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 bool gtIsRecursiveCall(GenTreeCall* call)
2143 return (call->gtCallMethHnd == info.compMethodHnd);
2146 //-------------------------------------------------------------------------
2148 GenTreePtr gtFoldExpr(GenTreePtr tree);
2151 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2152 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2153 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2154 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2155 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2156 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2157 // optimizations for now.
2158 __attribute__((optnone))
2160 gtFoldExprConst(GenTreePtr tree);
2161 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2162 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2164 //-------------------------------------------------------------------------
2165 // Get the handle, if any.
2166 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2167 // Get the handle, and assert if not found.
2168 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2170 //-------------------------------------------------------------------------
2171 // Functions to display the trees
2174 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2176 void gtDispVN(GenTreePtr tree);
2177 void gtDispConst(GenTreePtr tree);
2178 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2179 void gtDispNodeName(GenTreePtr tree);
2180 void gtDispRegVal(GenTreePtr tree);
2192 void gtDispChild(GenTreePtr child,
2193 IndentStack* indentStack,
2195 __in_opt const char* msg = nullptr,
2196 bool topOnly = false);
2197 void gtDispTree(GenTreePtr tree,
2198 IndentStack* indentStack = nullptr,
2199 __in_opt const char* msg = nullptr,
2200 bool topOnly = false,
2201 bool isLIR = false);
2202 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2203 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2204 char* gtGetLclVarName(unsigned lclNum);
2205 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2206 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2207 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2208 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2209 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2210 void gtDispFieldSeq(FieldSeqNode* pfsn);
2212 void gtDispRange(LIR::ReadOnlyRange const& range);
2214 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2216 void gtDispLIRNode(GenTree* node);
2228 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2229 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2232 static fgWalkPreFn gtAssertColonCond;
2234 static fgWalkPreFn gtMarkColonCond;
2235 static fgWalkPreFn gtClearColonCond;
2237 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2238 bool gtHasCatchArg(GenTreePtr tree);
2239 bool gtHasUnmanagedCall(GenTreePtr tree);
2241 typedef ArrayStack<GenTree*> GenTreeStack;
2243 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2244 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2246 //=========================================================================
2247 // BasicBlock functions
2249 // This is a debug flag we will use to assert when creating block during codegen
2250 // as this interferes with procedure splitting. If you know what you're doing, set
2251 // it to true before creating the block. (DEBUG only)
2252 bool fgSafeBasicBlockCreation;
2255 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2258 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2259 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2263 XX The variables to be used by the code generator. XX
2265 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2266 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2270 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2271 // be placed in the stack frame and it's fields must be laid out sequentially.
2273 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2274 // a local variable that can be enregistered or placed in the stack frame.
2275 // The fields do not need to be laid out sequentially
2277 enum lvaPromotionType
2279 PROMOTION_TYPE_NONE, // The struct local is not promoted
2280 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2281 // and its field locals are independent of its parent struct local.
2282 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2283 // but its field locals depend on its parent struct local.
2286 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2287 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2289 /*****************************************************************************/
2291 enum FrameLayoutState
2294 INITIAL_FRAME_LAYOUT,
2295 PRE_REGALLOC_FRAME_LAYOUT,
2296 REGALLOC_FRAME_LAYOUT,
2297 TENTATIVE_FRAME_LAYOUT,
2302 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2303 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2304 bool lvaSortAgain; // true: We need to sort the lvaTable
2305 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2306 unsigned lvaCount; // total number of locals
2308 unsigned lvaRefCount; // total number of references to locals
2309 LclVarDsc* lvaTable; // variable descriptor table
2310 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2312 LclVarDsc** lvaRefSorted; // table sorted by refcount
2314 unsigned short lvaTrackedCount; // actual # of locals being tracked
2315 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2317 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2318 // Only for AMD64 System V cache the first caller stack homed argument.
2319 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2320 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2323 VARSET_TP lvaTrackedVars; // set of tracked variables
2325 #ifndef _TARGET_64BIT_
2326 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2328 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2330 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2331 // It that changes, this changes. VarSets from different epochs
2332 // cannot be meaningfully combined.
