74 class emitter; // defined in emit.h
75 struct ShadowParamVarInfo; // defined in GSChecks.cpp
76 struct InitVarDscInfo; // defined in register_arg_convention.h
77 class FgStack; // defined in flowgraph.cpp
78 #if FEATURE_STACK_FP_X87
79 struct FlatFPStateX87; // defined in fp.h
82 class CSE_DataFlow; // defined in OptCSE.cpp
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the Compiler::compGetMem() function for allocation.
104 // Or the more-general IAllocator interface.
105 void* __cdecl operator new(size_t n, IAllocator* alloc);
106 void* __cdecl operator new[](size_t n, IAllocator* alloc);
108 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
109 // caused these to be ambiguous with the global placement new operators.
110 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
111 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
112 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
114 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
115 #include "loopcloning.h"
117 /*****************************************************************************/
119 /* This is included here and not earlier as it needs the definition of "CSE"
120 * which is defined in the section above */
122 /*****************************************************************************/
124 unsigned genLog2(unsigned value);
125 unsigned genLog2(unsigned __int64 value);
127 var_types genActualType(var_types type);
128 var_types genUnsignedType(var_types type);
129 var_types genSignedType(var_types type);
131 unsigned ReinterpretHexAsDecimal(unsigned);
133 /*****************************************************************************/
136 #ifdef FEATURE_AVX_SUPPORT
137 const unsigned TEMP_MAX_SIZE = YMM_REGSIZE_BYTES;
138 #else // !FEATURE_AVX_SUPPORT
139 const unsigned TEMP_MAX_SIZE = XMM_REGSIZE_BYTES;
140 #endif // !FEATURE_AVX_SUPPORT
141 #else // !FEATURE_SIMD
142 const unsigned TEMP_MAX_SIZE = sizeof(double);
143 #endif // !FEATURE_SIMD
144 const unsigned TEMP_SLOT_COUNT = (TEMP_MAX_SIZE / sizeof(int));
146 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
149 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
152 // The following holds the Local var info (scope information)
153 typedef const char* VarName; // Actual ASCII string
156 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
157 IL_OFFSET vsdLifeEnd; // instr offset of end of life
158 unsigned vsdVarNum; // (remapped) LclVarDsc number
161 VarName vsdName; // name of the var
164 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
165 // Also, it is the index of this entry in the info.compVarScopes array,
166 // which is useful since the array is also accessed via the
167 // compEnterScopeList and compExitScopeList sorted arrays.
170 /*****************************************************************************
172 * The following holds the local variable counts and the descriptor table.
175 // This is the location of a definition.
181 DefLoc() : m_blk(nullptr), m_tree(nullptr)
186 // This class encapsulates all info about a local variable that may vary for different SSA names
191 ValueNumPair m_vnPair;
199 typedef ExpandArray<LclSsaVarDsc> PerSsaArray;
204 // The constructor. Most things can just be zero'ed.
205 LclVarDsc(Compiler* comp);
207 // note this only packs because var_types is a typedef of unsigned char
208 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
210 unsigned char lvIsParam : 1; // is this a parameter?
211 unsigned char lvIsRegArg : 1; // is this a register argument?
212 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
214 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
215 // is to help determine whether to use block init in the prolog.
216 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
217 unsigned char lvDependReg : 1; // did the predictor depend upon this being enregistered
218 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
219 // variable is in the same register for the entire function.
220 unsigned char lvTracked : 1; // is this a tracked variable?
221 bool lvTrackedNonStruct()
223 return lvTracked && lvType != TYP_STRUCT;
225 unsigned char lvPinned : 1; // is this a pinned variable?
227 unsigned char lvMustInit : 1; // must be initialized
228 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
229 // global location, etc.
230 // We cannot reason reliably about the value of the variable.
231 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
232 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
236 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
238 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
239 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
240 // read/write its value.
241 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
242 // the variable to be
243 // in the stack (at least at those boundaries.)
244 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
245 // particular byte from an int).
246 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
247 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
249 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
250 unsigned char lvRefAssign : 1; // involved in pointer assignment
251 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
252 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
255 unsigned char lvArgWrite : 1; // variable is a parameter and STARG was used on it
256 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp
258 unsigned char lvIsBoolean : 1; // set if variable is boolean
260 unsigned char lvRngOptDone : 1; // considered for range check opt?
261 unsigned char lvLoopInc : 1; // incremented in the loop?
262 unsigned char lvLoopAsg : 1; // reassigned in the loop (other than a monotonic inc/dec for the index var)?
263 unsigned char lvArrIndx : 1; // used as an array index?
264 unsigned char lvArrIndxOff : 1; // used as an array index with an offset?
265 unsigned char lvArrIndxDom : 1; // index dominates loop exit
267 unsigned char lvSingleDef : 1; // variable has a single def
268 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
269 unsigned char lvVolatileHint : 1; // hint for AssertionProp
272 unsigned char lvSpilled : 1; // enregistered variable was spilled
273 #ifndef _TARGET_64BIT_
274 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
275 #endif // !_TARGET_64BIT_
276 #ifdef _TARGET_64BIT_
277 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
280 unsigned char lvKeepType : 1; // Don't change the type of this variable
281 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
283 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
285 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
286 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
288 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
289 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
290 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
291 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
292 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
294 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
295 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
298 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
299 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
300 // with (lvIsRegArg && lvIsHfa())
301 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
302 #endif // FEATURE_HFA
305 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
306 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
307 // Consider cleaning this up so this workaround is not required.
308 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
309 // I.e. there is no longer any reference to the struct directly.
310 // In this case we can simply remove this struct local.
312 #ifndef LEGACY_BACKEND
313 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
314 #endif // !LEGACY_BACKEND
317 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
318 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
319 unsigned char lvSIMDType : 1; // This is a SIMD struct
320 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
321 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
322 #endif // FEATURE_SIMD
323 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
326 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
328 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
329 // Valid on promoted struct local fields.
332 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
333 unsigned char lvFldOffset;
334 unsigned char lvFldOrdinal;
336 #if FEATURE_MULTIREG_ARGS
337 regNumber lvRegNumForSlot(unsigned slotNum)
343 else if (slotNum == 1)
345 return lvOtherArgReg;
349 assert(false && "Invalid slotNum!");
354 #endif // FEATURE_MULTIREG_ARGS
372 bool lvIsHfaRegArg() const
375 return _lvIsHfaRegArg;
381 void lvSetIsHfaRegArg()
384 _lvIsHfaRegArg = true;
388 bool lvHfaTypeIsFloat() const
391 return _lvHfaTypeIsFloat;
397 void lvSetHfaTypeIsFloat(bool value)
400 _lvHfaTypeIsFloat = value;
404 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
405 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
407 unsigned lvHfaSlots() const
410 assert(lvType == TYP_STRUCT);
412 return lvExactSize / sizeof(float);
413 #else // _TARGET_ARM64_
414 if (lvHfaTypeIsFloat())
416 return lvExactSize / sizeof(float);
420 return lvExactSize / sizeof(double);
422 #endif // _TARGET_ARM64_
425 // lvIsMultiRegArgOrRet()
426 // returns true if this is a multireg LclVar struct used in an argument context
427 // or if this is a multireg LclVar struct assigned from a multireg call
428 bool lvIsMultiRegArgOrRet()
430 return lvIsMultiRegArg || lvIsMultiRegRet;
434 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
435 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
436 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
437 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
438 // to non-zero if the variable gets the same register assignment for its entire
440 #if !defined(_TARGET_64BIT_)
441 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
442 #endif // !defined(_TARGET_64BIT_)
444 regNumberSmall _lvArgReg; // The register in which this argument is passed.
446 #if FEATURE_MULTIREG_ARGS
447 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
448 // Note this is defined but not used by ARM32
449 #endif // FEATURE_MULTIREG_ARGS
451 #ifndef LEGACY_BACKEND
453 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
454 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
456 #endif // !LEGACY_BACKEND
459 // The register number is stored in a small format (8 bits), but the getters return and the setters take
460 // a full-size (unsigned) format, to localize the casts here.
462 /////////////////////
464 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
466 regNumber GetRegNum() const
468 return (regNumber)_lvRegNum;
471 void SetRegNum(regNumber reg)
473 _lvRegNum = (regNumberSmall)reg;
474 assert(_lvRegNum == reg);
477 /////////////////////
479 #if defined(_TARGET_64BIT_)
480 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
482 regNumber GetOtherReg() const
484 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
485 // "unreachable code" warnings
489 void SetOtherReg(regNumber reg)
491 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
492 // "unreachable code" warnings
494 #else // !_TARGET_64BIT_
495 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
497 regNumber GetOtherReg() const
499 return (regNumber)_lvOtherReg;
502 void SetOtherReg(regNumber reg)
504 _lvOtherReg = (regNumberSmall)reg;
505 assert(_lvOtherReg == reg);
507 #endif // !_TARGET_64BIT_
509 /////////////////////
511 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
513 regNumber GetArgReg() const
515 return (regNumber)_lvArgReg;
518 void SetArgReg(regNumber reg)
520 _lvArgReg = (regNumberSmall)reg;
521 assert(_lvArgReg == reg);
524 #if FEATURE_MULTIREG_ARGS
525 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
527 regNumber GetOtherArgReg() const
529 return (regNumber)_lvOtherArgReg;
532 void SetOtherArgReg(regNumber reg)
534 _lvOtherArgReg = (regNumberSmall)reg;
535 assert(_lvOtherArgReg == reg);
537 #endif // FEATURE_MULTIREG_ARGS
540 // Is this is a SIMD struct?
541 bool lvIsSIMDType() const
546 // Is this is a SIMD struct which is used for SIMD intrinsic?
547 bool lvIsUsedInSIMDIntrinsic() const
549 return lvUsedInSIMDIntrinsic;
552 // If feature_simd not enabled, return false
553 bool lvIsSIMDType() const
557 bool lvIsUsedInSIMDIntrinsic() const
563 /////////////////////
565 #ifndef LEGACY_BACKEND
566 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
568 regNumber GetArgInitReg() const
570 return (regNumber)_lvArgInitReg;
573 void SetArgInitReg(regNumber reg)
575 _lvArgInitReg = (regNumberSmall)reg;
576 assert(_lvArgInitReg == reg);
579 /////////////////////
581 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
583 regPairNo GetArgInitRegPair() const
585 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
586 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
590 void SetArgInitRegPair(regPairNo regPair)
592 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
593 _lvArgInitRegPair = (regPairNoSmall)regPair;
594 assert(_lvArgInitRegPair == regPair);
597 /////////////////////
599 bool lvIsRegCandidate() const
601 return lvLRACandidate != 0;
604 bool lvIsInReg() const
606 return lvIsRegCandidate() && (lvRegNum != REG_STK);
609 #else // LEGACY_BACKEND
611 bool lvIsRegCandidate() const
613 return lvTracked != 0;
616 bool lvIsInReg() const
618 return lvRegister != 0;
621 #endif // LEGACY_BACKEND
623 regMaskTP lvRegMask() const
625 regMaskTP regMask = RBM_NONE;
626 if (varTypeIsFloating(TypeGet()))
628 if (lvRegNum != REG_STK)
630 regMask = genRegMaskFloat(lvRegNum, TypeGet());
635 if (lvRegNum != REG_STK)
637 regMask = genRegMask(lvRegNum);
640 // For longs we may have two regs
641 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
643 regMask |= genRegMask(lvOtherReg);
649 regMaskSmall lvPrefReg; // set of regs it prefers to live in
651 unsigned short lvVarIndex; // variable tracking index
652 unsigned short lvRefCnt; // unweighted (real) reference count
653 unsigned lvRefCntWtd; // weighted reference count
654 int lvStkOffs; // stack offset of home
655 unsigned lvExactSize; // (exact) size of the type in bytes
657 // Is this a promoted struct?
658 // This method returns true only for structs (including SIMD structs), not for
659 // locals that are split on a 32-bit target.
660 // It is only necessary to use this:
661 // 1) if only structs are wanted, and
662 // 2) if Lowering has already been done.
663 // Otherwise lvPromoted is valid.
664 bool lvPromotedStruct()
666 #if !defined(_TARGET_64BIT_)
667 return (lvPromoted && !varTypeIsLong(lvType));
668 #else // defined(_TARGET_64BIT_)
670 #endif // defined(_TARGET_64BIT_)
673 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
675 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
676 // where the struct itself is no longer used because all access is via its member fields.
677 // When that happens, the struct is marked as unused and its type has been changed to
678 // TYP_INT (to keep the GC tracking code from looking at it).
679 // See Compiler::raAssignVars() for details. For example:
680 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
681 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
682 // float V03.f1 (offs=0x00) -> V12 tmp7
683 // f8 (last use) (last use) $345
684 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
685 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
687 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
689 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
690 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
691 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
692 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
693 if ((lvType == TYP_SIMD12) && !lvIsParam)
695 assert(lvExactSize == 12);
698 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
700 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
703 unsigned lvSlotNum; // original slot # (if remapped)
705 typeInfo lvVerTypeInfo; // type info needed for verification
707 BYTE* lvGcLayout; // GC layout info for structs
710 BlockSet lvRefBlks; // Set of blocks that contain refs
711 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
712 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
714 var_types TypeGet() const
716 return (var_types)lvType;
718 bool lvStackAligned() const
720 assert(lvIsStructField);
721 return ((lvFldOffset % sizeof(void*)) == 0);
723 bool lvNormalizeOnLoad() const
725 return varTypeIsSmall(TypeGet()) &&
726 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
727 (lvIsParam || lvAddrExposed || lvIsStructField);
730 bool lvNormalizeOnStore()
732 return varTypeIsSmall(TypeGet()) &&
733 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
734 !(lvIsParam || lvAddrExposed || lvIsStructField);
737 void lvaResetSortAgainFlag(Compiler* pComp);
738 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
739 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
740 void setPrefReg(regNumber regNum, Compiler* pComp);
741 void addPrefReg(regMaskTP regMask, Compiler* pComp);
742 bool IsFloatRegType() const
744 return isFloatRegType(lvType) || lvIsHfaRegArg();
746 var_types GetHfaType() const
748 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
750 void SetHfaType(var_types type)
752 assert(varTypeIsFloating(type));
753 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
756 #ifndef LEGACY_BACKEND
757 var_types lvaArgType();
760 PerSsaArray lvPerSsaData;
763 // Keep track of the # of SsaNames, for a bounds check.
764 unsigned lvNumSsaNames;
767 // Returns the address of the per-Ssa data for the given ssaNum (which is required
768 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
769 // not an SSA variable).
770 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
772 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
773 assert(SsaConfig::RESERVED_SSA_NUM == 0);
774 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
775 assert(zeroBased < lvNumSsaNames);
776 return &lvPerSsaData.GetRef(zeroBased);
781 void PrintVarReg() const
783 if (isRegPairType(TypeGet()))
785 printf("%s:%s", getRegName(lvOtherReg), // hi32
786 getRegName(lvRegNum)); // lo32
790 printf("%s", getRegName(lvRegNum));
795 }; // class LclVarDsc
798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
799 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
803 XX The temporary lclVars allocated by the compiler for code generation XX
805 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
809 /*****************************************************************************
811 * The following keeps track of temporaries allocated in the stack frame
812 * during code-generation (after register allocation). These spill-temps are
813 * only used if we run out of registers while evaluating a tree.
815 * These are different from the more common temps allocated by lvaGrabTemp().
826 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
834 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
838 0); // temps must have a negative number (so they have a different number from all local variables)
839 tdOffs = BAD_TEMP_OFFSET;
843 IMPL_LIMITATION("too many spill temps");
848 bool tdLegalOffset() const
850 return tdOffs != BAD_TEMP_OFFSET;
854 int tdTempOffs() const
856 assert(tdLegalOffset());
859 void tdSetTempOffs(int offs)
862 assert(tdLegalOffset());
864 void tdAdjustTempOffs(int offs)
867 assert(tdLegalOffset());
870 int tdTempNum() const
875 unsigned tdTempSize() const
879 var_types tdTempType() const
885 // interface to hide linearscan implementation from rest of compiler
886 class LinearScanInterface
889 virtual void doLinearScan() = 0;
890 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
893 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
895 // Information about arrays: their element type and size, and the offset of the first element.
896 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
897 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
898 // for example, in value numbering of array index expressions.
901 var_types m_elemType;
902 CORINFO_CLASS_HANDLE m_elemStructType;
904 unsigned m_elemOffset;
906 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
910 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
911 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
916 // This enumeration names the phases into which we divide compilation. The phases should completely
917 // partition a compilation.
920 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
921 #include "compphases.h"
925 extern const char* PhaseNames[];
926 extern const char* PhaseEnums[];
927 extern const LPCWSTR PhaseShortNames[];
929 // The following enum provides a simple 1:1 mapping to CLR API's
930 enum API_ICorJitInfo_Names
932 #define DEF_CLR_API(name) API_##name,
933 #include "ICorJitInfo_API_names.h"
937 //---------------------------------------------------------------
941 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
942 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
943 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
944 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
945 // by "m_timerFailure" being true.
946 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
949 #ifdef FEATURE_JIT_METHOD_PERF
950 // The string names of the phases.
951 static const char* PhaseNames[];
953 static bool PhaseHasChildren[];
954 static int PhaseParent[];
956 unsigned m_byteCodeBytes;
957 unsigned __int64 m_totalCycles;
958 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
959 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
960 #if MEASURE_CLRAPI_CALLS
961 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
962 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
964 // For better documentation, we call EndPhase on
965 // non-leaf phases. We should also call EndPhase on the
966 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
967 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
968 // We add all such "redundant end phase" intervals to this variable below; we print
969 // it out in a report, so we can verify that it is, indeed, very small. If it ever
970 // isn't, this means that we're doing something significant between the end of the last
971 // declared subphase and the end of its parent.
972 unsigned __int64 m_parentPhaseEndSlop;
975 #if MEASURE_CLRAPI_CALLS
976 // The following measures the time spent inside each individual CLR API call.
977 unsigned m_allClrAPIcalls;
978 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
979 unsigned __int64 m_allClrAPIcycles;
980 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
981 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
982 #endif // MEASURE_CLRAPI_CALLS
984 CompTimeInfo(unsigned byteCodeBytes);
988 #ifdef FEATURE_JIT_METHOD_PERF
990 #if MEASURE_CLRAPI_CALLS
991 struct WrapICorJitInfo;
994 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
995 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
996 // The operation of adding a single method's timing to the summary may be performed concurrently by several
997 // threads, so it is protected by a lock.
998 // This class is intended to be used as a singleton type, with only a single instance.
999 class CompTimeSummaryInfo
1001 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1002 static CritSecObject s_compTimeSummaryLock;
1006 CompTimeInfo m_total;
1007 CompTimeInfo m_maximum;
1009 int m_numFilteredMethods;
1010 CompTimeInfo m_filtered;
1012 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1013 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1014 // If any OS interaction fails, returns 0.0.
1015 double CyclesPerSecond();
1017 // This can use what ever data you want to determine if the value to be added
1018 // belongs in the filtered section (it's always included in the unfiltered section)
1019 bool IncludedInFilteredData(CompTimeInfo& info);
1022 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1023 static CompTimeSummaryInfo s_compTimeSummary;
1025 CompTimeSummaryInfo()
1026 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1030 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1031 // This is thread safe.
1032 void AddInfo(CompTimeInfo& info, bool includePhases);
1034 // Print the summary information to "f".
1035 // This is not thread-safe; assumed to be called by only one thread.
1036 void Print(FILE* f);
1039 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1040 // and when the current phase started. This is intended to be part of a Compilation object. This is
1041 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1045 unsigned __int64 m_start; // Start of the compilation.
1046 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1047 #if MEASURE_CLRAPI_CALLS
1048 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1049 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1050 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1051 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1052 static double s_cyclesPerSec; // Cached for speedier measurements
1055 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1057 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1059 static CritSecObject s_csvLock; // Lock to protect the time log file.
1060 void PrintCsvMethodStats(Compiler* comp);
1063 void* operator new(size_t);
1064 void* operator new[](size_t);
1065 void operator delete(void*);
1066 void operator delete[](void*);
1069 // Initialized the timer instance
1070 JitTimer(unsigned byteCodeSize);
1072 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1074 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1077 static void PrintCsvHeader();
1079 // Ends the current phase (argument is for a redundant check).
1080 void EndPhase(Phases phase);
1082 #if MEASURE_CLRAPI_CALLS
1083 // Start and end a timed CLR API call.
1084 void CLRApiCallEnter(unsigned apix);
1085 void CLRApiCallLeave(unsigned apix);
1086 #endif // MEASURE_CLRAPI_CALLS
1088 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1089 // and adds it to "sum".
1090 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1092 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1093 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1094 // "m_info" to true.
1095 bool GetThreadCycles(unsigned __int64* cycles)
1097 bool res = CycleTimer::GetThreadCyclesS(cycles);
1100 m_info.m_timerFailure = true;
1105 #endif // FEATURE_JIT_METHOD_PERF
1107 //------------------- Function/Funclet info -------------------------------
1108 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1110 FUNC_ROOT, // The main/root function (always id==0)
1111 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1112 FUNC_FILTER, // a funclet associated with an EH filter
1115 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1122 BYTE funFlags; // Currently unused, just here for padding
1123 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1124 // funclet. It is only valid if funKind field indicates this is a
1125 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1127 #if defined(_TARGET_AMD64_)
1129 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1130 emitLocation* startLoc;
1131 emitLocation* endLoc;
1132 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1133 emitLocation* coldEndLoc;
1134 UNWIND_INFO unwindHeader;
1135 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1136 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1137 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1138 unsigned unwindCodeSlot;
1140 #ifdef UNIX_AMD64_ABI
1141 jitstd::vector<CFI_CODE>* cfiCodes;
1142 #endif // UNIX_AMD64_ABI
1144 #elif defined(_TARGET_ARMARCH_)
1146 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1147 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1148 // Note: we only have a pointer here instead of the actual object,
1149 // to save memory in the JIT case (compared to the NGEN case),
1150 // where we don't have any cold section.
1151 // Note 2: we currently don't support hot/cold splitting in functions
1152 // with EH, so uwiCold will be NULL for all funclets.
1154 #endif // _TARGET_ARMARCH_
1156 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1157 // that isn't shared between the main function body and funclets.
1160 struct fgArgTabEntry
1163 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1166 otherRegNum = REG_NA;
1167 isStruct = false; // is this a struct arg
1169 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1171 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1173 // it will point at the actual argument in the gtCallLateArgs list.
1174 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1176 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1178 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1180 unsigned numRegs; // Count of number of registers that this argument uses
1182 // A slot is a pointer sized region in the OutArg area.
1183 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1184 unsigned numSlots; // Count of number of slots that this argument uses
1186 unsigned alignment; // 1 or 2 (slots/registers)
1187 unsigned lateArgInx; // index into gtCallLateArgs list
1188 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1189 #if defined(UNIX_X86_ABI)
1190 unsigned padStkAlign; // Count of number of padding slots for stack alignment. For each Call, only the first
1191 // argument may have a value to emit "sub esp, n" to adjust the stack before pushing
1195 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1196 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1197 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1198 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1199 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1200 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1201 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1202 // previous arguments.
1203 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1204 // to be on the stack despite its arg list position.
1206 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1207 bool isStruct : 1; // True if this is a struct arg
1209 regNumber otherRegNum; // The (second) register to use when passing this argument.
1211 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1212 #elif defined(_TARGET_X86_)
1213 __declspec(property(get = getIsStruct)) bool isStruct;
1216 return varTypeIsStruct(node);
1218 #endif // _TARGET_X86_
1221 void SetIsHfaRegArg(bool hfaRegArg)
1223 isHfaRegArg = hfaRegArg;
1226 void SetIsBackFilled(bool backFilled)
1228 isBackFilled = backFilled;
1231 bool IsBackFilled() const
1233 return isBackFilled;
1235 #else // !_TARGET_ARM_
1236 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1238 void SetIsHfaRegArg(bool hfaRegArg)
1242 void SetIsBackFilled(bool backFilled)
1246 bool IsBackFilled() const
1250 #endif // !_TARGET_ARM_
1256 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1258 //-------------------------------------------------------------------------
1260 // The class fgArgInfo is used to handle the arguments
1261 // when morphing a GT_CALL node.
1266 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1267 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1268 unsigned argCount; // Updatable arg count value
1269 unsigned nextSlotNum; // Updatable slot count value
1270 unsigned stkLevel; // Stack depth when we make this call (for x86)
1271 #if defined(UNIX_X86_ABI)
1272 unsigned padStkAlign; // Count of number of padding slots for stack alignment. This value is used to turn back
1273 // stack pointer before it was adjusted after each Call
1276 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1277 bool hasRegArgs; // true if we have one or more register arguments
1278 bool hasStackArgs; // true if we have one or more stack arguments
1279 bool argsComplete; // marker for state
1280 bool argsSorted; // marker for state
1281 fgArgTabEntryPtr* argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1284 void AddArg(fgArgTabEntryPtr curArgTabEntry);
1287 fgArgInfo(Compiler* comp, GenTreePtr call, unsigned argCount);
1288 fgArgInfo(GenTreePtr newCall, GenTreePtr oldCall);
1290 fgArgTabEntryPtr AddRegArg(
1291 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1293 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
1294 fgArgTabEntryPtr AddRegArg(
1301 const bool isStruct,
1302 const regNumber otherRegNum = REG_NA,
1303 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1304 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
1306 fgArgTabEntryPtr AddStkArg(unsigned argNum,
1310 unsigned alignment FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool isStruct));
1312 void RemorphReset();
1313 fgArgTabEntryPtr RemorphRegArg(
1314 unsigned argNum, GenTreePtr node, GenTreePtr parent, regNumber regNum, unsigned numRegs, unsigned alignment);
1316 void RemorphStkArg(unsigned argNum, GenTreePtr node, GenTreePtr parent, unsigned numSlots, unsigned alignment);
1318 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1320 void EvalToTmp(unsigned argNum, unsigned tmpNum, GenTreePtr newNode);
1322 void ArgsComplete();
1324 #if defined(UNIX_X86_ABI)
1325 void ArgsAlignPadding();
1330 void EvalArgsToTemps();
1332 void RecordStkLevel(unsigned stkLvl);
1333 unsigned RetrieveStkLevel();
1339 fgArgTabEntryPtr* ArgTable()
1343 unsigned GetNextSlotNum()
1347 #if defined(UNIX_X86_ABI)
1348 unsigned GetPadStackAlign()
1359 return hasStackArgs;
1361 bool AreArgsComplete() const
1363 return argsComplete;
1366 // Get the late arg for arg at position argIndex. Caller must ensure this position has a late arg.
1367 GenTreePtr GetLateArg(unsigned argIndex);
1371 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1372 // We have the ability to mark source expressions with "Test Labels."
1373 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1374 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1376 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1379 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1380 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1381 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1382 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1383 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1386 struct TestLabelAndNum
1391 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1396 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1398 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1401 // This class implements the "IAllocator" interface, so that we can use
1402 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1404 class CompAllocator : public IAllocator
1407 #if MEASURE_MEM_ALLOC
1411 CompAllocator(Compiler* comp, CompMemKind cmk)
1413 #if MEASURE_MEM_ALLOC
1419 inline void* Alloc(size_t sz);
1421 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1423 // For the compiler's no-release allocator, free operations are no-ops.
1430 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1431 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1433 XX The big guy. The sections are currently organized as : XX
1435 XX o GenTree and BasicBlock XX
1447 XX o PrologScopeInfo XX
1448 XX o CodeGenerator XX
1453 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1454 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1459 friend class emitter;
1460 friend class UnwindInfo;
1461 friend class UnwindFragmentInfo;
1462 friend class UnwindEpilogInfo;
1463 friend class JitTimer;
1464 friend class LinearScan;
1465 friend class fgArgInfo;
1466 friend class Rationalizer;
1468 friend class Lowering;
1469 friend class CSE_DataFlow;
1470 friend class CSE_Heuristic;
1471 friend class CodeGenInterface;
1472 friend class CodeGen;
1473 friend class LclVarDsc;
1474 friend class TempDsc;
1476 friend class ObjectAllocator;
1478 #ifndef _TARGET_64BIT_
1479 friend class DecomposeLongs;
1480 #endif // !_TARGET_64BIT_
1483 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1484 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1486 XX Misc structs definitions XX
1488 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1489 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1493 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1512 bool dumpIRDataflow;
1513 bool dumpIRBlockHeaders;
1515 LPCWSTR dumpIRPhase;
1516 LPCWSTR dumpIRFormat;
1518 bool shouldUseVerboseTrees();
1519 bool asciiTrees; // If true, dump trees using only ASCII characters
1520 bool shouldDumpASCIITrees();
1521 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1522 bool shouldUseVerboseSsa();
1523 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1524 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1526 const char* VarNameToStr(VarName name)
1531 DWORD expensiveDebugCheckLevel;
1534 #if FEATURE_MULTIREG_RET
1535 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1536 #endif // FEATURE_MULTIREG_RET
1539 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1540 #endif // ARM_SOFTFP
1542 //-------------------------------------------------------------------------
1543 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1544 // HFAs are one to four element structs where each element is the same
1545 // type, either all float or all double. They are treated specially
1546 // in the ARM Procedure Call Standard, specifically, they are passed in
1547 // floating-point registers instead of the general purpose registers.
1550 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1551 bool IsHfa(GenTreePtr tree);
1553 var_types GetHfaType(GenTreePtr tree);
1554 unsigned GetHfaCount(GenTreePtr tree);
1556 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1557 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1559 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1560 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1562 //-------------------------------------------------------------------------
1563 // The following is used for validating format of EH table
1567 typedef struct EHNodeDsc* pEHNodeDsc;
1569 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1570 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1583 EHBlockType ehnBlockType; // kind of EH block
1584 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1585 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1586 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1588 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1589 pEHNodeDsc ehnChild; // leftmost nested block
1591 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1592 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1594 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1595 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1597 inline void ehnSetTryNodeType()
1599 ehnBlockType = TryNode;
1601 inline void ehnSetFilterNodeType()
1603 ehnBlockType = FilterNode;
1605 inline void ehnSetHandlerNodeType()
1607 ehnBlockType = HandlerNode;
1609 inline void ehnSetFinallyNodeType()
1611 ehnBlockType = FinallyNode;
1613 inline void ehnSetFaultNodeType()
1615 ehnBlockType = FaultNode;
1618 inline BOOL ehnIsTryBlock()
1620 return ehnBlockType == TryNode;
1622 inline BOOL ehnIsFilterBlock()
1624 return ehnBlockType == FilterNode;
1626 inline BOOL ehnIsHandlerBlock()
1628 return ehnBlockType == HandlerNode;
1630 inline BOOL ehnIsFinallyBlock()
1632 return ehnBlockType == FinallyNode;
1634 inline BOOL ehnIsFaultBlock()
1636 return ehnBlockType == FaultNode;
1639 // returns true if there is any overlap between the two nodes
1640 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1642 if (node1->ehnStartOffset < node2->ehnStartOffset)
1644 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1648 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1652 // fails with BADCODE if inner is not completely nested inside outer
1653 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1655 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1659 //-------------------------------------------------------------------------
1660 // Exception handling functions
1663 #if !FEATURE_EH_FUNCLETS
1665 bool ehNeedsShadowSPslots()
1667 return (info.compXcptnsCount || opts.compDbgEnC);
1670 // 0 for methods with no EH
1671 // 1 for methods with non-nested EH, or where only the try blocks are nested
1672 // 2 for a method with a catch within a catch
1674 unsigned ehMaxHndNestingCount;
1676 #endif // !FEATURE_EH_FUNCLETS
1678 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1679 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1681 bool bbInCatchHandlerILRange(BasicBlock* blk);
1682 bool bbInFilterILRange(BasicBlock* blk);
1683 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1684 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1685 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1686 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1687 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1689 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1690 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1692 // Returns true if "block" is the start of a try region.
1693 bool bbIsTryBeg(BasicBlock* block);
1695 // Returns true if "block" is the start of a handler or filter region.
1696 bool bbIsHandlerBeg(BasicBlock* block);
1698 // Returns true iff "block" is where control flows if an exception is raised in the
1699 // try region, and sets "*regionIndex" to the index of the try for the handler.
1700 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1701 // block of the filter, but not for the filter's handler.
1702 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1704 bool ehHasCallableHandlers();
1706 // Return the EH descriptor for the given region index.
1707 EHblkDsc* ehGetDsc(unsigned regionIndex);
1709 // Return the EH index given a region descriptor.
1710 unsigned ehGetIndex(EHblkDsc* ehDsc);
1712 // Return the EH descriptor index of the enclosing try, for the given region index.
1713 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1715 // Return the EH descriptor index of the enclosing handler, for the given region index.
1716 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1718 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1719 // block is not in a 'try' region).
1720 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1722 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1723 // if this block is not in a filter or handler region).
1724 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1726 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1727 // nullptr if this block's exceptions propagate to caller).
1728 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1730 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1731 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1732 bool ehIsBlockEHLast(BasicBlock* block);
1734 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1736 // Return the region index of the most nested EH region this block is in.
1737 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1739 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1740 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1742 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1743 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1744 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1745 // (It can never be a filter.)
1746 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1748 // A block has been deleted. Update the EH table appropriately.
1749 void ehUpdateForDeletedBlock(BasicBlock* block);
1751 // Determine whether a block can be deleted while preserving the EH normalization rules.
1752 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1754 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1755 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1757 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1758 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1759 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1760 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1761 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1762 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1763 // lives in a filter.)
1764 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1766 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1767 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1768 // (nullptr if the last block is the last block in the program).
1769 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1770 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1773 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1774 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1775 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1778 #if FEATURE_EH_FUNCLETS
1779 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1780 // if there is a filter that protects a region with a nested EH clause (such as a
1781 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1782 // genFuncletProlog() for more details. However, the VM seems to use it for more
1783 // purposes, maybe including debugging. Until we are sure otherwise, always create
1784 // a PSPSym for functions with any EH.
1785 bool ehNeedsPSPSym() const
1789 #else // _TARGET_X86_
1790 return compHndBBtabCount > 0;
1791 #endif // _TARGET_X86_
1794 bool ehAnyFunclets(); // Are there any funclets in this function?
1795 unsigned ehFuncletCount(); // Return the count of funclets in the function
1797 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1798 #else // !FEATURE_EH_FUNCLETS
1799 bool ehAnyFunclets()
1803 unsigned ehFuncletCount()
1808 unsigned bbThrowIndex(BasicBlock* blk)
1810 return blk->bbTryIndex;
1811 } // Get the index to use as the cache key for sharing throw blocks
1812 #endif // !FEATURE_EH_FUNCLETS
1814 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1815 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1816 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1817 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1818 // convenient to also consider it a predecessor.)
1819 flowList* BlockPredsWithEH(BasicBlock* blk);
1821 // This table is useful for memoization of the method above.
1822 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1824 BlockToFlowListMap* m_blockToEHPreds;
1825 BlockToFlowListMap* GetBlockToEHPreds()
1827 if (m_blockToEHPreds == nullptr)
1829 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1831 return m_blockToEHPreds;
1834 void* ehEmitCookie(BasicBlock* block);
1835 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1837 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1839 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1841 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1843 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1845 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1847 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1849 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1851 void fgAllocEHTable();
1853 void fgRemoveEHTableEntry(unsigned XTnum);
1855 #if FEATURE_EH_FUNCLETS
1857 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1859 #endif // FEATURE_EH_FUNCLETS
1863 #endif // !FEATURE_EH
1865 void fgSortEHTable();
1867 // Causes the EH table to obey some well-formedness conditions, by inserting
1868 // empty BB's when necessary:
1869 // * No block is both the first block of a handler and the first block of a try.
1870 // * No block is the first block of multiple 'try' regions.
1871 // * No block is the last block of multiple EH regions.
1872 void fgNormalizeEH();
1873 bool fgNormalizeEHCase1();
1874 bool fgNormalizeEHCase2();
1875 bool fgNormalizeEHCase3();
1878 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1879 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1880 void fgVerifyHandlerTab();
1881 void fgDispHandlerTab();
1884 bool fgNeedToSortEHTable;
1886 void verInitEHTree(unsigned numEHClauses);
1887 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1888 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1889 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1890 void verCheckNestingLevel(EHNodeDsc* initRoot);
1893 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1894 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1896 XX GenTree and BasicBlock XX
1898 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1900 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1901 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1904 // Functions to create nodes
1905 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1908 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1910 // For binary opers.
1911 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1913 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1915 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1916 var_types type = TYP_I_IMPL,
1917 GenTreePtr op1 = nullptr,
1918 GenTreePtr op2 = nullptr);
1920 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1922 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1924 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1926 GenTreePtr gtNewJmpTableNode();
1927 GenTreePtr gtNewIconHandleNode(
1928 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1930 unsigned gtTokenToIconFlags(unsigned token);
1932 GenTreePtr gtNewIconEmbHndNode(void* value,
1935 unsigned handle1 = 0,
1936 void* handle2 = nullptr,
1937 void* compileTimeHandle = nullptr);
1939 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1940 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1941 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1942 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1944 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1946 GenTreePtr gtNewLconNode(__int64 value);
1948 GenTreePtr gtNewDconNode(double value);
1950 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1952 GenTreePtr gtNewZeroConNode(var_types type);
1954 GenTreePtr gtNewOneConNode(var_types type);
1957 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1958 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1961 GenTreeBlk* gtNewBlkOpNode(
1962 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1964 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1967 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1970 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1971 void gtSetObjGcInfo(GenTreeObj* objNode);
1972 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1973 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1975 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1977 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1979 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1980 CORINFO_METHOD_HANDLE handle,
1982 GenTreeArgList* args,
1983 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1985 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1987 GenTreeArgList* args,
1988 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1990 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1993 GenTreeArgList* args = nullptr);
1995 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1998 GenTreeSIMD* gtNewSIMDNode(
1999 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2000 GenTreeSIMD* gtNewSIMDNode(var_types type,
2003 SIMDIntrinsicID simdIntrinsicID,
2006 void SetOpLclRelatedToSIMDIntrinsic(GenTreePtr op);
2009 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2010 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2011 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
2013 GenTreePtr gtNewCodeRef(BasicBlock* block);
2015 GenTreePtr gtNewFieldRef(
2016 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2018 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2020 GenTreeArgList* gtNewArgList(GenTreePtr op);
2021 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2022 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2024 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2025 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2026 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2027 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2029 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2031 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2033 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2034 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2035 CORINFO_ACCESS_FLAGS access,
2036 CORINFO_FIELD_INFO* pFieldInfo,
2038 CORINFO_CLASS_HANDLE structType,
2041 GenTreePtr gtNewNothingNode();
2043 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2045 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2047 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2049 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2051 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2053 //------------------------------------------------------------------------
2054 // Other GenTree functions
2056 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2058 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2059 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2060 // IntCnses with value `deepVarVal`.
2061 GenTreePtr gtCloneExpr(
2062 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2064 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2065 // `varNum` to int constants with value `varVal`.
2066 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2068 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2071 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2073 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2075 // Returns "true" iff the complexity (not formally defined, but first interpretation
2076 // is #of nodes in subtree) of "tree" is greater than "limit".
2077 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2078 // before they have been set.)
2079 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2081 bool gtCompareTree(GenTree* op1, GenTree* op2);
2083 GenTreePtr gtReverseCond(GenTree* tree);
2085 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2087 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2089 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2091 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2094 unsigned gtHashValue(GenTree* tree);
2096 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2099 void gtPrepareCost(GenTree* tree);
2100 bool gtIsLikelyRegVar(GenTree* tree);
2102 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2104 // Returns true iff the secondNode can be swapped with firstNode.
2105 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2107 unsigned gtSetEvalOrder(GenTree* tree);
2109 #if FEATURE_STACK_FP_X87
2111 void gtComputeFPlvls(GenTreePtr tree);
2112 #endif // FEATURE_STACK_FP_X87
2114 void gtSetStmtInfo(GenTree* stmt);
2116 // Returns "true" iff "node" has any of the side effects in "flags".
2117 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2119 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2120 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2122 // Appends 'expr' in front of 'list'
2123 // 'list' will typically start off as 'nullptr'
2124 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2125 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2127 void gtExtractSideEffList(GenTreePtr expr,
2129 unsigned flags = GTF_SIDE_EFFECT,
2130 bool ignoreRoot = false);
2132 GenTreePtr gtGetThisArg(GenTreePtr call);
2134 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2135 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2136 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2137 // the given "fldHnd", is such an object pointer.
2138 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2140 // Return true if call is a recursive call; return false otherwise.
2141 // Note when inlining, this looks for calls back to the root method.
2142 bool gtIsRecursiveCall(GenTreeCall* call)
2144 return (call->gtCallMethHnd == impInlineRoot()->info.compMethodHnd);
2147 //-------------------------------------------------------------------------
2149 GenTreePtr gtFoldExpr(GenTreePtr tree);
2152 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2153 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2154 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2155 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2156 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2157 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2158 // optimizations for now.
2159 __attribute__((optnone))
2161 gtFoldExprConst(GenTreePtr tree);
2162 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2163 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2165 //-------------------------------------------------------------------------
2166 // Get the handle, if any.
2167 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2168 // Get the handle, and assert if not found.
2169 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2171 //-------------------------------------------------------------------------
2172 // Functions to display the trees
2175 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2177 void gtDispVN(GenTreePtr tree);
2178 void gtDispConst(GenTreePtr tree);
2179 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2180 void gtDispNodeName(GenTreePtr tree);
2181 void gtDispRegVal(GenTreePtr tree);
2193 void gtDispChild(GenTreePtr child,
2194 IndentStack* indentStack,
2196 __in_opt const char* msg = nullptr,
2197 bool topOnly = false);
2198 void gtDispTree(GenTreePtr tree,
2199 IndentStack* indentStack = nullptr,
2200 __in_opt const char* msg = nullptr,
2201 bool topOnly = false,
2202 bool isLIR = false);
2203 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2204 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2205 char* gtGetLclVarName(unsigned lclNum);
2206 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2207 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2208 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2209 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2210 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2211 void gtDispFieldSeq(FieldSeqNode* pfsn);
2213 void gtDispRange(LIR::ReadOnlyRange const& range);
2215 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2217 void gtDispLIRNode(GenTree* node);
2229 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2230 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2233 static fgWalkPreFn gtAssertColonCond;
2235 static fgWalkPreFn gtMarkColonCond;
2236 static fgWalkPreFn gtClearColonCond;
2238 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2239 bool gtHasCatchArg(GenTreePtr tree);
2240 bool gtHasUnmanagedCall(GenTreePtr tree);
2242 typedef ArrayStack<GenTree*> GenTreeStack;
2244 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2245 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2247 //=========================================================================
2248 // BasicBlock functions
2250 // This is a debug flag we will use to assert when creating block during codegen
2251 // as this interferes with procedure splitting. If you know what you're doing, set
2252 // it to true before creating the block. (DEBUG only)
2253 bool fgSafeBasicBlockCreation;
2256 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2259 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2260 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2264 XX The variables to be used by the code generator. XX
2266 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2267 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2271 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2272 // be placed in the stack frame and it's fields must be laid out sequentially.
2274 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2275 // a local variable that can be enregistered or placed in the stack frame.
2276 // The fields do not need to be laid out sequentially
2278 enum lvaPromotionType
2280 PROMOTION_TYPE_NONE, // The struct local is not promoted
2281 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2282 // and its field locals are independent of its parent struct local.
2283 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2284 // but its field locals depend on its parent struct local.
2287 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2288 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2290 /*****************************************************************************/
2292 enum FrameLayoutState
2295 INITIAL_FRAME_LAYOUT,
2296 PRE_REGALLOC_FRAME_LAYOUT,
2297 REGALLOC_FRAME_LAYOUT,
2298 TENTATIVE_FRAME_LAYOUT,
2303 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2304 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2305 bool lvaSortAgain; // true: We need to sort the lvaTable
2306 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2307 unsigned lvaCount; // total number of locals
2309 unsigned lvaRefCount; // total number of references to locals
2310 LclVarDsc* lvaTable; // variable descriptor table
2311 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2313 LclVarDsc** lvaRefSorted; // table sorted by refcount
2315 unsigned short lvaTrackedCount; // actual # of locals being tracked
2316 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2318 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2319 // Only for AMD64 System V cache the first caller stack homed argument.
2320 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2321 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2324 VARSET_TP lvaTrackedVars; // set of tracked variables
2326 #ifndef _TARGET_64BIT_
2327 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2329 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2331 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2332 // It that changes, this changes. VarSets from different epochs
2333 // cannot be meaningfully combined.
2335 unsigned GetCurLVEpoch()
2340 // reverse map of tracked number to var number
2341 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2343 #ifdef LEGACY_BACKEND
2344 // variable interference graph
2345 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2348 // variable preference graph
2349 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2353 // # of procs compiled a with double-aligned stack
2354 static unsigned s_lvaDoubleAlignedProcsCount;
2358 // Getters and setters for address-exposed and do-not-enregister local var properties.
2359 bool lvaVarAddrExposed(unsigned varNum);
2360 void lvaSetVarAddrExposed(unsigned varNum);
2361 bool lvaVarDoNotEnregister(unsigned varNum);
2363 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2364 enum DoNotEnregisterReason
2369 DNER_VMNeedsStackAddr,
2370 DNER_LiveInOutOfHandler,
2371 DNER_LiveAcrossUnmanagedCall,
2372 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2373 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2374 #ifdef JIT32_GCENCODER
2379 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2381 unsigned lvaVarargsHandleArg;
2383 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2385 #endif // _TARGET_X86_
2387 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2388 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2389 #if FEATURE_FIXED_OUT_ARGS
2390 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2392 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2393 // that tracks whether the lock has been taken
2395 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2396 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2397 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2399 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2400 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2402 #if FEATURE_FIXED_OUT_ARGS
2403 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2404 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2405 #endif // FEATURE_FIXED_OUT_ARGS
2408 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2409 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2410 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2411 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2412 // this variable to be this scratch word whenever struct promotion occurs.
2413 unsigned lvaPromotedStructAssemblyScratchVar;
2414 #endif // _TARGET_ARM_
2417 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2418 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2421 bool lvaGenericsContextUsed;
2423 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2424 // CORINFO_GENERICS_CTXT_FROM_THIS?
2425 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2427 //-------------------------------------------------------------------------
2428 // All these frame offsets are inter-related and must be kept in sync
2430 #if !FEATURE_EH_FUNCLETS
2431 // This is used for the callable handlers
2432 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2433 #endif // FEATURE_EH_FUNCLETS
2435 unsigned lvaCachedGenericContextArgOffs;
2436 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2439 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2441 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2443 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2444 // after the reg predict we will use a computed maxTmpSize
2445 // which is based upon the number of spill temps predicted by reg predict
2446 // All this is necessary because if we under-estimate the size of the spill
2447 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2449 // Pre codegen max spill temp size.