2334 unsigned GetCurLVEpoch()
2339 // reverse map of tracked number to var number
2340 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2342 #ifdef LEGACY_BACKEND
2343 // variable interference graph
2344 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2347 // variable preference graph
2348 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2352 // # of procs compiled a with double-aligned stack
2353 static unsigned s_lvaDoubleAlignedProcsCount;
2357 // Getters and setters for address-exposed and do-not-enregister local var properties.
2358 bool lvaVarAddrExposed(unsigned varNum);
2359 void lvaSetVarAddrExposed(unsigned varNum);
2360 bool lvaVarDoNotEnregister(unsigned varNum);
2362 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2363 enum DoNotEnregisterReason
2368 DNER_VMNeedsStackAddr,
2369 DNER_LiveInOutOfHandler,
2370 DNER_LiveAcrossUnmanagedCall,
2371 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2372 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2373 #ifdef JIT32_GCENCODER
2378 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2380 unsigned lvaVarargsHandleArg;
2382 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2384 #endif // _TARGET_X86_
2386 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2387 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2388 #if FEATURE_FIXED_OUT_ARGS
2389 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2391 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2392 // that tracks whether the lock has been taken
2394 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2395 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2396 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2398 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2399 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2401 #if FEATURE_FIXED_OUT_ARGS
2402 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2403 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2404 #endif // FEATURE_FIXED_OUT_ARGS
2407 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2408 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2409 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2410 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2411 // this variable to be this scratch word whenever struct promotion occurs.
2412 unsigned lvaPromotedStructAssemblyScratchVar;
2413 #endif // _TARGET_ARM_
2416 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2417 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2420 bool lvaGenericsContextUsed;
2422 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2423 // CORINFO_GENERICS_CTXT_FROM_THIS?
2424 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2426 //-------------------------------------------------------------------------
2427 // All these frame offsets are inter-related and must be kept in sync
2429 #if !FEATURE_EH_FUNCLETS
2430 // This is used for the callable handlers
2431 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2432 #endif // FEATURE_EH_FUNCLETS
2434 unsigned lvaCachedGenericContextArgOffs;
2435 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2438 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2440 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2442 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2443 // after the reg predict we will use a computed maxTmpSize
2444 // which is based upon the number of spill temps predicted by reg predict
2445 // All this is necessary because if we under-estimate the size of the spill
2446 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2448 // Pre codegen max spill temp size.
2449 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2451 //-------------------------------------------------------------------------
2453 unsigned lvaGetMaxSpillTempSize();
2455 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2456 #endif // _TARGET_ARM_
2457 void lvaAssignFrameOffsets(FrameLayoutState curState);
2458 void lvaFixVirtualFrameOffsets();
2460 #ifndef LEGACY_BACKEND
2461 void lvaUpdateArgsWithInitialReg();
2462 #endif // !LEGACY_BACKEND
2464 void lvaAssignVirtualFrameOffsetsToArgs();
2465 #ifdef UNIX_AMD64_ABI
2466 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2467 #else // !UNIX_AMD64_ABI
2468 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2469 #endif // !UNIX_AMD64_ABI
2470 void lvaAssignVirtualFrameOffsetsToLocals();
2471 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2472 #ifdef _TARGET_AMD64_
2473 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2474 bool lvaIsCalleeSavedIntRegCountEven();
2476 void lvaAlignFrame();
2477 void lvaAssignFrameOffsetsToPromotedStructs();
2478 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2481 void lvaDumpRegLocation(unsigned lclNum);
2482 void lvaDumpFrameLocation(unsigned lclNum);
2483 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2484 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2485 // layout state defined by lvaDoneFrameLayout
2488 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2489 // to avoid bugs from borderline cases.
2490 #define MAX_FrameSize 0x3FFFFFFF
2491 void lvaIncrementFrameSize(unsigned size);
2493 unsigned lvaFrameSize(FrameLayoutState curState);
2495 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2496 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2498 // Returns the caller-SP-relative offset for the local variable "varNum."