2450 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2452 //-------------------------------------------------------------------------
2454 unsigned lvaGetMaxSpillTempSize();
2456 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2457 #endif // _TARGET_ARM_
2458 void lvaAssignFrameOffsets(FrameLayoutState curState);
2459 void lvaFixVirtualFrameOffsets();
2461 #ifndef LEGACY_BACKEND
2462 void lvaUpdateArgsWithInitialReg();
2463 #endif // !LEGACY_BACKEND
2465 void lvaAssignVirtualFrameOffsetsToArgs();
2466 #ifdef UNIX_AMD64_ABI
2467 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2468 #else // !UNIX_AMD64_ABI
2469 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2470 #endif // !UNIX_AMD64_ABI
2471 void lvaAssignVirtualFrameOffsetsToLocals();
2472 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2473 #ifdef _TARGET_AMD64_
2474 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2475 bool lvaIsCalleeSavedIntRegCountEven();
2477 void lvaAlignFrame();
2478 void lvaAssignFrameOffsetsToPromotedStructs();
2479 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2482 void lvaDumpRegLocation(unsigned lclNum);
2483 void lvaDumpFrameLocation(unsigned lclNum);
2484 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2485 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2486 // layout state defined by lvaDoneFrameLayout
2489 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2490 // to avoid bugs from borderline cases.
2491 #define MAX_FrameSize 0x3FFFFFFF
2492 void lvaIncrementFrameSize(unsigned size);
2494 unsigned lvaFrameSize(FrameLayoutState curState);
2496 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2497 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2499 // Returns the caller-SP-relative offset for the local variable "varNum."
2500 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2502 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2503 int lvaGetSPRelativeOffset(unsigned varNum);
2505 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2506 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2508 //------------------------ For splitting types ----------------------------
2510 void lvaInitTypeRef();
2512 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2513 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2514 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2515 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2516 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2517 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2519 void lvaInitVarDsc(LclVarDsc* varDsc,
2521 CorInfoType corInfoType,
2522 CORINFO_CLASS_HANDLE typeHnd,
2523 CORINFO_ARG_LIST_HANDLE varList,
2524 CORINFO_SIG_INFO* varSig);
2526 static unsigned lvaTypeRefMask(var_types type);
2528 var_types lvaGetActualType(unsigned lclNum);
2529 var_types lvaGetRealType(unsigned lclNum);
2531 //-------------------------------------------------------------------------
2535 unsigned lvaLclSize(unsigned varNum);
2536 unsigned lvaLclExactSize(unsigned varNum);
2538 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2540 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2541 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2542 // the return result.
2543 bool lvaLclVarRefsAccum(
2544 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2546 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2547 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2548 // and (destructively) unions "trkedVars" into "*result".
2549 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2551 ALLVARSET_VALARG_TP allVars,
2552 VARSET_VALARG_TP trkdVars);
2554 bool lvaHaveManyLocals() const;
2556 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2557 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2558 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2561 void lvaSortByRefCount();
2562 void lvaDumpRefCounts();
2564 void lvaMarkLocalVars(BasicBlock* block);
2566 void lvaMarkLocalVars(); // Local variable ref-counting
2568 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2570 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2572 static fgWalkPreFn lvaIncRefCntsCB;
2573 void lvaIncRefCnts(GenTreePtr tree);
2575 static fgWalkPreFn lvaDecRefCntsCB;
2576 void lvaDecRefCnts(GenTreePtr tree);
2577 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2578 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2579 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2582 struct lvaStressLclFldArgs
2584 Compiler* m_pCompiler;
2588 static fgWalkPreFn lvaStressLclFldCB;
2589 void lvaStressLclFld();
2591 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2592 void lvaDispVarSet(VARSET_VALARG_TP set);
2597 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2599 int lvaFrameAddress(int varNum, bool* pFPbased);
2602 bool lvaIsParameter(unsigned varNum);
2603 bool lvaIsRegArgument(unsigned varNum);
2604 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2605 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2606 // that writes to arg0
2608 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2609 // (this is an overload of lvIsTemp because there are no temp parameters).
2610 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2611 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2612 bool lvaIsImplicitByRefLocal(unsigned varNum)
2614 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2615 LclVarDsc* varDsc = &(lvaTable[varNum]);
2616 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2618 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2621 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2625 // Returns true if this local var is a multireg struct
2626 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2628 // If the class is a TYP_STRUCT, get/set a class handle describing it
2630 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2631 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2633 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2635 // Info about struct fields
2636 struct lvaStructFieldInfo
2638 CORINFO_FIELD_HANDLE fldHnd;
2639 unsigned char fldOffset;
2640 unsigned char fldOrdinal;
2643 CORINFO_CLASS_HANDLE fldTypeHnd;
2646 // Info about struct to be promoted.
2647 struct lvaStructPromotionInfo
2649 CORINFO_CLASS_HANDLE typeHnd;
2651 bool requiresScratchVar;
2654 unsigned char fieldCnt;
2655 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2657 lvaStructPromotionInfo()
2658 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2663 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2664 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2665 lvaStructPromotionInfo* StructPromotionInfo,
2667 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2668 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2669 #if !defined(_TARGET_64BIT_)
2670 void lvaPromoteLongVars();
2671 #endif // !defined(_TARGET_64BIT_)
2672 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2673 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2674 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2675 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2676 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2677 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2678 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2680 #if defined(FEATURE_SIMD)
2681 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2683 assert(varDsc->lvType == TYP_SIMD12);
2684 assert(varDsc->lvExactSize == 12);
2686 #if defined(_TARGET_64BIT_)
2687 assert(varDsc->lvSize() == 16);
2689 #else // !defined(_TARGET_64BIT_)
2691 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2692 // already does this calculation. However, we also need to prevent mapping types if the var is a
2693 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2694 // However, we don't know this until late, so we may have already pretended the field is bigger
2696 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2705 #endif // !defined(_TARGET_64BIT_)
2707 #endif // defined(FEATURE_SIMD)
2709 BYTE* lvaGetGcLayout(unsigned varNum);
2710 bool lvaTypeIsGC(unsigned varNum);
2711 unsigned lvaGSSecurityCookie; // LclVar number
2712 bool lvaTempsHaveLargerOffsetThanVars();
2714 unsigned lvaSecurityObject; // variable representing the security object on the stack
2715 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2717 #if FEATURE_EH_FUNCLETS
2718 unsigned lvaPSPSym; // variable representing the PSPSym
2721 InlineInfo* impInlineInfo;
2722 InlineStrategy* m_inlineStrategy;
2724 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2725 Compiler* impInlineRoot();
2727 #if defined(DEBUG) || defined(INLINE_DATA)
2728 unsigned __int64 getInlineCycleCount()
2730 return m_compCycles;
2732 #endif // defined(DEBUG) || defined(INLINE_DATA)
2734 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2735 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2737 //=========================================================================
2739 //=========================================================================
2742 //---------------- Local variable ref-counting ----------------------------
2745 BasicBlock* lvaMarkRefsCurBlock;
2746 GenTreePtr lvaMarkRefsCurStmt;
2748 BasicBlock::weight_t lvaMarkRefsWeight;
2750 static fgWalkPreFn lvaMarkLclRefsCallback;
2751 void lvaMarkLclRefs(GenTreePtr tree);
2753 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2754 PerSsaArray lvMemoryPerSsaData;
2755 unsigned lvMemoryNumSsaNames;
2758 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2759 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2760 // not an SSA variable).
2761 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2763 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2764 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2766 assert(ssaNum < lvMemoryNumSsaNames);
2767 return &lvMemoryPerSsaData.GetRef(ssaNum);
2771 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2772 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2776 XX Imports the given method and converts it to semantic trees XX
2778 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2779 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2785 void impImport(BasicBlock* method);
2787 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2788 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2789 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2790 CORINFO_CLASS_HANDLE impGetStringClass();
2791 CORINFO_CLASS_HANDLE impGetObjectClass();
2793 //=========================================================================
2795 //=========================================================================
2798 //-------------------- Stack manipulation ---------------------------------
2800 unsigned impStkSize; // Size of the full stack
2802 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2804 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2806 struct SavedStack // used to save/restore stack contents.
2808 unsigned ssDepth; // number of values on stack
2809 StackEntry* ssTrees; // saved tree values
2812 bool impIsPrimitive(CorInfoType type);
2813 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2815 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2816 void impPushOnStackNoType(GenTreePtr tree);
2818 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2819 void impPushNullObjRefOnStack();
2820 StackEntry impPopStack();
2821 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2822 GenTreePtr impPopStack(typeInfo& ti);
2823 StackEntry& impStackTop(unsigned n = 0);
2825 void impSaveStackState(SavedStack* savePtr, bool copy);
2826 void impRestoreStackState(SavedStack* savePtr);
2828 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2829 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2830 CORINFO_CALL_INFO* pCallInfo);
2832 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2834 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2836 bool impCanPInvokeInline();
2837 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2838 void impCheckForPInvokeCall(
2839 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2840 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2841 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2843 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2844 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2845 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2847 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2848 CORINFO_METHOD_HANDLE calleeMethodHnd,
2849 CORINFO_CLASS_HANDLE delegateTypeHnd);
2851 var_types impImportCall(OPCODE opcode,
2852 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2853 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2855 GenTreePtr newobjThis,
2857 CORINFO_CALL_INFO* callInfo,
2858 IL_OFFSET rawILOffset);
2860 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2862 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2864 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2867 var_types impImportJitTestLabelMark(int numArgs);
2870 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2872 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2874 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2875 CORINFO_ACCESS_FLAGS access,
2876 CORINFO_FIELD_INFO* pFieldInfo,
2879 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2881 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2883 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2885 void impImportLeave(BasicBlock* block);
2886 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2887 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2888 CORINFO_CLASS_HANDLE clsHnd,
2889 CORINFO_METHOD_HANDLE method,
2890 CORINFO_SIG_INFO* sig,
2894 CorInfoIntrinsics* pIntrinsicID);
2895 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2896 CORINFO_SIG_INFO* sig,
2899 CorInfoIntrinsics intrinsicID);
2900 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2902 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2904 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2905 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2906 CORINFO_THIS_TRANSFORM transform);
2908 //----------------- Manipulating the trees and stmts ----------------------
2910 GenTreePtr impTreeList; // Trees for the BB being imported
2911 GenTreePtr impTreeLast; // The last tree for the current BB
2915 CHECK_SPILL_ALL = -1,
2916 CHECK_SPILL_NONE = -2
2920 void impBeginTreeList();
2921 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2922 void impEndTreeList(BasicBlock* block);
2923 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2924 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2925 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2926 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2927 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2928 void impAssignTempGen(unsigned tmp,
2931 GenTreePtr* pAfterStmt = nullptr,
2932 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2933 BasicBlock* block = nullptr);
2934 void impAssignTempGen(unsigned tmpNum,
2936 CORINFO_CLASS_HANDLE structHnd,
2938 GenTreePtr* pAfterStmt = nullptr,
2939 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2940 BasicBlock* block = nullptr);
2941 GenTreePtr impCloneExpr(GenTreePtr tree,
2943 CORINFO_CLASS_HANDLE structHnd,
2945 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2946 GenTreePtr impAssignStruct(GenTreePtr dest,
2948 CORINFO_CLASS_HANDLE structHnd,
2950 GenTreePtr* pAfterStmt = nullptr,
2951 BasicBlock* block = nullptr);
2952 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2954 CORINFO_CLASS_HANDLE structHnd,
2956 GenTreePtr* pAfterStmt = nullptr,
2957 BasicBlock* block = nullptr);
2959 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2960 CORINFO_CLASS_HANDLE structHnd,
2964 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2965 BYTE* gcLayout = nullptr,
2966 unsigned* numGCVars = nullptr,
2967 var_types* simdBaseType = nullptr);
2969 GenTreePtr impNormStructVal(GenTreePtr structVal,
2970 CORINFO_CLASS_HANDLE structHnd,
2972 bool forceNormalization = false);
2974 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2975 BOOL* pRuntimeLookup = nullptr,
2976 BOOL mustRestoreHandle = FALSE,
2977 BOOL importParent = FALSE);
2979 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2980 BOOL* pRuntimeLookup = nullptr,
2981 BOOL mustRestoreHandle = FALSE)
2983 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2986 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2987 CORINFO_LOOKUP* pLookup,
2989 void* compileTimeHandle);
2991 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2993 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2994 CORINFO_LOOKUP* pLookup,
2995 void* compileTimeHandle);
2997 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2999 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3000 CorInfoHelpFunc helper,
3002 GenTreeArgList* arg = nullptr,
3003 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3005 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3007 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3010 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3011 CORINFO_CLASS_HANDLE typeClass,
3015 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3016 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3017 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3018 static bool IsMathIntrinsic(GenTreePtr tree);
3021 //----------------- Importing the method ----------------------------------
3023 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3026 unsigned impCurOpcOffs;
3027 const char* impCurOpcName;
3028 bool impNestedStackSpill;
3030 // For displaying instrs with generated native code (-n:B)
3031 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3032 void impNoteLastILoffs();
3035 /* IL offset of the stmt currently being imported. It gets set to
3036 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3037 updated at IL offsets for which we have to report mapping info.
3038 It also includes flag bits, so use jitGetILoffs()
3039 to get the actual IL offset value.
3042 IL_OFFSETX impCurStmtOffs;
3043 void impCurStmtOffsSet(IL_OFFSET offs);
3045 void impNoteBranchOffs();
3047 unsigned impInitBlockLineInfo();
3049 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3050 bool impIsThis(GenTreePtr obj);
3051 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3052 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3053 bool impIsAnySTLOC(OPCODE opcode)
3055 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3056 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3059 GenTreeArgList* impPopList(unsigned count,
3061 CORINFO_SIG_INFO* sig,
3062 GenTreeArgList* prefixTree = nullptr);
3064 GenTreeArgList* impPopRevList(unsigned count,
3066 CORINFO_SIG_INFO* sig,
3067 unsigned skipReverseCount = 0);
3070 * Get current IL offset with stack-empty info incoporated
3072 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3074 //---------------- Spilling the importer stack ----------------------------
3080 SavedStack pdSavedStack;
3081 ThisInitState pdThisPtrInit;
3084 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3085 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3087 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3088 ExpandArray<BYTE> impPendingBlockMembers;
3090 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3091 // Operates on the map in the top-level ancestor.
3092 BYTE impGetPendingBlockMember(BasicBlock* blk)
3094 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3097 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3098 // Operates on the map in the top-level ancestor.
3099 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3101 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3104 bool impCanReimport;
3106 bool impSpillStackEntry(unsigned level,
3110 bool bAssertOnRecursion,
3115 void impSpillStackEnsure(bool spillLeaves = false);
3116 void impEvalSideEffects();
3117 void impSpillSpecialSideEff();
3118 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3119 void impSpillValueClasses();
3120 void impSpillEvalStack();
3121 static fgWalkPreFn impFindValueClasses;
3122 void impSpillLclRefs(ssize_t lclNum);
3124 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3126 void impImportBlockCode(BasicBlock* block);
3128 void impReimportMarkBlock(BasicBlock* block);
3129 void impReimportMarkSuccessors(BasicBlock* block);
3131 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3133 void impImportBlockPending(BasicBlock* block);
3135 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3136 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3137 // for the block, but instead, just re-uses the block's existing EntryState.
3138 void impReimportBlockPending(BasicBlock* block);
3140 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3142 void impImportBlock(BasicBlock* block);
3144 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3145 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3146 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3147 // the variables that will be used -- and for all the predecessors of those successors, and the
3148 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3149 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3150 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3151 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3152 // of local variable numbers, so we represent them with the base local variable number), returns that.
3153 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3154 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3155 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3156 // on which kind of member of the clique the block is).
3157 unsigned impGetSpillTmpBase(BasicBlock* block);
3159 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3160 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3161 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3162 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3163 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3164 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3165 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3166 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3167 // successors receive a native int. Similarly float and double are unified to double.
3168 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3169 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3170 // predecessors, so they insert an upcast if needed).
3171 void impReimportSpillClique(BasicBlock* block);
3173 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3174 // block, and represent the predecessor and successor members of the clique currently being computed.
3175 // *** Access to these will need to be locked in a parallel compiler.
3176 ExpandArray<BYTE> impSpillCliquePredMembers;
3177 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3185 // Abstract class for receiving a callback while walking a spill clique
3186 class SpillCliqueWalker
3189 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3192 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3193 class SetSpillTempsBase : public SpillCliqueWalker
3198 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3201 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3204 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3205 class ReimportSpillClique : public SpillCliqueWalker
3210 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3213 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3216 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3217 // predecessor or successor within the spill clique
3218 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3220 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3221 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3222 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3223 void impRetypeEntryStateTemps(BasicBlock* blk);
3225 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3226 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3228 void impPushVar(GenTree* op, typeInfo tiRetVal);
3229 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3230 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3232 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3234 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3235 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3236 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3239 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3242 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3243 struct BlockListNode
3246 BlockListNode* m_next;
3247 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3250 void* operator new(size_t sz, Compiler* comp);
3252 BlockListNode* impBlockListNodeFreeList;
3254 BlockListNode* AllocBlockListNode();
3255 void FreeBlockListNode(BlockListNode* node);
3257 bool impIsValueType(typeInfo* pTypeInfo);
3258 var_types mangleVarArgsType(var_types type);
3261 regNumber getCallArgIntRegister(regNumber floatReg);
3262 regNumber getCallArgFloatRegister(regNumber intReg);
3263 #endif // FEATURE_VARARG
3266 static unsigned jitTotalMethodCompiled;
3270 static LONG jitNestingLevel;
3273 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3275 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3277 // STATIC inlining decision based on the IL code.
3278 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3279 CORINFO_METHOD_INFO* methInfo,
3281 InlineResult* inlineResult);
3283 void impCheckCanInline(GenTreePtr call,
3284 CORINFO_METHOD_HANDLE fncHandle,
3286 CORINFO_CONTEXT_HANDLE exactContextHnd,
3287 InlineCandidateInfo** ppInlineCandidateInfo,
3288 InlineResult* inlineResult);
3290 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3291 GenTreePtr curArgVal,
3293 InlineResult* inlineResult);
3295 void impInlineInitVars(InlineInfo* pInlineInfo);
3297 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3299 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3301 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3303 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3304 GenTreePtr variableBeingDereferenced,
3305 InlArgInfo* inlArgInfo);
3307 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3309 bool impTailCallRetTypeCompatible(var_types callerRetType,
3310 CORINFO_CLASS_HANDLE callerRetTypeClass,
3311 var_types calleeRetType,
3312 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3314 bool impIsTailCallILPattern(bool tailPrefixed,
3316 const BYTE* codeAddrOfNextOpcode,
3317 const BYTE* codeEnd,
3319 bool* IsCallPopRet = nullptr);
3321 bool impIsImplicitTailCallCandidate(
3322 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3325 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3326 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3330 XX Info about the basic-blocks, their contents and the flow analysis XX
3332 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3333 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3337 BasicBlock* fgFirstBB; // Beginning of the basic block list
3338 BasicBlock* fgLastBB; // End of the basic block list
3339 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3340 #if FEATURE_EH_FUNCLETS
3341 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3343 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3345 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3346 unsigned fgEdgeCount; // # of control flow edges between the BBs
3347 unsigned fgBBcount; // # of BBs in the method
3349 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3351 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3352 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3353 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3354 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3356 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3357 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3358 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3359 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3360 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3361 // index). The arrays are of size fgBBNumMax + 1.
3362 unsigned* fgDomTreePreOrder;
3363 unsigned* fgDomTreePostOrder;
3365 bool fgBBVarSetsInited;
3367 // Allocate array like T* a = new T[fgBBNumMax + 1];
3368 // Using helper so we don't keep forgetting +1.
3369 template <typename T>
3370 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3372 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3375 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3376 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3377 // cannot be meaningfully combined. Note that new blocks can be created with higher
3378 // block numbers without changing the basic block epoch. These blocks *cannot*
3379 // participate in a block set until the blocks are all renumbered, causing the epoch
3380 // to change. This is useful if continuing to use previous block sets is valuable.
3381 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3382 unsigned fgCurBBEpoch;
3384 unsigned GetCurBasicBlockEpoch()
3386 return fgCurBBEpoch;
3389 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3390 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3391 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3392 unsigned fgCurBBEpochSize;
3394 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3395 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3396 unsigned fgBBSetCountInSizeTUnits;
3398 void NewBasicBlockEpoch()
3400 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3402 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3404 fgCurBBEpochSize = fgBBNumMax + 1;
3405 fgBBSetCountInSizeTUnits =
3406 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3409 // All BlockSet objects are now invalid!
3410 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3411 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3415 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3416 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3417 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3418 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3420 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3421 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3422 // array of size_t bitsets), then print that out.
3423 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3430 void EnsureBasicBlockEpoch()
3432 if (fgCurBBEpochSize != fgBBNumMax + 1)
3434 NewBasicBlockEpoch();
3438 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3439 void fgEnsureFirstBBisScratch();
3440 bool fgFirstBBisScratch();
3441 bool fgBBisScratch(BasicBlock* block);
3443 void fgExtendEHRegionBefore(BasicBlock* block);
3444 void fgExtendEHRegionAfter(BasicBlock* block);
3446 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3448 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3450 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3453 BasicBlock* nearBlk,
3454 bool putInFilter = false,
3455 bool runRarely = false,
3456 bool insertAtEnd = false);
3458 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3460 bool runRarely = false,
3461 bool insertAtEnd = false);
3463 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3465 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3466 BasicBlock* afterBlk,
3467 unsigned xcptnIndex,
3468 bool putInTryRegion);
3470 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3471 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3472 void fgUnlinkBlock(BasicBlock* block);
3474 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3475 bool fgMultipleNots;
3478 bool fgModified; // True if the flow graph has been modified recently
3479 bool fgComputePredsDone; // Have we computed the bbPreds list
3480 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3481 bool fgDomsComputed; // Have we computed the dominator sets?
3482 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3484 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3485 bool fgHasPostfix; // any postfix ++/-- found?
3486 unsigned fgIncrCount; // number of increment nodes found
3488 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3492 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3493 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3496 bool fgRemoveRestOfBlock; // true if we know that we will throw
3497 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3499 // There are two modes for ordering of the trees.
3500 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3501 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3502 // by traversing the tree according to the order of the operands.
3503 // - In FGOrderLinear, the dominant ordering is the linear order.
3510 FlowGraphOrder fgOrder;
3512 // The following are boolean flags that keep track of the state of internal data structures
3514 bool fgStmtListThreaded;
3515 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3516 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3517 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3518 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3519 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3520 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3521 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3522 // This is derived from the profile data
3523 // or is BB_UNITY_WEIGHT when we don't have profile data
3525 #if FEATURE_EH_FUNCLETS
3526 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3527 #endif // FEATURE_EH_FUNCLETS
3529 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3530 // since fgMorphTree can be called from several places
3531 bool fgExpandInline; // indicates that we are creating tree for the inliner
3533 bool impBoxTempInUse; // the temp below is valid and available
3534 unsigned impBoxTemp; // a temporary that is used for boxing
3537 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3538 // and we are trying to compile again in a "safer", minopts mode?
3542 unsigned impInlinedCodeSize;
3545 //-------------------------------------------------------------------------
3551 void fgTransformFatCalli();
3555 void fgRemoveEmptyTry();
3557 void fgRemoveEmptyFinally();
3559 void fgCloneFinally();
3561 void fgCleanupContinuation(BasicBlock* continuation);
3563 void fgUpdateFinallyTargetFlags();
3565 GenTreePtr fgGetCritSectOfStaticMethod();
3567 #if !defined(_TARGET_X86_)
3569 void fgAddSyncMethodEnterExit();
3571 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3573 void fgConvertSyncReturnToLeave(BasicBlock* block);
3575 #endif // !_TARGET_X86_
3577 void fgAddReversePInvokeEnterExit();
3579 bool fgMoreThanOneReturnBlock();
3581 // The number of separate return points in the method.
3582 unsigned fgReturnCount;
3584 void fgAddInternal();
3586 bool fgFoldConditional(BasicBlock* block);
3588 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3589 void fgMorphBlocks();
3591 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3593 void fgCheckArgCnt();
3594 void fgSetOptions();
3597 static fgWalkPreFn fgAssertNoQmark;
3598 void fgPreExpandQmarkChecks(GenTreePtr expr);
3599 void fgPostExpandQmarkChecks();
3600 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3603 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3605 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3606 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3607 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3608 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3609 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3611 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3612 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3613 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3614 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3616 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3617 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3618 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3619 void fgExpandQmarkNodes();
3623 // Do "simple lowering." This functionality is (conceptually) part of "general"
3624 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3625 void fgSimpleLowering();
3627 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3629 GenTreePtr fgInitThisClass();
3631 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3633 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3635 void fgLocalVarLiveness();
3637 void fgLocalVarLivenessInit();
3639 #ifdef LEGACY_BACKEND
3640 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3642 void fgPerNodeLocalVarLiveness(GenTree* node);
3644 void fgPerBlockLocalVarLiveness();
3646 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3648 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3650 // This is used in the liveness computation, as a temporary. When we use the
3651 // arbitrary-length VarSet representation, it is better not to allocate a new one
3653 VARSET_TP fgMarkIntfUnionVS;
3655 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3657 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3659 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3661 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3663 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3665 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3667 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3668 GenTreePtr startNode,
3670 VARSET_VALARG_TP volatileVars,
3671 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3673 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3675 bool fgRemoveDeadStore(GenTree** pTree,
3679 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3681 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3683 // For updating liveset during traversal AFTER fgComputeLife has completed
3684 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3685 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3687 // Returns the set of live variables after endTree,
3688 // assuming that liveSet is the set of live variables BEFORE tree.
3689 // Requires that fgComputeLife has completed, and that tree is in the same
3690 // statement as endTree, and that it comes before endTree in execution order
3692 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3694 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3695 while (tree != nullptr && tree != endTree->gtNext)
3697 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3698 tree = tree->gtNext;
3700 assert(tree == endTree->gtNext);
3704 void fgInterBlockLocalVarLiveness();
3706 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3707 // "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
3708 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3709 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3710 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3711 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3712 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3714 if (m_opAsgnVarDefSsaNums == nullptr)
3716 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3718 return m_opAsgnVarDefSsaNums;
3721 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3722 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3723 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3725 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3727 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3728 // Except: assumes that lcl is a def, and if it is
3729 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3730 // rather than the "use" SSA number recorded in the tree "lcl".
3731 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3733 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3734 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3735 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3736 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3737 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3739 // (byref addrS1 = &s1,
3740 // *(addrS1 * offsetof(f0)) = s2f0,
3742 // *(addrS1 * offsetof(fn)) = s2fn)
3744 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3745 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3746 // give it SSA names and value numbers?
3748 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3749 // end with an instance of the structure below, whose fields are described in the declaration.
3750 struct IndirectAssignmentAnnotation
3752 unsigned m_lclNum; // The local num that is being indirectly assigned.
3753 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3754 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3755 // be the singleton field sequence "g". The individual assignments would
3756 // further append the fields of "s.g" to that.
3757 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3758 // structure has a single field).
3759 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3760 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3763 IndirectAssignmentAnnotation(unsigned lclNum,
3764 FieldSeqNode* fldSeq,
3766 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3767 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3768 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3772 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3773 NodeToIndirAssignMap;
3774 NodeToIndirAssignMap* m_indirAssignMap;
3775 NodeToIndirAssignMap* GetIndirAssignMap()
3777 if (m_indirAssignMap == nullptr)
3779 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3780 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3781 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3783 return m_indirAssignMap;
3786 // Performs SSA conversion.
3789 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3790 void fgResetForSsa();
3792 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3794 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3795 inline bool fgExcludeFromSsa(unsigned lclNum);
3797 // The value numbers for this compilation.
3798 ValueNumStore* vnStore;
3801 ValueNumStore* GetValueNumStore()
3806 // Do value numbering (assign a value number to each
3808 void fgValueNumber();
3810 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3811 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3812 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3813 // match the element type of the array or fldSeq. When this type doesn't match
3814 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3816 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3819 FieldSeqNode* fldSeq,
3823 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3824 // has been parsed to yield the other input arguments. If evaluation of the address
3825 // can raise exceptions, those should be captured in the exception set "excVN."
3826 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3827 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3828 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3829 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3830 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3832 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3833 CORINFO_CLASS_HANDLE elemTypeEq,
3837 FieldSeqNode* fldSeq);
3839 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3840 // by evaluating the array index expression "tree". Returns the value number resulting from
3841 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3842 // "GT_IND" that does the dereference, and it is given the returned value number.
3843 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3845 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3846 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3848 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3850 // Utility functions for fgValueNumber.
3852 // Perform value-numbering for the trees in "blk".
3853 void fgValueNumberBlock(BasicBlock* blk);
3855 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3856 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3857 // assumed for the memoryKind at the start "entryBlk".
3858 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3860 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3861 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3862 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3864 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3866 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3868 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3869 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3870 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3872 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3873 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3875 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3876 // value in that SSA #.
3877 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3879 // The input 'tree' is a leaf node that is a constant
3880 // Assign the proper value number to the tree
3881 void fgValueNumberTreeConst(GenTreePtr tree);
3883 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3884 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3886 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3888 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3890 // Does value-numbering for a block assignment.
3891 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3893 // Does value-numbering for a cast tree.
3894 void fgValueNumberCastTree(GenTreePtr tree);
3896 // Does value-numbering for an intrinsic tree.
3897 void fgValueNumberIntrinsic(GenTreePtr tree);
3899 // Does value-numbering for a call. We interpret some helper calls.
3900 void fgValueNumberCall(GenTreeCall* call);
3902 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3903 void fgUpdateArgListVNs(GenTreeArgList* args);
3905 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3906 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3908 // Requires "helpCall" to be a helper call. Assigns it a value number;
3909 // we understand the semantics of some of the calls. Returns "true" if
3910 // the call may modify the heap (we assume arbitrary memory side effects if so).
3911 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3913 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3914 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3916 // These are the current value number for the memory implicit variables while
3917 // doing value numbering. These are the value numbers under the "liberal" interpretation
3918 // of memory values; the "conservative" interpretation needs no VN, since every access of
3919 // memory yields an unknown value.
3920 ValueNum fgCurMemoryVN[MemoryKindCount];
3922 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3923 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3924 // is 1, and the rest is an encoding of "elemTyp".
3925 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3927 if (elemStructType != nullptr)
3929 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3930 varTypeIsIntegral(elemTyp));
3931 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3932 return elemStructType;
3936 elemTyp = varTypeUnsignedToSigned(elemTyp);
3937 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3940 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3941 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3942 // the struct type of the element).
3943 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3945 size_t clsHndVal = size_t(clsHnd);
3946 if (clsHndVal & 0x1)
3948 return var_types(clsHndVal >> 1);
3956 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3957 var_types getJitGCType(BYTE gcType);
3959 enum structPassingKind
3961 SPK_Unknown, // Invalid value, never returned
3962 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3963 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3964 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3965 // parameters registers are used, then the stack will be used)
3966 // for X86 passed on the stack, for ARM32 passed in registers
3967 // or the stack or split between registers and the stack.
3968 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3970 }; // The struct is passed/returned by reference to a copy/buffer.
3972 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3973 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3974 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3975 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3977 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3979 // Get the type that is used to pass values of the given struct type.
3980 // If you have already retrieved the struct size then pass it as the optional third argument
3982 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3983 structPassingKind* wbPassStruct,
3984 unsigned structSize = 0);
3986 // Get the type that is used to return values of the given struct type.
3987 // If you have already retrieved the struct size then pass it as the optional third argument
3989 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3990 structPassingKind* wbPassStruct = nullptr,
3991 unsigned structSize = 0);
3994 // Print a representation of "vnp" or "vn" on standard output.
3995 // If "level" is non-zero, we also print out a partial expansion of the value.
3996 void vnpPrint(ValueNumPair vnp, unsigned level);
3997 void vnPrint(ValueNum vn, unsigned level);
4000 // Dominator computation member functions
4001 // Not exposed outside Compiler
4003 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4005 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4007 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4008 // flow graph. We first assume the fields bbIDom on each
4009 // basic block are invalid. This computation is needed later
4010 // by fgBuildDomTree to build the dominance tree structure.
4011 // Based on: A Simple, Fast Dominance Algorithm
4012 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4014 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4015 // Note: this is relatively slow compared to calling fgDominate(),
4016 // especially if dealing with a single block versus block check.
4018 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4020 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4022 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4024 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4026 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4028 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4029 // processed in topological sort, this function takes care of that.
4031 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4033 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4034 // Returns this as a set.
4036 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4037 // root nodes. Returns this as a set.
4040 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4043 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4044 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4047 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4048 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4049 // && postOrder(A) >= postOrder(B) making the computation O(1).
4050 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4052 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4054 void fgUpdateChangedFlowGraph();
4057 // Compute the predecessors of the blocks in the control flow graph.
4058 void fgComputePreds();
4060 // Remove all predecessor information.
4061 void fgRemovePreds();
4063 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4064 // before the full predecessors lists are computed.
4065 void fgComputeCheapPreds();
4068 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4070 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4080 // Initialize the per-block variable sets (used for liveness analysis).
4081 void fgInitBlockVarSets();
4083 // true if we've gone through and created GC Poll calls.
4084 bool fgGCPollsCreated;
4085 void fgMarkGCPollBlocks();
4086 void fgCreateGCPolls();
4087 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4089 // Requires that "block" is a block that returns from
4090 // a finally. Returns the number of successors (jump targets of
4091 // of blocks in the covered "try" that did a "LEAVE".)
4092 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4094 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4095 // a finally. Returns its "i"th successor (jump targets of
4096 // of blocks in the covered "try" that did a "LEAVE".)
4097 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4098 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4101 // Factor out common portions of the impls of the methods above.
4102 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4105 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4106 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4107 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4108 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4109 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4110 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4111 // we leave the entry associated with the block, but it will no longer be accessed.)
4112 struct SwitchUniqueSuccSet
4114 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4115 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4118 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4119 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4120 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4121 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4124 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4125 BlockToSwitchDescMap;
4128 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4129 // iteration over only the distinct successors.
4130 BlockToSwitchDescMap* m_switchDescMap;
4133 BlockToSwitchDescMap* GetSwitchDescMap()
4135 if (m_switchDescMap == nullptr)
4137 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4139 return m_switchDescMap;
4142 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4143 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4144 // we don't accidentally look up and return the wrong switch data.
4145 void InvalidateUniqueSwitchSuccMap()
4147 m_switchDescMap = nullptr;
4150 // Requires "switchBlock" to be a block that ends in a switch. Returns
4151 // the corresponding SwitchUniqueSuccSet.
4152 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4154 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4155 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4156 // remove it from "this", and ensure that "to" is a member.
4157 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4159 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4160 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4162 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4164 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4166 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4168 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4170 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4172 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4174 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4176 void fgRemoveBlockAsPred(BasicBlock* block);
4178 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4180 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4182 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4184 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4186 flowList* fgAddRefPred(BasicBlock* block,
4187 BasicBlock* blockPred,
4188 flowList* oldEdge = nullptr,
4189 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4192 void fgFindBasicBlocks();
4194 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4196 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4198 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4199 bool putInTryRegion,
4200 BasicBlock* startBlk,
4202 BasicBlock* nearBlk,
4203 BasicBlock* jumpBlk,
4206 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4208 void fgRemoveEmptyBlocks();
4210 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4212 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4214 void fgCreateLoopPreHeader(unsigned lnum);
4216 void fgUnreachableBlock(BasicBlock* block);
4218 void fgRemoveConditionalJump(BasicBlock* block);
4220 BasicBlock* fgLastBBInMainFunction();
4222 BasicBlock* fgEndBBAfterMainFunction();
4224 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4226 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4228 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4230 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4232 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4234 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4236 bool fgRenumberBlocks();
4238 bool fgExpandRarelyRunBlocks();
4240 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4242 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4244 enum FG_RELOCATE_TYPE
4246 FG_RELOCATE_TRY, // relocate the 'try' region
4247 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4249 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4251 #if FEATURE_EH_FUNCLETS
4252 #if defined(_TARGET_ARM_)
4253 void fgClearFinallyTargetBit(BasicBlock* block);
4254 #endif // defined(_TARGET_ARM_)
4255 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4256 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4257 void fgInsertFuncletPrologBlock(BasicBlock* block);
4258 void fgCreateFuncletPrologBlocks();
4259 void fgCreateFunclets();
4260 #else // !FEATURE_EH_FUNCLETS
4261 bool fgRelocateEHRegions();
4262 #endif // !FEATURE_EH_FUNCLETS
4264 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4266 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4268 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4270 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4272 bool fgOptimizeEmptyBlock(BasicBlock* block);
4274 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4276 bool fgOptimizeBranch(BasicBlock* bJump);
4278 bool fgOptimizeSwitchBranches(BasicBlock* block);
4280 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4282 bool fgOptimizeSwitchJumps();
4284 void fgPrintEdgeWeights();
4286 void fgComputeEdgeWeights();
4288 void fgReorderBlocks();
4290 void fgDetermineFirstColdBlock();
4292 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4294 bool fgUpdateFlowGraph(bool doTailDup = false);
4296 void fgFindOperOrder();
4298 // method that returns if you should split here
4299 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4301 void fgSetBlockOrder();
4303 void fgRemoveReturnBlock(BasicBlock* block);
4305 /* Helper code that has been factored out */
4306 inline void fgConvertBBToThrowBB(BasicBlock* block);
4308 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4309 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4310 GenTreePtr fgMakeTmpArgNode(
4311 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4313 // The following check for loops that don't execute calls
4314 bool fgLoopCallMarked;
4316 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4317 void fgLoopCallMark();
4319 void fgMarkLoopHead(BasicBlock* block);
4321 unsigned fgGetCodeEstimate(BasicBlock* block);
4324 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4325 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4326 bool fgDumpFlowGraph(Phases phase);
4328 #endif // DUMP_FLOWGRAPHS
4333 void fgDispBBLiveness(BasicBlock* block);
4334 void fgDispBBLiveness();
4335 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4336 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4337 void fgDispBasicBlocks(bool dumpTrees = false);
4338 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4339 void fgDumpBlock(BasicBlock* block);
4340 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4342 static fgWalkPreFn fgStress64RsltMulCB;
4343 void fgStress64RsltMul();
4344 void fgDebugCheckUpdate();
4345 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4346 void fgDebugCheckBlockLinks();
4347 void fgDebugCheckLinks(bool morphTrees = false);
4348 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4349 void fgDebugCheckFlags(GenTreePtr tree);
4350 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4351 void fgDebugCheckTryFinallyExits();
4354 #ifdef LEGACY_BACKEND
4355 static void fgOrderBlockOps(GenTreePtr tree,
4359 GenTreePtr* opsPtr, // OUT
4360 regMaskTP* regsPtr); // OUT
4361 #endif // LEGACY_BACKEND
4363 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4364 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4366 inline bool fgIsInlining()
4368 return fgExpandInline;
4371 void fgTraverseRPO();
4373 //--------------------- Walking the trees in the IR -----------------------
4378 fgWalkPreFn* wtprVisitorFn;
4379 fgWalkPostFn* wtpoVisitorFn;
4380 void* pCallbackData; // user-provided data
4381 bool wtprLclsOnly; // whether to only visit lclvar nodes
4382 GenTreePtr parent; // parent of current node, provided to callback
4383 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4385 bool printModified; // callback can use this
4389 template <bool computeStack>
4390 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4392 // general purpose tree-walker that is capable of doing pre- and post- order
4393 // callbacks at the same time
4394 template <bool doPreOrder, bool doPostOrder>
4395 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4397 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4398 fgWalkPreFn* visitor,
4399 void* pCallBackData = nullptr,
4400 bool lclVarsOnly = false,
4401 bool computeStack = false);
4403 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4404 fgWalkPreFn* preVisitor,
4405 fgWalkPostFn* postVisitor,
4406 void* pCallBackData = nullptr);
4408 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4412 template <bool computeStack>
4413 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4415 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4416 fgWalkPostFn* visitor,
4417 void* pCallBackData = nullptr,
4418 bool computeStack = false);
4420 // An fgWalkPreFn that looks for expressions that have inline throws in
4421 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4422 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4423 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4424 // properly propagated to parent trees). It returns WALK_CONTINUE
4426 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4427 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4428 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4430 /**************************************************************************
4432 *************************************************************************/
4435 friend class SsaBuilder;
4436 friend struct ValueNumberState;
4438 //--------------------- Detect the basic blocks ---------------------------
4440 BasicBlock** fgBBs; // Table of pointers to the BBs
4442 void fgInitBBLookup();
4443 BasicBlock* fgLookupBB(unsigned addr);
4445 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4447 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4449 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4451 void fgLinkBasicBlocks();
4453 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4455 void fgCheckBasicBlockControlFlow();
4457 void fgControlFlowPermitted(BasicBlock* blkSrc,
4458 BasicBlock* blkDest,
4459 BOOL IsLeave = false /* is the src a leave block */);
4461 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4463 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4465 void fgAdjustForAddressExposedOrWrittenThis();
4467 bool fgProfileData_ILSizeMismatch;
4468 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4469 ULONG fgProfileBufferCount;
4470 ULONG fgNumProfileRuns;
4472 unsigned fgStressBBProf()
4475 unsigned result = JitConfig.JitStressBBProf();
4478 if (compStressCompile(STRESS_BB_PROFILE, 15))
4489 bool fgHaveProfileData();
4490 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4492 bool fgIsUsingProfileWeights()
4494 return (fgHaveProfileData() || fgStressBBProf());
4496 void fgInstrumentMethod();
4498 //-------- Insert a statement at the start or end of a basic block --------
4502 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4506 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4508 public: // Used by linear scan register allocation
4509 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4512 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4513 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4515 public: // Used by linear scan register allocation
4516 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4519 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4521 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4523 // Create a new temporary variable to hold the result of *ppTree,
4524 // and transform the graph accordingly.
4525 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4526 GenTree* fgMakeMultiUse(GenTree** ppTree);
4529 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4530 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4531 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4533 //-------- Determine the order in which the trees will be evaluated -------
4535 unsigned fgTreeSeqNum;
4536 GenTree* fgTreeSeqLst;
4537 GenTree* fgTreeSeqBeg;
4539 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4540 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4541 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4542 void fgSetStmtSeq(GenTree* tree);
4543 void fgSetBlockOrder(BasicBlock* block);
4545 //------------------------- Morphing --------------------------------------
4547 unsigned fgPtrArgCntCur;
4548 unsigned fgPtrArgCntMax;
4549 hashBv* fgOutgoingArgTemps;
4550 hashBv* fgCurrentlyInUseArgTemps;
4552 bool compCanEncodePtrArgCntMax();
4554 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4557 void fgMoveOpsLeft(GenTreePtr tree);
4560 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4562 bool fgIsThrow(GenTreePtr tree);
4564 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4565 bool fgIsBlockCold(BasicBlock* block);
4567 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4569 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4571 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4573 bool fgMorphRelopToQmark(GenTreePtr tree);
4575 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4576 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4577 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4578 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4579 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4580 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4581 // small; hence the other fields of MorphAddrContext.
4582 enum MorphAddrContextKind
4587 struct MorphAddrContext
4589 MorphAddrContextKind m_kind;
4590 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4591 // top-level indirection and here have been constants.
4592 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4593 // In that case, is the sum of those constant offsets.
4595 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4600 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4601 static MorphAddrContext s_CopyBlockMAC;
4604 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4605 var_types* baseTypeOut,
4607 unsigned* simdSizeOut,
4608 bool ignoreUsedInSIMDIntrinsic = false);
4609 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4610 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4611 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4612 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4614 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4615 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4616 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4618 #endif // FEATURE_SIMD
4619 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4620 GenTreePtr fgMorphCast(GenTreePtr tree);
4621 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4622 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4624 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4627 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4628 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4630 void fgFixupStructReturn(GenTreePtr call);
4631 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4632 bool fgAddrCouldBeNull(GenTreePtr addr);
4633 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4634 bool fgCanFastTailCall(GenTreeCall* call);
4635 void fgMorphTailCall(GenTreeCall* call);
4636 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4637 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4638 fgArgTabEntryPtr argTabEntry,
4640 IL_OFFSETX callILOffset,
4641 GenTreePtr tmpAssignmentInsertionPoint,
4642 GenTreePtr paramAssignmentInsertionPoint);
4643 static int fgEstimateCallStackSize(GenTreeCall* call);
4644 GenTreePtr fgMorphCall(GenTreeCall* call);
4645 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4646 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4648 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4649 static fgWalkPreFn fgFindNonInlineCandidate;
4651 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4652 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4653 void fgAssignSetVarDef(GenTreePtr tree);
4654 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4655 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4656 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4657 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4658 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4659 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4660 void fgMorphUnsafeBlk(GenTreeObj* obj);
4661 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4662 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4663 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4664 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4665 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4666 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4667 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4669 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4670 GenTreePtr fgMorphConst(GenTreePtr tree);
4673 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4676 #if LOCAL_ASSERTION_PROP
4677 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4678 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4680 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4682 GenTreeStmt* fgMorphStmt;
4684 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4685 // used when morphing big offset.
4687 //----------------------- Liveness analysis -------------------------------
4689 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4690 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4692 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4693 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4694 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4696 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4698 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4700 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4701 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4703 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4704 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4706 void fgExtendDbgScopes();
4707 void fgExtendDbgLifetimes();
4710 void fgDispDebugScopes();
4713 //-------------------------------------------------------------------------
4715 // The following keeps track of any code we've added for things like array
4716 // range checking or explicit calls to enable GC, and so on.
4721 AddCodeDsc* acdNext;
4722 BasicBlock* acdDstBlk; // block to which we jump
4724 SpecialCodeKind acdKind; // what kind of a special block is this?
4725 unsigned short acdStkLvl;
4729 static unsigned acdHelper(SpecialCodeKind codeKind);
4731 AddCodeDsc* fgAddCodeList;
4733 bool fgRngChkThrowAdded;
4734 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4736 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4738 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4741 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4744 bool fgIsCodeAdded();
4746 bool fgIsThrowHlpBlk(BasicBlock* block);
4747 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4749 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4751 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4752 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4753 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4754 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4755 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4757 #if FEATURE_MULTIREG_RET
4758 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4759 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4760 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4761 #endif // FEATURE_MULTIREG_RET
4763 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4766 static fgWalkPreFn fgDebugCheckInlineCandidates;
4768 void CheckNoFatPointerCandidatesLeft();
4769 static fgWalkPreFn fgDebugCheckFatPointerCandidates;
4772 void fgPromoteStructs();
4773 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4774 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4775 void fgMarkImplicitByRefArgs();
4776 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4777 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4778 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4779 void fgMarkAddressExposedLocals();
4780 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4782 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4784 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4786 // The given local variable, required to be a struct variable, is being assigned via
4787 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4788 // the variable is not enregistered, and is therefore not promoted independently.
4789 void fgLclFldAssign(unsigned lclNum);
4791 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4792 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4793 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4794 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4797 bool fgPrintInlinedMethods;
4800 bool fgIsBigOffset(size_t offset);
4802 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4803 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4804 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4805 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4806 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4809 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4810 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4814 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4815 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4822 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4825 void optRemoveRangeCheck(
4826 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4827 bool optIsRangeCheckRemovable(GenTreePtr tree);
4830 static fgWalkPreFn optValidRangeCheckIndex;
4831 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4834 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4836 /**************************************************************************
4838 *************************************************************************/
4841 // Do hoisting for all loops.
4842 void optHoistLoopCode();
4844 // To represent sets of VN's that have already been hoisted in outer loops.
4845 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4846 typedef VNToBoolMap VNSet;
4848 struct LoopHoistContext
4851 // The set of variables hoisted in the current loop (or nullptr if there are none).