2499 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2501 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2502 int lvaGetSPRelativeOffset(unsigned varNum);
2504 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2505 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2507 //------------------------ For splitting types ----------------------------
2509 void lvaInitTypeRef();
2511 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2512 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2513 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2514 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2515 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2516 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2518 void lvaInitVarDsc(LclVarDsc* varDsc,
2520 CorInfoType corInfoType,
2521 CORINFO_CLASS_HANDLE typeHnd,
2522 CORINFO_ARG_LIST_HANDLE varList,
2523 CORINFO_SIG_INFO* varSig);
2525 static unsigned lvaTypeRefMask(var_types type);
2527 var_types lvaGetActualType(unsigned lclNum);
2528 var_types lvaGetRealType(unsigned lclNum);
2530 //-------------------------------------------------------------------------
2534 unsigned lvaLclSize(unsigned varNum);
2535 unsigned lvaLclExactSize(unsigned varNum);
2537 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2539 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2540 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2541 // the return result.
2542 bool lvaLclVarRefsAccum(
2543 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2545 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2546 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2547 // and (destructively) unions "trkedVars" into "*result".
2548 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2550 ALLVARSET_VALARG_TP allVars,
2551 VARSET_VALARG_TP trkdVars);
2553 bool lvaHaveManyLocals() const;
2555 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2556 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2557 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2560 void lvaSortByRefCount();
2561 void lvaDumpRefCounts();
2563 void lvaMarkLocalVars(BasicBlock* block);
2565 void lvaMarkLocalVars(); // Local variable ref-counting
2567 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2569 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2571 static fgWalkPreFn lvaIncRefCntsCB;
2572 void lvaIncRefCnts(GenTreePtr tree);
2574 static fgWalkPreFn lvaDecRefCntsCB;
2575 void lvaDecRefCnts(GenTreePtr tree);
2576 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2577 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2578 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2581 struct lvaStressLclFldArgs
2583 Compiler* m_pCompiler;
2587 static fgWalkPreFn lvaStressLclFldCB;
2588 void lvaStressLclFld();
2590 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2591 void lvaDispVarSet(VARSET_VALARG_TP set);
2596 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2598 int lvaFrameAddress(int varNum, bool* pFPbased);
2601 bool lvaIsParameter(unsigned varNum);
2602 bool lvaIsRegArgument(unsigned varNum);
2603 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2604 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2605 // that writes to arg0
2607 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2608 // (this is an overload of lvIsTemp because there are no temp parameters).
2609 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2610 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2611 bool lvaIsImplicitByRefLocal(unsigned varNum)
2613 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2614 LclVarDsc* varDsc = &(lvaTable[varNum]);
2615 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2617 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2620 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2624 // Returns true if this local var is a multireg struct
2625 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2627 // If the class is a TYP_STRUCT, get/set a class handle describing it
2629 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2630 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2632 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2634 // Info about struct fields
2635 struct lvaStructFieldInfo
2637 CORINFO_FIELD_HANDLE fldHnd;
2638 unsigned char fldOffset;
2639 unsigned char fldOrdinal;
2642 CORINFO_CLASS_HANDLE fldTypeHnd;
2645 // Info about struct to be promoted.
2646 struct lvaStructPromotionInfo
2648 CORINFO_CLASS_HANDLE typeHnd;
2650 bool requiresScratchVar;
2653 unsigned char fieldCnt;
2654 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2656 lvaStructPromotionInfo()
2657 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2662 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2663 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2664 lvaStructPromotionInfo* StructPromotionInfo,
2666 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2667 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2668 #if !defined(_TARGET_64BIT_)
2669 void lvaPromoteLongVars();
2670 #endif // !defined(_TARGET_64BIT_)
2671 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2672 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2673 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2674 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2675 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2676 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2677 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2679 #if defined(FEATURE_SIMD)
2680 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
2682 assert(varDsc->lvType == TYP_SIMD12);
2683 assert(varDsc->lvExactSize == 12);
2685 #if defined(_TARGET_64BIT_)
2686 assert(varDsc->lvSize() == 16);
2688 #else // !defined(_TARGET_64BIT_)
2690 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
2691 // already does this calculation. However, we also need to prevent mapping types if the var is a
2692 // depenendently promoted struct field, which must remain its exact size within its parent struct.