4852 VNSet* m_pHoistedInCurLoop;
4855 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4856 VNSet m_hoistedInParentLoops;
4857 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4858 // Previous decisions on loop-invariance of value numbers in the current loop.
4859 VNToBoolMap m_curLoopVnInvariantCache;
4861 VNSet* GetHoistedInCurLoop(Compiler* comp)
4863 if (m_pHoistedInCurLoop == nullptr)
4865 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4867 return m_pHoistedInCurLoop;
4870 VNSet* ExtractHoistedInCurLoop()
4872 VNSet* res = m_pHoistedInCurLoop;
4873 m_pHoistedInCurLoop = nullptr;
4877 LoopHoistContext(Compiler* comp)
4878 : m_pHoistedInCurLoop(nullptr)
4879 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4880 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4885 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4886 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4887 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4888 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4890 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4891 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4892 // "m_hoistedInParentLoops".
4894 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4896 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4897 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4898 // expressions to "hoistInLoop".
4899 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4901 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4902 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4904 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4905 // that are invariant in loop "lnum" (an index into the optLoopTable)
4906 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4907 // expressions to "hoistInLoop".
4908 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4909 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4910 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4911 bool optHoistLoopExprsForTree(GenTreePtr tree,
4913 LoopHoistContext* hoistCtxt,
4914 bool* firstBlockAndBeforeSideEffect,
4917 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4918 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4920 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4921 // Constants and init values are always loop invariant.
4922 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4923 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4925 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4926 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4927 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4928 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4929 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4931 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4932 // in the loop table.
4933 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4935 // Records the set of "side effects" of all loops: fields (object instance and static)
4936 // written to, and SZ-array element type equivalence classes updated.
4937 void optComputeLoopSideEffects();
4940 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4941 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4942 // static) written to, and SZ-array element type equivalence classes updated.
4943 void optComputeLoopNestSideEffects(unsigned lnum);
4945 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4946 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4948 // Hoist the expression "expr" out of loop "lnum".
4949 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4952 void optOptimizeBools();
4955 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4957 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4960 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4962 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4963 // the loop into a "do-while" loop
4964 // Also finds all natural loops and records them in the loop table
4966 // Optionally clone loops in the loop table.
4967 void optCloneLoops();
4969 // Clone loop "loopInd" in the loop table.
4970 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4972 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4973 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4974 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4976 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4978 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4981 // This enumeration describes what is killed by a call.
4985 CALLINT_NONE, // no interference (most helpers)
4986 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4987 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4988 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4989 CALLINT_ALL, // kills everything (normal method call)
4993 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4994 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4995 // in bbNext order; we use comparisons on the bbNum to decide order.)
4996 // The blocks that define the body are
4997 // first <= top <= entry <= bottom .
4998 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4999 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5000 // Compiler::optFindNaturalLoops().
5003 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5004 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5005 // loop, but not the outer loop.)
5006 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5008 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5009 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5010 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5012 callInterf lpAsgCall; // "callInterf" for calls in the loop
5013 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5014 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5016 unsigned short lpFlags; // Mask of the LPFLG_* constants
5018 unsigned char lpExitCnt; // number of exits from the loop
5020 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5021 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5022 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5023 // (Actually, an "immediately" nested loop --
5024 // no other child of this loop is a parent of lpChild.)
5025 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5026 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5027 // by following "lpChild" then "lpSibling" links.
5029 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5030 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5032 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5033 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5034 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5036 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5037 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5039 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5040 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5041 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5042 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5044 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5045 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5046 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5048 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5049 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5050 // type are assigned to.
5052 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5053 // memory side effects. If this is set, the fields below
5054 // may not be accurate (since they become irrelevant.)
5055 bool lpContainsCall; // True if executing the loop body *may* execute a call
5057 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5058 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5060 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5062 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5063 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5065 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5067 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5068 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5070 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5071 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5073 JitSimplerHashBehavior>
5075 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5076 // instance fields modified
5079 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5080 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5082 JitSimplerHashBehavior>
5084 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5085 // arrays of that type are modified
5088 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5089 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5091 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5092 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5093 // (shifted left, with a low-order bit set to distinguish.)
5094 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5095 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5097 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5099 GenTreePtr lpIterTree; // The "i <op>= const" tree
5100 unsigned lpIterVar(); // iterator variable #
5101 int lpIterConst(); // the constant with which the iterator is incremented
5102 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5103 void VERIFY_lpIterTree();
5105 var_types lpIterOperType(); // For overflow instructions
5108 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5109 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5113 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5115 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5116 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5117 void VERIFY_lpTestTree();
5119 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5120 GenTreePtr lpIterator(); // the iterator node in the loop test
5121 GenTreePtr lpLimit(); // the limit node in the loop test
5123 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5124 // LPFLG_CONST_LIMIT
5125 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5127 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5128 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5129 // LPFLG_ARRLEN_LIMIT
5131 // Returns "true" iff "*this" contains the blk.
5132 bool lpContains(BasicBlock* blk)
5134 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5136 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5137 // to be equal, but requiring bottoms to be different.)
5138 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5140 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5143 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5144 // bottoms to be different.)
5145 bool lpContains(const LoopDsc& lp2)
5147 return lpContains(lp2.lpFirst, lp2.lpBottom);
5150 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5151 // (allowing firsts to be equal, but requiring bottoms to be different.)
5152 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5154 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5157 // Returns "true" iff "*this" is (properly) contained by "lp2"
5158 // (allowing firsts to be equal, but requiring bottoms to be different.)
5159 bool lpContainedBy(const LoopDsc& lp2)
5161 return lpContains(lp2.lpFirst, lp2.lpBottom);
5164 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5165 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5167 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5169 // Returns "true" iff "*this" is disjoint from "lp2".
5170 bool lpDisjoint(const LoopDsc& lp2)
5172 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5174 // Returns "true" iff the loop is well-formed (see code for defn).
5177 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5178 lpEntry->bbNum <= lpBottom->bbNum &&
5179 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5184 bool fgMightHaveLoop(); // returns true if there are any backedges
5185 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5188 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5189 unsigned char optLoopCount; // number of tracked loops
5192 unsigned optCallCount; // number of calls made in the method
5193 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5194 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5195 unsigned optLoopsCloned; // number of loops cloned in the current method.
5198 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5199 void optPrintLoopInfo(unsigned loopNum,
5201 BasicBlock* lpFirst,
5203 BasicBlock* lpEntry,
5204 BasicBlock* lpBottom,
5205 unsigned char lpExitCnt,
5207 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5208 void optPrintLoopInfo(unsigned lnum);
5209 void optPrintLoopRecording(unsigned lnum);
5211 void optCheckPreds();
5214 void optSetBlockWeights();
5216 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5218 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5220 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5222 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5223 unsigned optIsLoopIncrTree(GenTreePtr incr);
5224 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5225 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5226 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5227 bool optExtractInitTestIncr(BasicBlock* head,
5232 GenTreePtr* ppIncr);
5234 void optRecordLoop(BasicBlock* head,
5240 unsigned char exitCnt);
5242 void optFindNaturalLoops();
5244 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5245 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5246 bool optCanonicalizeLoopNest(unsigned char loopInd);
5248 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5249 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5250 bool optCanonicalizeLoop(unsigned char loopInd);
5252 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5253 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5254 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5255 bool optLoopContains(unsigned l1, unsigned l2);
5257 // Requires "loopInd" to be a valid index into the loop table.
5258 // Updates the loop table by changing loop "loopInd", whose head is required
5259 // to be "from", to be "to". Also performs this transformation for any
5260 // loop nested in "loopInd" that shares the same head as "loopInd".
5261 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5263 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5264 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5265 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5267 // Marks the containsCall information to "lnum" and any parent loops.
5268 void AddContainsCallAllContainingLoops(unsigned lnum);
5269 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5270 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5271 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5272 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5273 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5274 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5276 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5277 // of "from".) Copies the jump destination from "from" to "to".
5278 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5280 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5281 unsigned optLoopDepth(unsigned lnum)
5283 unsigned par = optLoopTable[lnum].lpParent;
5284 if (par == BasicBlock::NOT_IN_LOOP)
5290 return 1 + optLoopDepth(par);
5294 void fgOptWhileLoop(BasicBlock* block);
5296 bool optComputeLoopRep(int constInit,
5299 genTreeOps iterOper,
5301 genTreeOps testOper,
5304 unsigned* iterCount);
5305 #if FEATURE_STACK_FP_X87
5308 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5309 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5310 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5311 #endif // FEATURE_STACK_FP_X87
5314 static fgWalkPreFn optIsVarAssgCB;
5317 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5319 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5321 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5323 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5325 /**************************************************************************
5326 * Optimization conditions
5327 *************************************************************************/
5329 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5330 bool optPentium4(void);
5331 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5332 bool optAvoidIntMult(void);
5337 // The following is the upper limit on how many expressions we'll keep track
5338 // of for the CSE analysis.
5340 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5342 static const int MIN_CSE_COST = 2;
5344 // Keeps tracked cse indices
5345 BitVecTraits* cseTraits;
5349 /* Generic list of nodes - used by the CSE logic */
5357 typedef struct treeLst* treeLstPtr;
5361 treeStmtLst* tslNext;
5362 GenTreePtr tslTree; // tree node
5363 GenTreePtr tslStmt; // statement containing the tree
5364 BasicBlock* tslBlock; // block containing the statement
5367 typedef struct treeStmtLst* treeStmtLstPtr;
5369 // The following logic keeps track of expressions via a simple hash table.
5373 CSEdsc* csdNextInBucket; // used by the hash table
5375 unsigned csdHashValue; // the orginal hashkey
5377 unsigned csdIndex; // 1..optCSECandidateCount
5378 char csdLiveAcrossCall; // 0 or 1
5380 unsigned short csdDefCount; // definition count
5381 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5383 unsigned csdDefWtCnt; // weighted def count
5384 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5386 GenTreePtr csdTree; // treenode containing the 1st occurance
5387 GenTreePtr csdStmt; // stmt containing the 1st occurance
5388 BasicBlock* csdBlock; // block containing the 1st occurance
5390 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5391 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5393 ValueNum defConservativeVN; // if all def occurrences share the same conservative value
5394 // number, this will reflect it; otherwise, NoVN.
5397 static const size_t s_optCSEhashSize;
5398 CSEdsc** optCSEhash;
5403 CSEdsc* optCSEfindDsc(unsigned index);
5404 void optUnmarkCSE(GenTreePtr tree);
5406 // user defined callback data for the tree walk function optCSE_MaskHelper()
5407 struct optCSE_MaskData
5409 EXPSET_TP CSE_defMask;
5410 EXPSET_TP CSE_useMask;
5413 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5414 static fgWalkPreFn optCSE_MaskHelper;
5416 // This function walks all the node for an given tree
5417 // and return the mask of CSE definitions and uses for the tree
5419 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5421 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5422 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5423 bool optCSE_canSwap(GenTree* tree);
5425 static fgWalkPostFn optPropagateNonCSE;
5426 static fgWalkPreFn optHasNonCSEChild;
5428 static fgWalkPreFn optUnmarkCSEs;
5430 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5431 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5433 void optCleanupCSEs();
5436 void optEnsureClearCSEInfo();
5439 #endif // FEATURE_ANYCSE
5441 #if FEATURE_VALNUM_CSE
5442 /**************************************************************************
5443 * Value Number based CSEs
5444 *************************************************************************/
5447 void optOptimizeValnumCSEs();
5450 void optValnumCSE_Init();
5451 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5452 unsigned optValnumCSE_Locate();
5453 void optValnumCSE_InitDataFlow();
5454 void optValnumCSE_DataFlow();
5455 void optValnumCSE_Availablity();
5456 void optValnumCSE_Heuristic();
5457 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5459 #endif // FEATURE_VALNUM_CSE
5462 bool optDoCSE; // True when we have found a duplicate CSE tree
5463 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5464 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5465 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5466 unsigned optCSEstart; // The first local variable number that is a CSE
5467 unsigned optCSEcount; // The total count of CSE's introduced.
5468 unsigned optCSEweight; // The weight of the current block when we are
5469 // scanning for CSE expressions
5471 bool optIsCSEcandidate(GenTreePtr tree);
5473 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5475 bool lclNumIsTrueCSE(unsigned lclNum) const
5477 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5480 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5482 bool lclNumIsCSE(unsigned lclNum) const
5484 return lvaTable[lclNum].lvIsCSE;
5488 bool optConfigDisableCSE();
5489 bool optConfigDisableCSE2();
5491 void optOptimizeCSEs();
5493 #endif // FEATURE_ANYCSE
5501 unsigned ivaVar; // Variable we are interested in, or -1
5502 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5503 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5504 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5505 callInterf ivaMaskCall; // What kind of calls are there?
5508 static callInterf optCallInterf(GenTreePtr call);
5511 // VN based copy propagation.
5512 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5513 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5514 LclNumToGenTreePtrStack;
5516 // Kill set to track variables with intervening definitions.
5517 VARSET_TP optCopyPropKillSet;
5519 // Copy propagation functions.
5520 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5521 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5522 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5523 bool optIsSsaLocal(GenTreePtr tree);
5524 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5525 void optVnCopyProp();
5527 /**************************************************************************
5528 * Early value propagation
5529 *************************************************************************/
5535 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5539 static unsigned GetHashCode(SSAName ssaNm)
5541 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5544 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5546 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5550 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5551 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5552 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5553 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5554 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5555 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
5557 bool doesMethodHaveFatPointer()
5559 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
5562 void setMethodHasFatPointer()
5564 optMethodFlags |= OMF_HAS_FATPOINTER;
5567 void clearMethodHasFatPointer()
5569 optMethodFlags &= ~OMF_HAS_FATPOINTER;
5572 unsigned optMethodFlags;
5574 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5575 // No throughput diff was found with backward walk bound between 3-8.
5576 static const int optEarlyPropRecurBound = 5;
5578 enum class optPropKind
5586 bool gtIsVtableRef(GenTreePtr tree);
5587 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5588 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5589 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5590 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5591 bool optEarlyPropRewriteTree(GenTreePtr tree);
5592 bool optDoEarlyPropForBlock(BasicBlock* block);
5593 bool optDoEarlyPropForFunc();
5594 void optEarlyProp();
5595 void optFoldNullCheck(GenTreePtr tree);
5596 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5599 /**************************************************************************
5600 * Value/Assertion propagation
5601 *************************************************************************/
5603 // Data structures for assertion prop
5604 BitVecTraits* apTraits;
5608 enum optAssertionKind
5623 O1K_ARRLEN_OPER_BND,
5624 O1K_ARRLEN_LOOP_BND,
5625 O1K_CONSTANT_LOOP_BND,
5646 optAssertionKind assertionKind;
5649 unsigned lclNum; // assigned to or property of this local var number
5657 struct AssertionDscOp1
5659 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5666 struct AssertionDscOp2
5668 optOp2Kind kind; // a const or copy assignment
5672 ssize_t iconVal; // integer
5673 unsigned iconFlags; // gtFlags
5675 struct Range // integer subrange
5689 bool IsArrLenArithBound()
5691 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5693 bool IsArrLenBound()
5695 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5697 bool IsConstantBound()
5699 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5700 op1.kind == O1K_CONSTANT_LOOP_BND);
5702 bool IsBoundsCheckNoThrow()
5704 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5707 bool IsCopyAssertion()
5709 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5712 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5714 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5715 a1->op2.kind == a2->op2.kind;
5718 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5720 if (kind == OAK_EQUAL)
5722 return kind2 == OAK_NOT_EQUAL;
5724 else if (kind == OAK_NOT_EQUAL)
5726 return kind2 == OAK_EQUAL;
5731 static ssize_t GetLowerBoundForIntegralType(var_types type)
5751 static ssize_t GetUpperBoundForIntegralType(var_types type)
5775 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5777 return (op1.kind == that->op1.kind) &&
5778 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5781 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5783 if (op2.kind != that->op2.kind)
5789 case O2K_IND_CNS_INT:
5791 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5793 case O2K_CONST_LONG:
5794 return (op2.lconVal == that->op2.lconVal);
5796 case O2K_CONST_DOUBLE:
5797 // exact match because of positive and negative zero.
5798 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5800 case O2K_LCLVAR_COPY:
5802 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5803 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5806 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5809 // we will return false
5813 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5819 bool Complementary(AssertionDsc* that, bool vnBased)
5821 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5822 HasSameOp2(that, vnBased);
5825 bool Equals(AssertionDsc* that, bool vnBased)
5827 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5831 typedef unsigned short AssertionIndex;
5834 static fgWalkPreFn optAddCopiesCallback;
5835 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5836 unsigned optAddCopyLclNum;
5837 GenTreePtr optAddCopyAsgnNode;
5839 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5840 bool optAssertionPropagated; // set to true if we modified the trees
5841 bool optAssertionPropagatedCurrentStmt;
5843 GenTreePtr optAssertionPropCurrentTree;
5845 AssertionIndex* optComplementaryAssertionMap;
5846 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5847 // using the value of a local var) for each local var
5848 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5849 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5850 AssertionIndex optMaxAssertionCount;
5853 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5854 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5855 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5856 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5857 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5858 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5860 AssertionIndex GetAssertionCount()
5862 return optAssertionCount;
5864 ASSERT_TP* bbJtrueAssertionOut;
5865 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5866 ValueNumToAssertsMap;
5867 ValueNumToAssertsMap* optValueNumToAsserts;
5869 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5871 // Assertion prop helpers.
5872 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5873 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5874 void optAssertionInit(bool isLocalProp);
5875 void optAssertionTraitsInit(AssertionIndex assertionCount);
5876 #if LOCAL_ASSERTION_PROP
5877 void optAssertionReset(AssertionIndex limit);
5878 void optAssertionRemove(AssertionIndex index);
5881 // Assertion prop data flow functions.
5882 void optAssertionPropMain();
5883 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5884 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5885 ASSERT_TP* optInitAssertionDataflowFlags();
5886 ASSERT_TP* optComputeAssertionGen();
5888 // Assertion Gen functions.
5889 void optAssertionGen(GenTreePtr tree);
5890 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5891 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5892 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5893 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5894 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5895 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5897 // Assertion creation functions.
5898 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5899 AssertionIndex optCreateAssertion(GenTreePtr op1,
5901 optAssertionKind assertionKind,
5902 AssertionDsc* assertion);
5903 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5905 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5906 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5907 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5909 void optPrintVnAssertionMapping();
5911 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5913 // Used for respective assertion propagations.
5914 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5915 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5916 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5917 bool optAssertionIsNonNull(GenTreePtr op,
5918 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5920 // Used for Relop propagation.
5921 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5922 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5923 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5925 // Assertion prop for lcl var functions.
5926 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5927 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5929 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5930 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5931 const GenTreePtr tree,
5932 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5933 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5935 // Assertion propagation functions.
5936 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5937 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5938 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5939 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5940 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5941 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5942 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5943 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5944 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5945 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5946 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5947 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5949 // Implied assertion functions.
5950 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5951 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5952 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5953 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5955 ASSERT_VALRET_TP optNewFullAssertSet();
5956 ASSERT_VALRET_TP optNewEmptyAssertSet();
5959 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5960 void optDebugCheckAssertion(AssertionDsc* assertion);
5961 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5963 void optAddCopies();
5964 #endif // ASSERTION_PROP
5966 /**************************************************************************
5968 *************************************************************************/
5971 struct LoopCloneVisitorInfo
5973 LoopCloneContext* context;
5976 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5977 : context(context), loopNum(loopNum), stmt(nullptr)
5982 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5983 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5984 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5985 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5986 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5987 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5988 void optObtainLoopCloningOpts(LoopCloneContext* context);
5989 bool optIsLoopClonable(unsigned loopInd);
5991 bool optCanCloneLoops();
5994 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5996 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5997 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5998 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5999 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6003 void optInsertLoopCloningStress(BasicBlock* head);
6005 #if COUNT_RANGECHECKS
6006 static unsigned optRangeChkRmv;
6007 static unsigned optRangeChkAll;
6016 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
6021 RngChkDsc* rcdNextInBucket; // used by the hash table
6023 unsigned short rcdHashValue; // to make matching faster
6024 unsigned short rcdIndex; // 0..optRngChkCount-1
6026 GenTreePtr rcdTree; // the array index tree
6029 unsigned optRngChkCount;
6030 static const size_t optRngChkHashSize;
6032 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6033 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6035 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6038 bool optLoopsMarked;
6041 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6042 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6046 XX Does the register allocation and puts the remaining lclVars on the stack XX
6048 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6049 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6053 #ifndef LEGACY_BACKEND
6058 #else // LEGACY_BACKEND
6063 #endif // LEGACY_BACKEND
6065 #ifdef LEGACY_BACKEND
6067 void raAssignVars(); // register allocation
6068 #endif // LEGACY_BACKEND
6070 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6072 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6074 void raMarkStkVars();
6077 // Some things are used by both LSRA and regpredict allocators.
6079 FrameType rpFrameType;
6080 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6082 #ifdef LEGACY_BACKEND
6083 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6085 #endif // LEGACY_BACKEND
6087 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6089 #if FEATURE_FP_REGALLOC
6090 enum enumConfigRegisterFP
6092 CONFIG_REGISTER_FP_NONE = 0x0,
6093 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6094 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6095 CONFIG_REGISTER_FP_FULL = 0x3,
6097 enumConfigRegisterFP raConfigRegisterFP();
6098 #endif // FEATURE_FP_REGALLOC
6101 regMaskTP raConfigRestrictMaskFP();
6104 #ifndef LEGACY_BACKEND
6105 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6106 #else // LEGACY_BACKEND
6107 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6108 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6109 bool raNewBlocks; // True is we added killing blocks for FPU registers
6110 unsigned rpPasses; // Number of passes made by the register predicter
6111 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6112 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6113 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6114 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6115 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6116 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6117 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6118 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6119 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6120 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6121 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6122 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6124 bool rpRegAllocDone; // Set to true after we have completed register allocation
6126 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6128 void raSetupArgMasks(RegState* r);
6130 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6132 void raDumpVarIntf(); // Dump the variable to variable interference graph
6133 void raDumpRegIntf(); // Dump the variable to register interference graph
6135 void raAdjustVarIntf();
6137 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6139 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6141 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6142 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6144 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6146 static fgWalkPreFn rpMarkRegIntf;
6148 regMaskTP rpPredictAddressMode(
6149 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6151 void rpPredictRefAssign(unsigned lclNum);
6153 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6155 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6157 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6159 void rpPredictRegUse(); // Entry point
6161 unsigned raPredictTreeRegUse(GenTreePtr tree);
6162 unsigned raPredictListRegUse(GenTreePtr list);
6164 void raSetRegVarOrder(var_types regType,
6165 regNumber* customVarOrder,
6166 unsigned* customVarOrderSize,
6168 regMaskTP avoidReg);
6170 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6171 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6172 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6173 void raAddToStkPredict(unsigned val)
6175 unsigned newStkPredict = rpStkPredict + val;
6176 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6177 rpStkPredict = UINT_MAX - 1;
6179 rpStkPredict = newStkPredict;
6183 #if !FEATURE_FP_REGALLOC
6184 void raDispFPlifeInfo();
6188 regMaskTP genReturnRegForTree(GenTreePtr tree);
6189 #endif // LEGACY_BACKEND
6191 /* raIsVarargsStackArg is called by raMaskStkVars and by
6192 lvaSortByRefCount. It identifies the special case
6193 where a varargs function has a parameter passed on the
6194 stack, other than the special varargs handle. Such parameters
6195 require special treatment, because they cannot be tracked
6196 by the GC (their offsets in the stack are not known
6200 bool raIsVarargsStackArg(unsigned lclNum)
6204 LclVarDsc* varDsc = &lvaTable[lclNum];
6206 assert(varDsc->lvIsParam);
6208 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6210 #else // _TARGET_X86_
6214 #endif // _TARGET_X86_
6217 #ifdef LEGACY_BACKEND
6218 // Records the current prediction, if it's better than any previous recorded prediction.
6219 void rpRecordPrediction();
6220 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6221 void rpUseRecordedPredictionIfBetter();
6223 // Data members used in the methods above.
6224 unsigned rpBestRecordedStkPredict;
6225 struct VarRegPrediction
6227 bool m_isEnregistered;
6228 regNumberSmall m_regNum;
6229 regNumberSmall m_otherReg;
6231 VarRegPrediction* rpBestRecordedPrediction;
6232 #endif // LEGACY_BACKEND
6235 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6236 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6240 XX Get to the class and method info from the Execution Engine given XX
6241 XX tokens for the class and method XX
6243 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6244 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6248 /* These are the different addressing modes used to access a local var.
6249 * The JIT has to report the location of the locals back to the EE
6250 * for debugging purposes.
6256 VLT_REG_BYREF, // this type is currently only used for value types on X64
6259 VLT_STK_BYREF, // this type is currently only used for value types on X64
6273 siVarLocType vlType;
6276 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6278 // VLT_REG_BYREF -- the specified register contains the address of the variable
6286 // VLT_STK -- Any 32 bit value which is on the stack
6287 // eg. [ESP+0x20], or [EBP-0x28]
6288 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6289 // eg. mov EAX, [ESP+0x20]; [EAX]
6293 regNumber vlsBaseReg;
6294 NATIVE_OFFSET vlsOffset;
6297 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6306 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6307 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6315 regNumber vlrssBaseReg;
6316 NATIVE_OFFSET vlrssOffset;
6320 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6321 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6327 regNumber vlsrsBaseReg;
6328 NATIVE_OFFSET vlsrsOffset;
6334 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6335 // eg 2 DWords at [ESP+0x10]
6339 regNumber vls2BaseReg;
6340 NATIVE_OFFSET vls2Offset;
6343 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6344 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6351 // VLT_FIXED_VA -- fixed argument of a varargs function.
6352 // The argument location depends on the size of the variable
6353 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6354 // location of the first arg. This argument can then be accessed
6355 // relative to the position of the first arg
6359 unsigned vlfvOffset;
6366 void* rpValue; // pointer to the in-process
6367 // location of the value.
6373 bool vlIsInReg(regNumber reg);
6374 bool vlIsOnStk(regNumber reg, signed offset);
6377 /*************************************************************************/
6382 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6383 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6384 CORINFO_CALLINFO_FLAGS flags,
6385 CORINFO_CALL_INFO* pResult);
6386 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6388 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6389 CORINFO_ACCESS_FLAGS flags,
6390 CORINFO_FIELD_INFO* pResult);
6394 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6396 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6398 bool IsSuperPMIException(unsigned code)
6400 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6402 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6403 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6404 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6405 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6406 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6407 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6408 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6409 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6413 case EXCEPTIONCODE_DebugBreakorAV:
6414 case EXCEPTIONCODE_MC:
6415 case EXCEPTIONCODE_LWM:
6416 case EXCEPTIONCODE_SASM:
6417 case EXCEPTIONCODE_SSYM:
6418 case EXCEPTIONCODE_CALLUTILS:
6419 case EXCEPTIONCODE_TYPEUTILS:
6420 case EXCEPTIONCODE_ASSERT:
6427 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6428 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6430 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6431 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6434 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6435 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6436 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6438 // VOM info, method sigs
6440 void eeGetSig(unsigned sigTok,
6441 CORINFO_MODULE_HANDLE scope,
6442 CORINFO_CONTEXT_HANDLE context,
6443 CORINFO_SIG_INFO* retSig);
6445 void eeGetCallSiteSig(unsigned sigTok,
6446 CORINFO_MODULE_HANDLE scope,
6447 CORINFO_CONTEXT_HANDLE context,
6448 CORINFO_SIG_INFO* retSig);
6450 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6452 // Method entry-points, instrs
6454 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6456 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6458 CORINFO_EE_INFO eeInfo;
6459 bool eeInfoInitialized;
6461 CORINFO_EE_INFO* eeGetEEInfo();
6463 // Gets the offset of a SDArray's first element
6464 unsigned eeGetArrayDataOffset(var_types type);
6465 // Gets the offset of a MDArray's first element
6466 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6468 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6470 // Returns the page size for the target machine as reported by the EE.
6471 inline size_t eeGetPageSize()
6473 #if COR_JIT_EE_VERSION > 460
6474 return eeGetEEInfo()->osPageSize;
6475 #else // COR_JIT_EE_VERSION <= 460
6476 return CORINFO_PAGE_SIZE;
6477 #endif // COR_JIT_EE_VERSION > 460
6480 // Returns the frame size at which we will generate a loop to probe the stack.
6481 inline size_t getVeryLargeFrameSize()
6484 // The looping probe code is 40 bytes, whereas the straight-line probing for
6485 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6486 // or greater, to generate smaller code.
6487 return 2 * eeGetPageSize();
6489 return 3 * eeGetPageSize();
6493 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6495 #if COR_JIT_EE_VERSION > 460
6496 return eeGetEEInfo()->targetAbi == abi;
6498 return CORINFO_DESKTOP_ABI == abi;
6502 inline bool generateCFIUnwindCodes()
6504 #ifdef UNIX_AMD64_ABI
6505 return IsTargetAbi(CORINFO_CORERT_ABI);
6513 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6515 // Debugging support - Line number info
6517 void eeGetStmtOffsets();
6519 unsigned eeBoundariesCount;
6521 struct boundariesDsc
6523 UNATIVE_OFFSET nativeIP;
6525 unsigned sourceReason;
6526 } * eeBoundaries; // Boundaries to report to EE
6527 void eeSetLIcount(unsigned count);
6528 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6532 static void eeDispILOffs(IL_OFFSET offs);
6533 static void eeDispLineInfo(const boundariesDsc* line);
6534 void eeDispLineInfos();
6537 // Debugging support - Local var info
6541 unsigned eeVarsCount;
6543 struct VarResultInfo
6545 UNATIVE_OFFSET startOffset;
6546 UNATIVE_OFFSET endOffset;
6550 void eeSetLVcount(unsigned count);
6551 void eeSetLVinfo(unsigned which,
6552 UNATIVE_OFFSET startOffs,
6553 UNATIVE_OFFSET length,
6558 const siVarLoc& loc);
6562 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6563 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6566 // ICorJitInfo wrappers
6568 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6570 void eeAllocUnwindInfo(BYTE* pHotCode,
6576 CorJitFuncKind funcKind);
6578 void eeSetEHcount(unsigned cEH);
6580 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6582 WORD eeGetRelocTypeHint(void* target);
6584 // ICorStaticInfo wrapper functions
6586 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6588 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6590 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6593 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6594 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6595 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6596 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6598 template <typename ParamType>
6599 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6601 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6604 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6606 // Utility functions
6608 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6611 const wchar_t* eeGetCPString(size_t stringHandle);
6614 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6616 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6617 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6619 static fgWalkPreFn CountSharedStaticHelper;
6620 static bool IsSharedStaticHelper(GenTreePtr tree);
6621 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6623 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6624 // returns true/false if 'field' is a Jit Data offset
6625 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6626 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6627 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6629 /*****************************************************************************/
6634 enum TEMP_USAGE_TYPE
6640 static var_types tmpNormalizeType(var_types type);
6641 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6642 void tmpRlsTemp(TempDsc* temp);
6643 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6646 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6647 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6651 bool tmpAllFree() const;
6654 #ifndef LEGACY_BACKEND
6655 void tmpPreAllocateTemps(var_types type, unsigned count);
6656 #endif // !LEGACY_BACKEND
6659 #ifdef LEGACY_BACKEND
6660 unsigned tmpIntSpillMax; // number of int-sized spill temps
6661 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6662 #endif // LEGACY_BACKEND
6664 unsigned tmpCount; // Number of temps
6665 unsigned tmpSize; // Size of all the temps
6668 // Used by RegSet::rsSpillChk()
6669 unsigned tmpGetCount; // Temps which haven't been released yet
6672 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6674 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6675 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6678 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6679 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6683 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6684 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6688 CodeGenInterface* codeGen;
6690 // The following holds information about instr offsets in terms of generated code.
6694 IPmappingDsc* ipmdNext; // next line# record
6695 IL_OFFSETX ipmdILoffsx; // the instr offset
6696 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6697 bool ipmdIsLabel; // Can this code be a branch label?
6700 // Record the instr offset mapping to the generated code
6702 IPmappingDsc* genIPmappingList;
6703 IPmappingDsc* genIPmappingLast;
6705 // Managed RetVal - A side hash table meant to record the mapping from a
6706 // GT_CALL node to its IL offset. This info is used to emit sequence points
6707 // that can be used by debugger to determine the native offset at which the
6708 // managed RetVal will be available.
6710 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6711 // favor of a side table for two reasons: 1) We need IL offset for only those
6712 // GT_CALL nodes (created during importation) that correspond to an IL call and
6713 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6714 // structure and IL offset is needed only when generating debuggable code. Therefore
6715 // it is desirable to avoid memory size penalty in retail scenarios.
6716 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6717 CallSiteILOffsetTable;
6718 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6720 unsigned genReturnLocal; // Local number for the return value when applicable.
6721 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6723 // The following properties are part of CodeGenContext. Getters are provided here for
6724 // convenience and backward compatibility, but the properties can only be set by invoking
6725 // the setter on CodeGenContext directly.
6727 __declspec(property(get = getEmitter)) emitter* genEmitter;
6728 emitter* getEmitter()
6730 return codeGen->getEmitter();
6733 const bool isFramePointerUsed()
6735 return codeGen->isFramePointerUsed();
6738 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6739 bool getInterruptible()
6741 return codeGen->genInterruptible;
6743 void setInterruptible(bool value)
6745 codeGen->setInterruptible(value);
6749 const bool genDoubleAlign()
6751 return codeGen->doDoubleAlign();
6753 DWORD getCanDoubleAlign();
6754 bool shouldDoubleAlign(unsigned refCntStk,
6756 unsigned refCntWtdReg,
6757 unsigned refCntStkParam,
6758 unsigned refCntWtdStkDbl);
6759 #endif // DOUBLE_ALIGN
6761 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6762 bool getFullPtrRegMap()
6764 return codeGen->genFullPtrRegMap;
6766 void setFullPtrRegMap(bool value)
6768 codeGen->setFullPtrRegMap(value);
6771 // Things that MAY belong either in CodeGen or CodeGenContext
6773 #if FEATURE_EH_FUNCLETS
6774 FuncInfoDsc* compFuncInfos;
6775 unsigned short compCurrFuncIdx;
6776 unsigned short compFuncInfoCount;
6778 unsigned short compFuncCount()
6780 assert(fgFuncletsCreated);
6781 return compFuncInfoCount;
6784 #else // !FEATURE_EH_FUNCLETS
6786 // This is a no-op when there are no funclets!
6787 void genUpdateCurrentFunclet(BasicBlock* block)
6792 FuncInfoDsc compFuncInfoRoot;
6794 static const unsigned compCurrFuncIdx = 0;
6796 unsigned short compFuncCount()
6801 #endif // !FEATURE_EH_FUNCLETS
6803 FuncInfoDsc* funCurrentFunc();
6804 void funSetCurrentFunc(unsigned funcIdx);
6805 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6806 unsigned int funGetFuncIdx(BasicBlock* block);
6810 VARSET_TP compCurLife; // current live variables
6811 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6813 template <bool ForCodeGen>
6814 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6816 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6818 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6821 template <bool ForCodeGen>
6822 void compUpdateLife(GenTreePtr tree);
6824 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6825 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6826 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6827 template <bool ForCodeGen>
6828 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6830 template <bool ForCodeGen>
6831 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6833 // Gets a register mask that represent the kill set for a helper call since
6834 // not all JIT Helper calls follow the standard ABI on the target architecture.
6835 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6837 // Gets a register mask that represent the kill set for a NoGC helper call.
6838 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6841 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6842 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6843 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6844 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6845 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6846 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6847 #endif // _TARGET_ARM_
6849 // 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
6851 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6853 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6854 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6855 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6856 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6857 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6858 // for the tracked var indices of the field vars, as in a live var set).
6859 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6861 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6863 if (m_promotedStructDeathVars == nullptr)
6865 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6867 return m_promotedStructDeathVars;
6871 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6872 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6876 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6877 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6880 #if !defined(__GNUC__)
6881 #pragma region Unwind information
6886 // Infrastructure functions: start/stop/reserve/emit.
6889 void unwindBegProlog();
6890 void unwindEndProlog();
6891 void unwindBegEpilog();
6892 void unwindEndEpilog();
6893 void unwindReserve();
6894 void unwindEmit(void* pHotCode, void* pColdCode);
6897 // Specific unwind information functions: called by code generation to indicate a particular
6898 // prolog or epilog unwindable instruction has been generated.
6901 void unwindPush(regNumber reg);
6902 void unwindAllocStack(unsigned size);
6903 void unwindSetFrameReg(regNumber reg, unsigned offset);
6904 void unwindSaveReg(regNumber reg, unsigned offset);
6906 #if defined(_TARGET_ARM_)
6907 void unwindPushMaskInt(regMaskTP mask);
6908 void unwindPushMaskFloat(regMaskTP mask);
6909 void unwindPopMaskInt(regMaskTP mask);
6910 void unwindPopMaskFloat(regMaskTP mask);
6911 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6912 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6913 // called via unwindPadding().
6914 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6915 // instruction and the current location.
6916 #endif // _TARGET_ARM_
6918 #if defined(_TARGET_ARM64_)
6920 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6921 // instruction and the current location.
6922 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6923 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6924 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6925 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6926 void unwindSaveNext(); // unwind code: save_next
6927 void unwindReturn(regNumber reg); // ret lr
6928 #endif // defined(_TARGET_ARM64_)
6931 // Private "helper" functions for the unwind implementation.
6935 #if FEATURE_EH_FUNCLETS
6936 void unwindGetFuncLocations(FuncInfoDsc* func,
6937 bool getHotSectionData,
6938 /* OUT */ emitLocation** ppStartLoc,
6939 /* OUT */ emitLocation** ppEndLoc);
6940 #endif // FEATURE_EH_FUNCLETS
6942 void unwindReserveFunc(FuncInfoDsc* func);
6943 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6945 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6947 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6948 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6950 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6952 #if defined(_TARGET_AMD64_)
6954 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6956 void unwindBegPrologWindows();
6957 void unwindPushWindows(regNumber reg);
6958 void unwindAllocStackWindows(unsigned size);
6959 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6960 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6962 #ifdef UNIX_AMD64_ABI
6963 void unwindBegPrologCFI();
6964 void unwindPushCFI(regNumber reg);
6965 void unwindAllocStackCFI(unsigned size);
6966 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6967 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6968 int mapRegNumToDwarfReg(regNumber reg);
6969 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6970 #endif // UNIX_AMD64_ABI
6971 #elif defined(_TARGET_ARM_)
6973 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6974 void unwindPushPopMaskFloat(regMaskTP mask);
6975 void unwindSplit(FuncInfoDsc* func);
6977 #endif // _TARGET_ARM_
6979 #if !defined(__GNUC__)
6980 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6984 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6985 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6989 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6990 XX that contains the distinguished, well-known SIMD type definitions). XX
6992 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6993 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6996 // Get highest available instruction set for floating point codegen
6997 InstructionSet getFloatingPointInstructionSet()
6999 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7002 return InstructionSet_AVX;
7007 return InstructionSet_SSE3_4;
7011 assert(canUseSSE2());
7012 return InstructionSet_SSE2;
7014 assert(!"getFPInstructionSet() is not implemented for target arch");
7016 return InstructionSet_NONE;
7020 // Get highest available instruction set for SIMD codegen
7021 InstructionSet getSIMDInstructionSet()
7023 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7024 return getFloatingPointInstructionSet();
7026 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7028 return InstructionSet_NONE;
7034 // Should we support SIMD intrinsics?
7037 // Have we identified any SIMD types?
7038 // This is currently used by struct promotion to avoid getting type information for a struct
7039 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7041 bool _usesSIMDTypes;
7042 bool usesSIMDTypes()
7044 return _usesSIMDTypes;
7046 void setUsesSIMDTypes(bool value)
7048 _usesSIMDTypes = value;
7051 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7052 // that require indexed access to the individual fields of the vector, which is not well supported
7053 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7054 unsigned lvaSIMDInitTempVarNum;
7057 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7058 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7059 CORINFO_CLASS_HANDLE SIMDIntHandle;
7060 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7061 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7062 CORINFO_CLASS_HANDLE SIMDShortHandle;
7063 CORINFO_CLASS_HANDLE SIMDByteHandle;
7064 CORINFO_CLASS_HANDLE SIMDLongHandle;
7065 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7066 CORINFO_CLASS_HANDLE SIMDULongHandle;
7067 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7068 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7069 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7070 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7072 // Get the handle for a SIMD type.
7073 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7075 if (simdBaseType == TYP_FLOAT)
7080 return SIMDVector2Handle;
7082 return SIMDVector3Handle;
7084 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7086 return SIMDVector4Handle;
7095 assert(simdType == getSIMDVectorType());
7096 switch (simdBaseType)
7099 return SIMDFloatHandle;
7101 return SIMDDoubleHandle;
7103 return SIMDIntHandle;
7105 return SIMDUShortHandle;
7107 return SIMDUShortHandle;
7109 return SIMDUByteHandle;
7111 return SIMDShortHandle;
7113 return SIMDByteHandle;
7115 return SIMDLongHandle;
7117 return SIMDUIntHandle;
7119 return SIMDULongHandle;
7121 assert(!"Didn't find a class handle for simdType");
7123 return NO_CLASS_HANDLE;
7127 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7128 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7129 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7131 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7132 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7133 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7134 bool isSIMDTypeLocal(GenTree* tree)
7136 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7139 // Returns true if the type of the tree is a byref of TYP_SIMD
7140 bool isAddrOfSIMDType(GenTree* tree)
7142 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7144 switch (tree->OperGet())
7147 return varTypeIsSIMD(tree->gtGetOp1());
7149 case GT_LCL_VAR_ADDR:
7150 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7153 return isSIMDTypeLocal(tree);
7160 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7162 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7163 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7164 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7167 // Returns base type of a TYP_SIMD local.
7168 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7169 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7171 if (isSIMDTypeLocal(tree))
7173 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7179 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7181 return info.compCompHnd->isInSIMDModule(clsHnd);
7184 bool isSIMDClass(typeInfo* pTypeInfo)
7186 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7189 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7190 // if it is not a SIMD type or is an unsupported base type.
7191 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7193 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7195 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7198 // Get SIMD Intrinsic info given the method handle.
7199 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7200 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7201 CORINFO_METHOD_HANDLE methodHnd,
7202 CORINFO_SIG_INFO* sig,
7205 var_types* baseType,
7206 unsigned* sizeBytes);
7208 // Pops and returns GenTree node from importers type stack.
7209 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7210 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7212 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7213 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7215 // Creates a GT_SIMD tree for Select operation
7216 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7218 unsigned simdVectorSize,
7223 // Creates a GT_SIMD tree for Min/Max operation
7224 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7225 CORINFO_CLASS_HANDLE typeHnd,
7227 unsigned simdVectorSize,
7231 // Transforms operands and returns the SIMD intrinsic to be applied on
7232 // transformed operands to obtain given relop result.
7233 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7234 CORINFO_CLASS_HANDLE typeHnd,
7235 unsigned simdVectorSize,
7236 var_types* baseType,
7240 // Creates a GT_SIMD tree for Abs intrinsic.
7241 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7243 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7244 // Transforms operands and returns the SIMD intrinsic to be applied on
7245 // transformed operands to obtain == comparison result.
7246 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7247 unsigned simdVectorSize,
7251 // Transforms operands and returns the SIMD intrinsic to be applied on
7252 // transformed operands to obtain > comparison result.
7253 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7254 unsigned simdVectorSize,
7258 // Transforms operands and returns the SIMD intrinsic to be applied on
7259 // transformed operands to obtain >= comparison result.
7260 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7261 unsigned simdVectorSize,
7265 // Transforms operands and returns the SIMD intrinsic to be applied on
7266 // transformed operands to obtain >= comparison result in case of int32
7267 // and small int base type vectors.
7268 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7269 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7270 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7272 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7273 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7274 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7275 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7276 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7278 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7279 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7280 GenTreePtr newobjThis,
7281 CORINFO_CLASS_HANDLE clsHnd,
7282 CORINFO_METHOD_HANDLE method,
7283 CORINFO_SIG_INFO* sig,
7286 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7288 // Whether SIMD vector occupies part of SIMD register.
7289 // SSE2: vector2f/3f are considered sub register SIMD types.
7290 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7291 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7293 unsigned sizeBytes = 0;
7294 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7295 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7298 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7300 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7303 // Get the type for the hardware SIMD vector.
7304 // This is the maximum SIMD type supported for this target.
7305 var_types getSIMDVectorType()
7307 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7314 assert(canUseSSE2());
7318 assert(!"getSIMDVectorType() unimplemented on target arch");
7323 // Get the size of the SIMD type in bytes
7324 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7326 unsigned sizeBytes = 0;
7327 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7331 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7332 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7334 // Get the the number of elements of basetype of SIMD vector given by its type handle
7335 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7337 // Get preferred alignment of SIMD type.
7338 int getSIMDTypeAlignment(var_types simdType);
7340 // Get the number of bytes in a SIMD Vector for the current compilation.
7341 unsigned getSIMDVectorRegisterByteLength()
7343 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7346 return YMM_REGSIZE_BYTES;
7350 assert(canUseSSE2());
7351 return XMM_REGSIZE_BYTES;
7354 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7359 // The minimum and maximum possible number of bytes in a SIMD vector.
7360 unsigned int maxSIMDStructBytes()
7362 return getSIMDVectorRegisterByteLength();
7364 unsigned int minSIMDStructBytes()
7366 return emitTypeSize(TYP_SIMD8);
7369 #ifdef FEATURE_AVX_SUPPORT
7370 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7371 static const unsigned maxPossibleSIMDStructBytes = 32;
7372 #else // !FEATURE_AVX_SUPPORT
7373 static const unsigned maxPossibleSIMDStructBytes = 16;
7374 #endif // !FEATURE_AVX_SUPPORT
7376 // Returns the codegen type for a given SIMD size.
7377 var_types getSIMDTypeForSize(unsigned size)
7379 var_types simdType = TYP_UNDEF;
7382 simdType = TYP_SIMD8;
7384 else if (size == 12)
7386 simdType = TYP_SIMD12;
7388 else if (size == 16)
7390 simdType = TYP_SIMD16;
7392 #ifdef FEATURE_AVX_SUPPORT
7393 else if (size == 32)
7395 simdType = TYP_SIMD32;
7397 #endif // FEATURE_AVX_SUPPORT
7400 noway_assert(!"Unexpected size for SIMD type");
7405 unsigned getSIMDInitTempVarNum()
7407 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7409 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7410 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7412 return lvaSIMDInitTempVarNum;
7415 #endif // FEATURE_SIMD
7418 //------------------------------------------------------------------------
7419 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7421 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7422 // candidate for enregistration.
7424 unsigned largestEnregisterableStructSize()
7427 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7428 if (vectorRegSize > TARGET_POINTER_SIZE)
7430 return vectorRegSize;
7433 #endif // FEATURE_SIMD
7435 return TARGET_POINTER_SIZE;
7440 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7441 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7442 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7444 // Is this var is of type simd struct?
7445 bool lclVarIsSIMDType(unsigned varNum)
7447 LclVarDsc* varDsc = lvaTable + varNum;
7448 return varDsc->lvIsSIMDType();
7451 // Is this Local node a SIMD local?
7452 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7454 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7457 // Returns true if the TYP_SIMD locals on stack are aligned at their
7458 // preferred byte boundary specified by getSIMDTypeAlignment().
7460 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7461 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7462 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7463 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7464 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7465 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7466 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7469 bool isSIMDTypeLocalAligned(unsigned varNum)
7471 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7472 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7475 int off = lvaFrameAddress(varNum, &ebpBased);
7476 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7477 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7478 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7481 #endif // FEATURE_SIMD
7486 // Whether SSE2 is available
7487 bool canUseSSE2() const
7489 #ifdef _TARGET_XARCH_
7490 return opts.compCanUseSSE2;
7496 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7497 bool CanUseSSE3_4() const
7499 #ifdef _TARGET_XARCH_
7500 return opts.compCanUseSSE3_4;
7506 bool canUseAVX() const
7508 #ifdef FEATURE_AVX_SUPPORT
7509 return opts.compCanUseAVX;
7516 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7517 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7521 XX Generic info about the compilation and the method being compiled. XX
7522 XX It is responsible for driving the other phases. XX
7523 XX It is also responsible for all the memory management. XX
7525 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7526 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7530 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7532 InlineResult* compInlineResult; // The result of importing the inlinee method.
7534 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7535 bool compJmpOpUsed; // Does the method do a JMP
7536 bool compLongUsed; // Does the method use TYP_LONG
7537 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7538 bool compTailCallUsed; // Does the method do a tailcall
7539 bool compLocallocUsed; // Does the method use localloc.
7540 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7541 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7542 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7544 // NOTE: These values are only reliable after
7545 // the importing is completely finished.
7547 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7548 // we can iterate over these efficiently.
7550 #if CPU_USES_BLOCK_MOVE
7551 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7555 // State information - which phases have completed?
7556 // These are kept together for easy discoverability
7558 bool bRangeAllowStress;
7559 bool compCodeGenDone;
7560 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7561 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7562 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7563 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7566 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7567 bool fgLocalVarLivenessChanged;
7569 bool compStackProbePrologDone;
7571 #ifndef LEGACY_BACKEND
7573 #endif // !LEGACY_BACKEND
7574 bool compRationalIRForm;
7576 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7578 bool compGeneratingProlog;
7579 bool compGeneratingEpilog;
7580 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7581 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7582 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7583 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7584 bool getNeedsGSSecurityCookie() const
7586 return compNeedsGSSecurityCookie;
7588 void setNeedsGSSecurityCookie()
7590 compNeedsGSSecurityCookie = true;
7593 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7594 // frame layout calculations, this is the level we are currently
7597 //---------------------------- JITing options -----------------------------
7610 JitFlags* jitFlags; // all flags passed from the EE
7611 unsigned compFlags; // method attributes
7613 codeOptimize compCodeOpt; // what type of code optimizations
7617 #ifdef _TARGET_XARCH_
7618 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7619 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7621 #ifdef FEATURE_AVX_SUPPORT
7622 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7623 #endif // FEATURE_AVX_SUPPORT
7624 #endif // _TARGET_XARCH_
7626 // optimize maximally and/or favor speed over size?
7628 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7629 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7630 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7631 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7632 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7634 // Maximun number of locals before turning off the inlining
7635 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7638 unsigned instrCount;
7639 unsigned lvRefCount;
7640 bool compMinOptsIsSet;
7642 bool compMinOptsIsUsed;
7644 inline bool MinOpts()
7646 assert(compMinOptsIsSet);
7647 compMinOptsIsUsed = true;
7650 inline bool IsMinOptsSet()
7652 return compMinOptsIsSet;
7655 inline bool MinOpts()
7659 inline bool IsMinOptsSet()
7661 return compMinOptsIsSet;
7664 inline void SetMinOpts(bool val)
7666 assert(!compMinOptsIsUsed);
7667 assert(!compMinOptsIsSet || (compMinOpts == val));
7669 compMinOptsIsSet = true;
7672 // true if the CLFLG_* for an optimization is set.
7673 inline bool OptEnabled(unsigned optFlag)
7675 return !!(compFlags & optFlag);
7678 #ifdef FEATURE_READYTORUN_COMPILER
7679 inline bool IsReadyToRun()
7681 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7684 inline bool IsReadyToRun()
7690 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7691 // PInvoke transitions inline (e.g. when targeting CoreRT).
7692 inline bool ShouldUsePInvokeHelpers()
7694 #if COR_JIT_EE_VERSION > 460
7695 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7701 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7703 inline bool IsReversePInvoke()
7705 #if COR_JIT_EE_VERSION > 460
7706 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7712 // true if we must generate code compatible with JIT32 quirks
7713 inline bool IsJit32Compat()
7715 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7716 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7722 // true if we must generate code compatible with Jit64 quirks
7723 inline bool IsJit64Compat()
7725 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7726 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7727 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7734 bool compScopeInfo; // Generate the LocalVar info ?
7735 bool compDbgCode; // Generate debugger-friendly code?
7736 bool compDbgInfo; // Gather debugging info?
7739 #ifdef PROFILING_SUPPORTED
7740 bool compNoPInvokeInlineCB;
7742 static const bool compNoPInvokeInlineCB;
7746 bool compGcChecks; // Check arguments and return values to ensure they are sane
7747 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7748 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7752 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7753 // to be allocated on the stack.
7754 // It will be set to true in the following cases:
7755 // 1. When the method being compiled has a declarative security
7756 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7757 // This is also the case when we inject a prolog and epilog in the method.
7759 // 2. When the method being compiled has imperative security (i.e. the method
7760 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7762 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7764 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7765 // which gets reported as a GC root to stackwalker.
7766 // (See also ICodeManager::GetAddrOfSecurityObject.)
7773 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7774 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7778 #ifdef UNIX_AMD64_ABI
7779 // This flag is indicating if there is a need to align the frame.
7780 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7781 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7782 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7783 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7784 // there are calls and making sure the frame alignment logic is executed.
7785 bool compNeedToAlignFrame;
7786 #endif // UNIX_AMD64_ABI
7788 bool compProcedureSplitting; // Separate cold code from hot code
7790 bool genFPorder; // Preserve FP order (operations are non-commutative)
7791 bool genFPopt; // Can we do frame-pointer-omission optimization?
7792 bool altJit; // True if we are an altjit and are compiling this method
7795 bool optRepeat; // Repeat optimizer phases k times
7799 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7800 bool dspCode; // Display native code generated
7801 bool dspEHTable; // Display the EH table reported to the VM
7802 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7803 bool dspEmit; // Display emitter output
7804 bool dspLines; // Display source-code lines intermixed with native code output
7805 bool dmpHex; // Display raw bytes in hex of native code output
7806 bool varNames; // Display variables names in native code output
7807 bool disAsm; // Display native code as it is generated
7808 bool disAsmSpilled; // Display native code when any register spilling occurs
7809 bool disDiffable; // Makes the Disassembly code 'diff-able'
7810 bool disAsm2; // Display native code after it is generated using external disassembler
7811 bool dspOrder; // Display names of each of the methods that we ngen/jit
7812 bool dspUnwind; // Display the unwind info output
7813 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7814 bool compLongAddress; // Force using large pseudo instructions for long address
7815 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7816 bool dspGCtbls; // Display the GC tables
7820 bool doLateDisasm; // Run the late disassembler
7821 #endif // LATE_DISASM
7823 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7824 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7825 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7826 static const bool dspGCtbls = true;
7829 // We need stack probes to guarantee that we won't trigger a stack overflow
7830 // when calling unmanaged code until they get a chance to set up a frame, because
7831 // the EE will have no idea where it is.
7833 // We will only be doing this currently for hosted environments. Unfortunately
7834 // we need to take care of stubs, so potentially, we will have to do the probes
7835 // for any call. We have a plan for not needing for stubs though
7836 bool compNeedStackProbes;
7838 #ifdef PROFILING_SUPPORTED
7839 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7840 // This option helps make the JIT behave as if it is running under a profiler.
7841 bool compJitELTHookEnabled;
7842 #endif // PROFILING_SUPPORTED
7844 #if FEATURE_TAILCALL_OPT
7845 // Whether opportunistic or implicit tail call optimization is enabled.
7846 bool compTailCallOpt;
7847 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7848 bool compTailCallLoopOpt;
7852 static const bool compUseSoftFP = true;
7853 #else // !ARM_SOFTFP
7854 static const bool compUseSoftFP = false;
7857 GCPollType compGCPollType;
7861 static bool s_pAltJitExcludeAssembliesListInitialized;
7862 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7867 template <typename T>
7870 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7873 template <typename T>
7876 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7879 static int dspTreeID(GenTree* tree)
7881 return tree->gtTreeID;
7883 static void printTreeID(GenTree* tree)
7885 if (tree == nullptr)
7891 printf("[%06d]", dspTreeID(tree));
7898 #define STRESS_MODES \
7902 /* "Variations" stress areas which we try to mix up with each other. */ \
7903 /* These should not be exhaustively used as they might */ \
7904 /* hide/trivialize other areas */ \
7907 STRESS_MODE(DBL_ALN) \
7908 STRESS_MODE(LCL_FLDS) \
7909 STRESS_MODE(UNROLL_LOOPS) \
7910 STRESS_MODE(MAKE_CSE) \
7911 STRESS_MODE(LEGACY_INLINE) \
7912 STRESS_MODE(CLONE_EXPR) \
7913 STRESS_MODE(USE_FCOMI) \
7914 STRESS_MODE(USE_CMOV) \
7916 STRESS_MODE(BB_PROFILE) \
7917 STRESS_MODE(OPT_BOOLS_GC) \
7918 STRESS_MODE(REMORPH_TREES) \
7919 STRESS_MODE(64RSLT_MUL) \
7920 STRESS_MODE(DO_WHILE_LOOPS) \
7921 STRESS_MODE(MIN_OPTS) \
7922 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7923 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7924 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7925 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7926 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7927 STRESS_MODE(NULL_OBJECT_CHECK) \
7928 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7929 STRESS_MODE(RANDOM_INLINE) \
7930 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
7931 STRESS_MODE(GENERIC_VARN) \
7933 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
7935 STRESS_MODE(COUNT_VARN) \
7937 /* "Check" stress areas that can be exhaustively used if we */ \
7938 /* dont care about performance at all */ \
7940 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7941 STRESS_MODE(CHK_FLOW_UPDATE) \
7942 STRESS_MODE(EMITTER) \
7943 STRESS_MODE(CHK_REIMPORT) \
7944 STRESS_MODE(FLATFP) \
7945 STRESS_MODE(GENERIC_CHECK) \
7950 #define STRESS_MODE(mode) STRESS_##mode,
7957 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7958 BYTE compActiveStressModes[STRESS_COUNT];
7961 #define MAX_STRESS_WEIGHT 100
7963 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7967 bool compInlineStress()
7969 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7972 bool compRandomInlineStress()
7974 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7979 bool compTailCallStress()
7982 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7988 codeOptimize compCodeOpt()
7991 // Switching between size & speed has measurable throughput impact
7992 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7993 // DEBUG, but should generate identical code between CHK & RET builds,
7994 // so that's not acceptable.
7995 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7996 // Investigate the cause of the throughput regression.
7998 return opts.compCodeOpt;
8000 return BLENDED_CODE;
8004 //--------------------- Info about the procedure --------------------------
8008 COMP_HANDLE compCompHnd;
8009 CORINFO_MODULE_HANDLE compScopeHnd;
8010 CORINFO_CLASS_HANDLE compClassHnd;
8011 CORINFO_METHOD_HANDLE compMethodHnd;
8012 CORINFO_METHOD_INFO* compMethodInfo;
8014 BOOL hasCircularClassConstraints;
8015 BOOL hasCircularMethodConstraints;
8017 #if defined(DEBUG) || defined(LATE_DISASM)
8018 const char* compMethodName;
8019 const char* compClassName;
8020 const char* compFullName;
8021 #endif // defined(DEBUG) || defined(LATE_DISASM)
8023 #if defined(DEBUG) || defined(INLINE_DATA)
8024 // Method hash is logcally const, but computed
8026 mutable unsigned compMethodHashPrivate;
8027 unsigned compMethodHash() const;
8028 #endif // defined(DEBUG) || defined(INLINE_DATA)
8030 #ifdef PSEUDORANDOM_NOP_INSERTION
8031 // things for pseudorandom nop insertion
8032 unsigned compChecksum;
8036 // The following holds the FLG_xxxx flags for the method we're compiling.
8039 // The following holds the class attributes for the method we're compiling.
8040 unsigned compClassAttr;
8042 const BYTE* compCode;
8043 IL_OFFSET compILCodeSize; // The IL code size
8044 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8045 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8046 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8047 // (2) the code is hot/cold split, and we issued less code than we expected
8048 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8050 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8051 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8052 bool compIsContextful : 1; // contextful method
8053 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8054 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8055 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8056 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8057 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8059 var_types compRetType; // Return type of the method as declared in IL
8060 var_types compRetNativeType; // Normalized return type as per target arch ABI
8061 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8062 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8063 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8064 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8065 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8066 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8067 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8068 unsigned compMaxStack;
8069 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8070 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8072 unsigned compCallUnmanaged; // count of unmanaged calls
8073 unsigned compLvFrameListRoot; // lclNum for the Frame root
8074 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8075 // You should generally use compHndBBtabCount instead: it is the
8076 // current number of EH clauses (after additions like synchronized
8077 // methods and funclets, and removals like unreachable code deletion).
8079 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8080 // and the VM expects that, or the JIT is a "self-host" compiler
8081 // (e.g., x86 hosted targeting x86) and the VM expects that.
8083 /* The following holds IL scope information about local variables.
8086 unsigned compVarScopesCount;
8087 VarScopeDsc* compVarScopes;
8089 /* The following holds information about instr offsets for
8090 * which we need to report IP-mappings
8093 IL_OFFSET* compStmtOffsets; // sorted
8094 unsigned compStmtOffsetsCount;
8095 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8097 #define CPU_X86 0x0100 // The generic X86 CPU
8098 #define CPU_X86_PENTIUM_4 0x0110
8100 #define CPU_X64 0x0200 // The generic x64 CPU
8101 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8102 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8104 #define CPU_ARM 0x0300 // The generic ARM CPU
8106 unsigned genCPU; // What CPU are we running on
8109 // Returns true if the method being compiled returns a non-void and non-struct value.
8110 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8111 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8112 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8113 // Methods returning such structs are considered to return non-struct return value and
8114 // this method returns true in that case.
8115 bool compMethodReturnsNativeScalarType()
8117 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8120 // Returns true if the method being compiled returns RetBuf addr as its return value
8121 bool compMethodReturnsRetBufAddr()
8123 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8124 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8126 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8127 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8128 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8129 // methods with hidden RetBufArg.
8131 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8132 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8133 // returning the address of RetBuf.
8135 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8136 // to be returned in RAX.
8137 CLANG_FORMAT_COMMENT_ANCHOR;
8139 #ifdef _TARGET_AMD64_
8140 return (info.compRetBuffArg != BAD_VAR_NUM);
8141 #else // !_TARGET_AMD64_
8142 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8143 #endif // !_TARGET_AMD64_
8146 // Returns true if the method returns a value in more than one return register
8147 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8148 // TODO-ARM64: Does this apply for ARM64 too?
8149 bool compMethodReturnsMultiRegRetType()
8151 #if FEATURE_MULTIREG_RET
8152 #if defined(_TARGET_X86_)
8153 // On x86 only 64-bit longs are returned in multiple registers
8154 return varTypeIsLong(info.compRetNativeType);
8155 #else // targets: X64-UNIX, ARM64 or ARM32
8156 // On all other targets that support multireg return values:
8157 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8158 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8159 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8160 #endif // TARGET_XXX
8162 #else // not FEATURE_MULTIREG_RET
8164 // For this architecture there are no multireg returns
8167 #endif // FEATURE_MULTIREG_RET
8170 #if FEATURE_MULTIREG_ARGS
8171 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8172 // return the gcPtr layout for the pointers sized fields
8173 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8174 #endif // FEATURE_MULTIREG_ARGS
8176 // Returns true if the method being compiled returns a value
8177 bool compMethodHasRetVal()
8179 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8180 compMethodReturnsMultiRegRetType();
8185 void compDispLocalVars();
8189 //-------------------------- Global Compiler Data ------------------------------------
8192 static unsigned s_compMethodsCount; // to produce unique label names
8193 unsigned compGenTreeID;
8196 BasicBlock* compCurBB; // the current basic block in process
8197 GenTreePtr compCurStmt; // the current statement in process
8199 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8202 // The following is used to create the 'method JIT info' block.
8203 size_t compInfoBlkSize;
8204 BYTE* compInfoBlkAddr;
8206 EHblkDsc* compHndBBtab; // array of EH data
8207 unsigned compHndBBtabCount; // element count of used elements in EH data array
8208 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8210 #if defined(_TARGET_X86_)
8212 //-------------------------------------------------------------------------
8213 // Tracking of region covered by the monitor in synchronized methods
8214 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8215 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8217 #endif // !_TARGET_X86_
8219 Phases previousCompletedPhase; // the most recently completed phase
8221 //-------------------------------------------------------------------------
8222 // The following keeps track of how many bytes of local frame space we've
8223 // grabbed so far in the current function, and how many argument bytes we
8224 // need to pop when we return.
8227 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8229 // Count of callee-saved regs we pushed in the prolog.
8230 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8231 // In case of Amd64 this doesn't include float regs saved on stack.
8232 unsigned compCalleeRegsPushed;
8234 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8235 // Mask of callee saved float regs on stack.
8236 regMaskTP compCalleeFPRegsSavedMask;
8238 #ifdef _TARGET_AMD64_
8239 // Quirk for VS debug-launch scenario to work:
8240 // Bytes of padding between save-reg area and locals.
8241 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8242 unsigned compVSQuirkStackPaddingNeeded;
8243 bool compQuirkForPPPflag;
8246 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8248 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8249 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8250 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8252 //-------------------------------------------------------------------------
8254 static void compStartup(); // One-time initialization
8255 static void compShutdown(); // One-time finalization
8257 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8260 static void compDisplayStaticSizes(FILE* fout);
8262 //------------ Some utility functions --------------
8264 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8265 void** ppIndirection); /* OUT */
8267 // Several JIT/EE interface functions return a CorInfoType, and also return a
8268 // class handle as an out parameter if the type is a value class. Returns the
8269 // size of the type these describe.
8270 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8273 // Components used by the compiler may write unit test suites, and
8274 // have them run within this method. They will be run only once per process, and only
8275 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8276 // These should fail by asserting.
8277 void compDoComponentUnitTestsOnce();
8280 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8281 CORINFO_MODULE_HANDLE classPtr,
8282 COMP_HANDLE compHnd,
8283 CORINFO_METHOD_INFO* methodInfo,
8284 void** methodCodePtr,
8285 ULONG* methodCodeSize,
8286 JitFlags* compileFlags);
8287 void compCompileFinish();
8288 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8289 COMP_HANDLE compHnd,
8290 CORINFO_METHOD_INFO* methodInfo,
8291 void** methodCodePtr,
8292 ULONG* methodCodeSize,
8293 JitFlags* compileFlags,
8294 CorInfoInstantiationVerification instVerInfo);
8296 ArenaAllocator* compGetAllocator();
8298 #if MEASURE_MEM_ALLOC
8300 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8304 unsigned allocCnt; // # of allocs
8305 UINT64 allocSz; // total size of those alloc.
8306 UINT64 allocSzMax; // Maximum single allocation.
8307 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8308 UINT64 nraTotalSizeAlloc;
8309 UINT64 nraTotalSizeUsed;
8311 static const char* s_CompMemKindNames[]; // Names of the kinds.
8313 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8315 for (int i = 0; i < CMK_Count; i++)
8317 allocSzByKind[i] = 0;
8320 MemStats(const MemStats& ms)
8321 : allocCnt(ms.allocCnt)
8322 , allocSz(ms.allocSz)
8323 , allocSzMax(ms.allocSzMax)
8324 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8325 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8327 for (int i = 0; i < CMK_Count; i++)
8329 allocSzByKind[i] = ms.allocSzByKind[i];
8333 // Until we have ubiquitous constructors.
8336 this->MemStats::MemStats();
8339 void AddAlloc(size_t sz, CompMemKind cmk)
8343 if (sz > allocSzMax)
8347 allocSzByKind[cmk] += sz;
8350 void Print(FILE* f); // Print these stats to f.
8351 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8353 MemStats genMemStats;
8355 struct AggregateMemStats : public MemStats
8359 AggregateMemStats() : MemStats(), nMethods(0)
8363 void Add(const MemStats& ms)
8366 allocCnt += ms.allocCnt;
8367 allocSz += ms.allocSz;
8368 allocSzMax = max(allocSzMax, ms.allocSzMax);
8369 for (int i = 0; i < CMK_Count; i++)
8371 allocSzByKind[i] += ms.allocSzByKind[i];
8373 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8374 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8377 void Print(FILE* f); // Print these stats to jitstdout.
8380 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8381 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8382 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8384 #endif // MEASURE_MEM_ALLOC
8386 #if LOOP_HOIST_STATS
8387 unsigned m_loopsConsidered;
8388 bool m_curLoopHasHoistedExpression;
8389 unsigned m_loopsWithHoistedExpressions;
8390 unsigned m_totalHoistedExpressions;
8392 void AddLoopHoistStats();
8393 void PrintPerMethodLoopHoistStats();
8395 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8396 static unsigned s_loopsConsidered;
8397 static unsigned s_loopsWithHoistedExpressions;
8398 static unsigned s_totalHoistedExpressions;
8400 static void PrintAggregateLoopHoistStats(FILE* f);
8401 #endif // LOOP_HOIST_STATS
8403 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8404 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8405 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8406 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8407 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8408 void compFreeMem(void*);
8410 bool compIsForImportOnly();
8411 bool compIsForInlining();
8412 bool compDonotInline();
8415 const char* compLocalVarName(unsigned varNum, unsigned offs);
8416 VarName compVarName(regNumber reg, bool isFloatReg = false);
8417 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8418 const char* compRegPairName(regPairNo regPair);
8419 const char* compRegNameForSize(regNumber reg, size_t size);
8420 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8421 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8422 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8425 //-------------------------------------------------------------------------
8427 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8429 struct VarScopeMapInfo
8431 VarScopeListNode* head;
8432 VarScopeListNode* tail;
8433 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8435 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8442 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8443 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8445 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8446 VarNumToScopeDscMap;
8448 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8449 VarNumToScopeDscMap* compVarScopeMap;
8451 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8453 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8455 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8457 void compInitVarScopeMap();
8459 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8460 // enter scope, sorted by instr offset
8461 unsigned compNextEnterScope;
8463 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8464 // go out of scope, sorted by instr offset
8465 unsigned compNextExitScope;
8467 void compInitScopeLists();
8469 void compResetScopeLists();
8471 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8473 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8475 void compProcessScopesUntil(unsigned offset,
8477 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8478 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8481 void compDispScopeLists();
8484 bool compIsProfilerHookNeeded();
8486 //-------------------------------------------------------------------------
8487 /* Statistical Data Gathering */
8489 void compJitStats(); // call this function and enable
8490 // various ifdef's below for statistical data
8493 void compCallArgStats();
8494 static void compDispCallArgStats(FILE* fout);
8497 //-------------------------------------------------------------------------
8504 ArenaAllocator* compAllocator;
8507 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8508 // suitable for use by utilcode collection types.
8509 IAllocator* compAsIAllocator;
8511 #if MEASURE_MEM_ALLOC
8512 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8513 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8514 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8516 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8518 #endif // MEASURE_MEM_ALLOC
8520 void compFunctionTraceStart();
8521 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8524 size_t compMaxUncheckedOffsetForNullObject;
8526 void compInitOptions(JitFlags* compileFlags);
8528 void compSetProcessor();
8529 void compInitDebuggingInfo();
8530 void compSetOptimizationLevel();
8531 #ifdef _TARGET_ARMARCH_
8532 bool compRsvdRegCheck(FrameLayoutState curState);
8534 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8536 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8537 void ResetOptAnnotations();
8539 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8540 void RecomputeLoopInfo();
8542 #ifdef PROFILING_SUPPORTED
8543 // Data required for generating profiler Enter/Leave/TailCall hooks
8545 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8546 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8547 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8550 #ifdef _TARGET_AMD64_
8551 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8554 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8555 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8557 IAllocator* getAllocator()
8559 return compAsIAllocator;
8562 #if MEASURE_MEM_ALLOC
8563 IAllocator* getAllocatorBitset()
8565 return compAsIAllocatorBitset;
8567 IAllocator* getAllocatorGC()
8569 return compAsIAllocatorGC;
8571 IAllocator* getAllocatorLoopHoist()
8573 return compAsIAllocatorLoopHoist;
8575 #else // !MEASURE_MEM_ALLOC
8576 IAllocator* getAllocatorBitset()
8578 return compAsIAllocator;
8580 IAllocator* getAllocatorGC()
8582 return compAsIAllocator;
8584 IAllocator* getAllocatorLoopHoist()
8586 return compAsIAllocator;
8588 #endif // !MEASURE_MEM_ALLOC
8591 IAllocator* getAllocatorDebugOnly()
8593 #if MEASURE_MEM_ALLOC
8594 return compAsIAllocatorDebugOnly;
8595 #else // !MEASURE_MEM_ALLOC
8596 return compAsIAllocator;
8597 #endif // !MEASURE_MEM_ALLOC
8602 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8603 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8607 XX Checks for type compatibility and merges types XX
8609 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8610 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8614 // Set to TRUE if verification cannot be skipped for this method
8615 // If we detect unverifiable code, we will lazily check
8616 // canSkipMethodVerification() to see if verification is REALLY needed.
8617 BOOL tiVerificationNeeded;
8619 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8620 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8621 BOOL tiIsVerifiableCode;
8623 // Set to TRUE if runtime callout is needed for this method
8624 BOOL tiRuntimeCalloutNeeded;
8626 // Set to TRUE if security prolog/epilog callout is needed for this method
8627 // Note: This flag is different than compNeedSecurityCheck.
8628 // compNeedSecurityCheck means whether or not a security object needs
8629 // to be allocated on the stack, which is currently true for EnC as well.
8630 // tiSecurityCalloutNeeded means whether or not security callouts need
8631 // to be inserted in the jitted code.
8632 BOOL tiSecurityCalloutNeeded;
8634 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8635 // This support is necessary to suport attributes that are not described in
8636 // for example, signatures. For example, the permanent home byref (byref that
8637 // points to the gc heap), isn't a property of method signatures, therefore,
8638 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8639 // but when deciding if we need to reimport a block, we need to take these
8641 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8643 // Returns TRUE if child is equal to or a subtype of parent.
8644 // normalisedForStack indicates that both types are normalised for the stack
8645 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8647 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8648 // *pDest is modified to represent the merged type. Sets "*changed" to true
8649 // if this changes "*pDest".
8650 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8652 // Set pDest from the primitive value type.
8653 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8655 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8658 // <BUGNUM> VSW 471305
8659 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8660 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8661 // We use a "short" as we need to push/pop this scope.
8663 short compRegSetCheckLevel;
8667 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8668 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8670 XX IL verification stuff XX
8673 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8674 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8678 // The following is used to track liveness of local variables, initialization
8679 // of valueclass constructors, and type safe use of IL instructions.
8681 // dynamic state info needed for verification
8682 EntryState verCurrentState;
8684 // this ptr of object type .ctors are considered intited only after
8685 // the base class ctor is called, or an alternate ctor is called.
8686 // An uninited this ptr can be used to access fields, but cannot
8687 // be used to call a member function.
8688 BOOL verTrackObjCtorInitState;
8690 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8692 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8693 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8694 void verInitCurrentState();
8695 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8697 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8698 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8699 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8701 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8702 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8703 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8704 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8705 typeInfo verMakeTypeInfo(CorInfoType ciType,
8706 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8707 BOOL verIsSDArray(typeInfo ti);
8708 typeInfo verGetArrayElemType(typeInfo ti);
8710 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8711 BOOL verNeedsVerification();
8712 BOOL verIsByRefLike(const typeInfo& ti);
8713 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8715 // generic type variables range over types that satisfy IsBoxable
8716 BOOL verIsBoxable(const typeInfo& ti);
8718 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8719 DEBUGARG(unsigned line));
8720 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8721 DEBUGARG(unsigned line));
8722 bool verCheckTailCallConstraint(OPCODE opcode,
8723 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8724 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8725 // on a type parameter?
8726 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8727 // return false to the caller.
8728 // If false, it will throw.
8730 bool verIsBoxedValueType(typeInfo ti);
8732 void verVerifyCall(OPCODE opcode,
8733 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8734 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8736 bool readonlyCall, // is this a "readonly." call?
8737 const BYTE* delegateCreateStart,
8738 const BYTE* codeAddr,
8739 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8741 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8743 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8744 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8745 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8746 const CORINFO_FIELD_INFO& fieldInfo,
8747 const typeInfo* tiThis,
8749 BOOL allowPlainStructAsThis = FALSE);
8750 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8751 void verVerifyThisPtrInitialised();
8752 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8754 // Register allocator
8755 void raInitStackFP();
8756 void raEnregisterVarsPrePassStackFP();
8757 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8758 void raEnregisterVarsPostPassStackFP();
8759 void raGenerateFPRefCounts();
8760 void raEnregisterVarsStackFP();
8761 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8763 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8764 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8766 // returns true if enregistering v1 would save more mem accesses than v2
8767 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8770 void raDumpHeightsStackFP();
8771 void raDumpVariableRegIntfFloat();
8774 #if FEATURE_STACK_FP_X87
8776 // Currently, we use FP transition blocks in only 2 situations:
8778 // -conditional jump on longs where FP stack differs with target: it's not strictly
8779 // necessary, but its low frequency and the code would get complicated if we try to
8780 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8781 // minimize the way we generate the jump code.
8782 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8783 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8785 // However, transition blocks have 2 problems
8787 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8788 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8789 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8790 // in the right place without preordering them), this causes us to have to generate the transition
8791 // blocks in the cold area if we want procedure splitting.
8794 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8795 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8796 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8797 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8798 // a big change in the exception.
8800 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8801 // optimizations. For these 2 cases:
8803 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8804 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8805 // a switch statement.
8807 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8808 // current procedure splitting and exception code have.
8809 bool compMayHaveTransitionBlocks;
8811 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8813 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8815 unsigned raCntStkStackFP;
8816 unsigned raCntWtdStkDblStackFP;
8817 unsigned raCntStkParamDblStackFP;
8819 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8820 // TODO: Do we want to put this in LclVarDsc?
8821 unsigned raPayloadStackFP[lclMAX_TRACKED];
8822 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8824 // Useful for debugging
8825 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8827 #endif // FEATURE_STACK_FP_X87
8830 // One line log function. Default level is 0. Increasing it gives you
8831 // more log information
8833 // levels are currently unused: #define JITDUMP(level,...) ();
8834 void JitLogEE(unsigned level, const char* fmt, ...);
8836 bool compDebugBreak;
8838 bool compJitHaltMethod();
8843 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8844 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8846 XX GS Security checks for unsafe buffers XX
8848 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8849 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8852 struct ShadowParamVarInfo
8854 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8855 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8857 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8859 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8860 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8861 // slots and update all trees to refer to shadow slots is done immediately after
8862 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8863 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8864 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8865 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8866 // creating a shadow slot even though this routine returns true.
8868 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8869 // required. There are two cases under which a reg arg could potentially be used from its
8871 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8872 // b) LSRA spills it
8874 // Possible solution to address case (a)
8875 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8876 // in this routine. Note that live out of exception handler is something we may not be
8877 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8878 // Therefore, for methods with exception handling and need GS cookie check we might have
8879 // to take conservative approach.
8881 // Possible solution to address case (b)
8882 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8883 // create a new spill temp if the method needs GS cookie check.
8884 return varDsc->lvIsParam;
8885 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8886 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8893 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8898 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8899 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8900 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8902 void gsGSChecksInitCookie(); // Grabs cookie variable
8903 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8904 bool gsFindVulnerableParams(); // Shadow param analysis code
8905 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8907 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8908 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8910 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8911 // This can be overwritten by setting complus_JITInlineSize env variable.
8913 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8916 #ifdef FEATURE_JIT_METHOD_PERF
8917 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8918 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8920 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8921 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8923 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8925 #if MEASURE_CLRAPI_CALLS
8926 // Thin wrappers that call into JitTimer (if present).
8927 inline void CLRApiCallEnter(unsigned apix);
8928 inline void CLRApiCallLeave(unsigned apix);
8931 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8932 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8937 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8938 // These variables are associated with maintaining SQM data about compile time.
8939 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8940 // in the current compilation.
8941 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8942 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8943 // the inlining phase in the current compilation.
8944 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8946 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8947 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8948 // type-loading and class initialization).
8949 void RecordStateAtEndOfInlining();
8950 // Assumes being called at the end of compilation. Update the SQM state.
8951 void RecordStateAtEndOfCompilation();
8953 #ifdef FEATURE_CLRSQM
8954 // Does anything SQM related necessary at process shutdown time.
8955 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8956 #endif // FEATURE_CLRSQM
8959 #if FUNC_INFO_LOGGING
8960 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8961 // filename to write it to.
8962 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8963 #endif // FUNC_INFO_LOGGING
8965 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8967 // Is the compilation in a full trust context?
8968 bool compIsFullTrust();
8970 #ifndef FEATURE_TRACELOGGING
8971 // Should we actually fire the noway assert body and the exception handler?
8972 bool compShouldThrowOnNoway();
8973 #else // FEATURE_TRACELOGGING
8974 // Should we actually fire the noway assert body and the exception handler?
8975 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8977 // Telemetry instance to use per method compilation.
8978 JitTelemetry compJitTelemetry;
8980 // Get common parameters that have to be logged with most telemetry data.
8981 void compGetTelemetryDefaults(const char** assemblyName,
8982 const char** scopeName,
8983 const char** methodName,
8984 unsigned* methodHash);
8985 #endif // !FEATURE_TRACELOGGING
8989 NodeToTestDataMap* m_nodeTestData;
8991 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8992 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8993 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8994 // Current kept in this.
8996 NodeToTestDataMap* GetNodeTestData()
8998 Compiler* compRoot = impInlineRoot();
8999 if (compRoot->m_nodeTestData == nullptr)
9001 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9003 return compRoot->m_nodeTestData;
9006 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
9008 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9009 // currently occur in the AST graph.
9010 NodeToIntMap* FindReachableNodesInNodeTestData();
9012 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9013 // test data, associate that data with "to".
9014 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
9016 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9017 // have annotations, attach similar annotations to the corresponding nodes in "to".
9018 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
9020 // These are the methods that test that the various conditions implied by the
9021 // test attributes are satisfied.
9022 void JitTestCheckSSA(); // SSA builder tests.
9023 void JitTestCheckVN(); // Value numbering tests.
9026 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9028 FieldSeqStore* m_fieldSeqStore;
9030 FieldSeqStore* GetFieldSeqStore()
9032 Compiler* compRoot = impInlineRoot();
9033 if (compRoot->m_fieldSeqStore == nullptr)
9035 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9036 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
9037 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9039 return compRoot->m_fieldSeqStore;
9042 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
9044 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9045 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9046 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9047 // attach the field sequence directly to the address node.
9048 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9050 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9052 // Don't need to worry about inlining here
9053 if (m_zeroOffsetFieldMap == nullptr)
9055 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9057 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9058 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9060 return m_zeroOffsetFieldMap;
9063 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9064 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9065 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9066 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9067 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9068 // record the the field sequence using the ZeroOffsetFieldMap described above.
9070 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9071 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9072 // CoreRT. Such case is handled same as the default case.
9073 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9075 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9077 NodeToArrayInfoMap* m_arrayInfoMap;
9079 NodeToArrayInfoMap* GetArrayInfoMap()
9081 Compiler* compRoot = impInlineRoot();
9082 if (compRoot->m_arrayInfoMap == nullptr)
9084 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9085 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9086 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9088 return compRoot->m_arrayInfoMap;
9091 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9093 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9094 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9095 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9096 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9098 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9100 // Use the same map for GCHeap and ByrefExposed when their states match.
9101 memoryKind = ByrefExposed;
9104 assert(memoryKind < MemoryKindCount);
9105 Compiler* compRoot = impInlineRoot();
9106 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9108 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9109 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9110 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9112 return compRoot->m_memorySsaMap[memoryKind];
9115 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9116 CORINFO_CLASS_HANDLE m_refAnyClass;
9117 CORINFO_FIELD_HANDLE GetRefanyDataField()
9119 if (m_refAnyClass == nullptr)
9121 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9123 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9125 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9127 if (m_refAnyClass == nullptr)
9129 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9131 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9135 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9137 #if ALLVARSET_COUNTOPS
9138 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9141 static HelperCallProperties s_helperCallProperties;
9143 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9144 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9145 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9147 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9150 unsigned __int8* offset0,
9151 unsigned __int8* offset1);
9152 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9153 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9155 void fgMorphMultiregStructArgs(GenTreeCall* call);
9156 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9158 }; // end of class Compiler
9160 // Inline methods of CompAllocator.
9161 void* CompAllocator::Alloc(size_t sz)
9163 #if MEASURE_MEM_ALLOC
9164 return m_comp->compGetMem(sz, m_cmk);
9166 return m_comp->compGetMem(sz);
9170 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9172 #if MEASURE_MEM_ALLOC
9173 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9175 return m_comp->compGetMemArray(elems, elemSize);
9179 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9180 inline LclVarDsc::LclVarDsc(Compiler* comp)
9181 : // Initialize the ArgRegs to REG_STK.
9182 // The morph will do the right thing to change
9183 // to the right register if passed in register.
9186 #if FEATURE_MULTIREG_ARGS
9187 _lvOtherArgReg(REG_STK)
9189 #endif // FEATURE_MULTIREG_ARGS
9191 lvRefBlks(BlockSetOps::UninitVal())
9193 #endif // ASSERTION_PROP
9194 lvPerSsaData(comp->getAllocator())
9199 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9200 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9202 XX Miscellaneous Compiler stuff XX
9204 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9205 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9208 // Values used to mark the types a stack slot is used for
9210 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9211 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9212 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9213 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9214 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9215 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9216 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9217 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9219 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9221 /*****************************************************************************
9223 * Variables to keep track of total code amounts.
9228 extern size_t grossVMsize;
9229 extern size_t grossNCsize;
9230 extern size_t totalNCsize;
9232 extern unsigned genMethodICnt;
9233 extern unsigned genMethodNCnt;
9234 extern size_t gcHeaderISize;
9235 extern size_t gcPtrMapISize;
9236 extern size_t gcHeaderNSize;
9237 extern size_t gcPtrMapNSize;
9239 #endif // DISPLAY_SIZES
9241 /*****************************************************************************
9243 * Variables to keep track of basic block counts (more data on 1 BB methods)
9246 #if COUNT_BASIC_BLOCKS
9247 extern Histogram bbCntTable;
9248 extern Histogram bbOneBBSizeTable;
9251 /*****************************************************************************
9253 * Used by optFindNaturalLoops to gather statistical information such as
9254 * - total number of natural loops
9255 * - number of loops with 1, 2, ... exit conditions
9256 * - number of loops that have an iterator (for like)
9257 * - number of loops that have a constant iterator
9262 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9263 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9264 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9265 extern unsigned totalLoopCount; // counts the total number of natural loops
9266 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9267 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9268 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9269 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9271 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9272 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9273 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9274 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9275 extern Histogram loopCountTable; // Histogram of loop counts
9276 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9278 #endif // COUNT_LOOPS
9280 /*****************************************************************************
9281 * variables to keep track of how many iterations we go in a dataflow pass
9286 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9287 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9289 #endif // DATAFLOW_ITER
9291 #if MEASURE_BLOCK_SIZE
9292 extern size_t genFlowNodeSize;
9293 extern size_t genFlowNodeCnt;
9294 #endif // MEASURE_BLOCK_SIZE
9296 #if MEASURE_NODE_SIZE
9297 struct NodeSizeStats
9302 genTreeNodeSize = 0;
9303 genTreeNodeActualSize = 0;
9306 size_t genTreeNodeCnt;
9307 size_t genTreeNodeSize; // The size we allocate
9308 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9309 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9310 // a smaller node to a larger one. TODO-Cleanup: add stats on
9311 // SetOper()/ChangeOper() usage to quanitfy this.
9313 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9314 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9315 extern Histogram genTreeNcntHist;
9316 extern Histogram genTreeNsizHist;
9317 #endif // MEASURE_NODE_SIZE
9319 /*****************************************************************************
9320 * Count fatal errors (including noway_asserts).
9324 extern unsigned fatal_badCode;
9325 extern unsigned fatal_noWay;
9326 extern unsigned fatal_NOMEM;
9327 extern unsigned fatal_noWayAssertBody;
9329 extern unsigned fatal_noWayAssertBodyArgs;
9331 extern unsigned fatal_NYI;
9332 #endif // MEASURE_FATAL
9334 /*****************************************************************************
9338 #ifdef _TARGET_XARCH_
9340 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9341 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9342 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9344 const instruction INS_AND = INS_and;
9345 const instruction INS_OR = INS_or;
9346 const instruction INS_XOR = INS_xor;
9347 const instruction INS_NEG = INS_neg;
9348 const instruction INS_TEST = INS_test;
9349 const instruction INS_MUL = INS_imul;
9350 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9351 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9352 const instruction INS_BREAKPOINT = INS_int3;
9353 const instruction INS_ADDC = INS_adc;
9354 const instruction INS_SUBC = INS_sbb;
9355 const instruction INS_NOT = INS_not;
9361 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9362 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9363 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9365 const instruction INS_AND = INS_and;
9366 const instruction INS_OR = INS_orr;
9367 const instruction INS_XOR = INS_eor;
9368 const instruction INS_NEG = INS_rsb;
9369 const instruction INS_TEST = INS_tst;
9370 const instruction INS_MUL = INS_mul;
9371 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9372 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9373 const instruction INS_BREAKPOINT = INS_bkpt;
9374 const instruction INS_ADDC = INS_adc;
9375 const instruction INS_SUBC = INS_sbc;
9376 const instruction INS_NOT = INS_mvn;
9380 #ifdef _TARGET_ARM64_
9382 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9383 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9384 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9386 const instruction INS_AND = INS_and;
9387 const instruction INS_OR = INS_orr;
9388 const instruction INS_XOR = INS_eor;
9389 const instruction INS_NEG = INS_neg;
9390 const instruction INS_TEST = INS_tst;
9391 const instruction INS_MUL = INS_mul;
9392 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9393 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9394 const instruction INS_BREAKPOINT = INS_bkpt;
9395 const instruction INS_ADDC = INS_adc;
9396 const instruction INS_SUBC = INS_sbc;
9397 const instruction INS_NOT = INS_mvn;
9401 /*****************************************************************************/
9403 extern const BYTE genTypeSizes[];
9404 extern const BYTE genTypeAlignments[];
9405 extern const BYTE genTypeStSzs[];
9406 extern const BYTE genActualTypes[];
9408 /*****************************************************************************/
9410 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9411 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9414 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9415 #elif defined(_TARGET_ARM64_)
9416 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9419 /*****************************************************************************/
9421 #define REG_CORRUPT regNumber(REG_NA + 1)
9422 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9423 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9425 /*****************************************************************************/
9427 extern BasicBlock dummyBB;
9429 /*****************************************************************************/
9430 /*****************************************************************************/
9432 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9433 // nodes of a statement in execution order.
9434 // __stmt: a GT_STMT type GenTree*
9435 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9437 #define foreach_treenode_execution_order(__node, __stmt) \
9438 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9440 // foreach_block: An iterator over all blocks in the function.
9441 // __compiler: the Compiler* object
9442 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9444 #define foreach_block(__compiler, __block) \
9445 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9447 /*****************************************************************************/
9448 /*****************************************************************************/
9452 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9454 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9455 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9457 XX Debugging helpers XX
9459 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9460 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9463 /*****************************************************************************/
9464 /* The following functions are intended to be called from the debugger, to dump
9465 * various data structures. The can be used in the debugger Watch or Quick Watch
9466 * windows. They are designed to be short to type and take as few arguments as
9467 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9468 * See the function definition comment for more details.
9471 void cBlock(Compiler* comp, BasicBlock* block);
9472 void cBlocks(Compiler* comp);
9473 void cBlocksV(Compiler* comp);
9474 void cTree(Compiler* comp, GenTree* tree);
9475 void cTrees(Compiler* comp);
9476 void cEH(Compiler* comp);
9477 void cVar(Compiler* comp, unsigned lclNum);
9478 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9479 void cVars(Compiler* comp);
9480 void cVarsFinal(Compiler* comp);
9481 void cBlockPreds(Compiler* comp, BasicBlock* block);
9482 void cReach(Compiler* comp);
9483 void cDoms(Compiler* comp);
9484 void cLiveness(Compiler* comp);
9485 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9487 void cFuncIR(Compiler* comp);
9488 void cBlockIR(Compiler* comp, BasicBlock* block);
9489 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9490 void cTreeIR(Compiler* comp, GenTree* tree);
9491 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9492 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9493 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9494 int cOperandIR(Compiler* comp, GenTree* operand);
9495 int cLeafIR(Compiler* comp, GenTree* tree);
9496 int cIndirIR(Compiler* comp, GenTree* tree);
9497 int cListIR(Compiler* comp, GenTree* list);
9498 int cSsaNumIR(Compiler* comp, GenTree* tree);
9499 int cValNumIR(Compiler* comp, GenTree* tree);
9500 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9502 void dBlock(BasicBlock* block);
9505 void dTree(GenTree* tree);
9508 void dVar(unsigned lclNum);
9509 void dVarDsc(LclVarDsc* varDsc);
9512 void dBlockPreds(BasicBlock* block);
9516 void dCVarSet(VARSET_VALARG_TP vars);
9518 void dVarSet(VARSET_VALARG_TP vars);
9519 void dRegMask(regMaskTP mask);
9522 void dBlockIR(BasicBlock* block);
9523 void dTreeIR(GenTree* tree);
9524 void dLoopIR(Compiler::LoopDsc* loop);
9525 void dLoopNumIR(unsigned loopNum);
9526 int dTabStopIR(int curr, int tabstop);
9527 int dTreeTypeIR(GenTree* tree);
9528 int dTreeKindsIR(GenTree* tree);
9529 int dTreeFlagsIR(GenTree* tree);
9530 int dOperandIR(GenTree* operand);
9531 int dLeafIR(GenTree* tree);
9532 int dIndirIR(GenTree* tree);
9533 int dListIR(GenTree* list);
9534 int dSsaNumIR(GenTree* tree);
9535 int dValNumIR(GenTree* tree);
9536 int dDependsIR(GenTree* comma);
9539 GenTree* dFindTree(GenTree* tree, unsigned id);
9540 GenTree* dFindTree(unsigned id);
9541 GenTreeStmt* dFindStmt(unsigned id);
9542 BasicBlock* dFindBlock(unsigned bbNum);
9546 #include "compiler.hpp" // All the shared inline functions
9548 /*****************************************************************************/
9549 #endif //_COMPILER_H_
9550 /*****************************************************************************/