2693 // However, we don't know this until late, so we may have already pretended the field is bigger
2695 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
2704 #endif // !defined(_TARGET_64BIT_)
2706 #endif // defined(FEATURE_SIMD)
2708 BYTE* lvaGetGcLayout(unsigned varNum);
2709 bool lvaTypeIsGC(unsigned varNum);
2710 unsigned lvaGSSecurityCookie; // LclVar number
2711 bool lvaTempsHaveLargerOffsetThanVars();
2713 unsigned lvaSecurityObject; // variable representing the security object on the stack
2714 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2716 #if FEATURE_EH_FUNCLETS
2717 unsigned lvaPSPSym; // variable representing the PSPSym
2720 InlineInfo* impInlineInfo;
2721 InlineStrategy* m_inlineStrategy;
2723 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2724 Compiler* impInlineRoot();
2726 #if defined(DEBUG) || defined(INLINE_DATA)
2727 unsigned __int64 getInlineCycleCount()
2729 return m_compCycles;
2731 #endif // defined(DEBUG) || defined(INLINE_DATA)
2733 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2734 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2736 //=========================================================================
2738 //=========================================================================
2741 //---------------- Local variable ref-counting ----------------------------
2744 BasicBlock* lvaMarkRefsCurBlock;
2745 GenTreePtr lvaMarkRefsCurStmt;
2747 BasicBlock::weight_t lvaMarkRefsWeight;
2749 static fgWalkPreFn lvaMarkLclRefsCallback;
2750 void lvaMarkLclRefs(GenTreePtr tree);
2752 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
2753 PerSsaArray lvMemoryPerSsaData;
2754 unsigned lvMemoryNumSsaNames;
2757 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
2758 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2759 // not an SSA variable).
2760 LclSsaVarDsc* GetMemoryPerSsaData(unsigned ssaNum)
2762 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2763 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2765 assert(ssaNum < lvMemoryNumSsaNames);
2766 return &lvMemoryPerSsaData.GetRef(ssaNum);
2770 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2771 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2775 XX Imports the given method and converts it to semantic trees XX
2777 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2778 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2784 void impImport(BasicBlock* method);
2786 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2787 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2788 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2789 CORINFO_CLASS_HANDLE impGetStringClass();
2790 CORINFO_CLASS_HANDLE impGetObjectClass();
2792 //=========================================================================
2794 //=========================================================================
2797 //-------------------- Stack manipulation ---------------------------------
2799 unsigned impStkSize; // Size of the full stack
2801 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2803 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2805 struct SavedStack // used to save/restore stack contents.
2807 unsigned ssDepth; // number of values on stack
2808 StackEntry* ssTrees; // saved tree values
2811 bool impIsPrimitive(CorInfoType type);
2812 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2814 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2815 void impPushOnStackNoType(GenTreePtr tree);
2817 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2818 void impPushNullObjRefOnStack();
2819 StackEntry impPopStack();
2820 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2821 GenTreePtr impPopStack(typeInfo& ti);
2822 StackEntry& impStackTop(unsigned n = 0);
2824 void impSaveStackState(SavedStack* savePtr, bool copy);
2825 void impRestoreStackState(SavedStack* savePtr);
2827 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2828 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2829 CORINFO_CALL_INFO* pCallInfo);
2831 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2833 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2835 bool impCanPInvokeInline();
2836 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2837 void impCheckForPInvokeCall(
2838 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2839 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2840 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2842 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2843 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2844 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2846 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2847 CORINFO_METHOD_HANDLE calleeMethodHnd,
2848 CORINFO_CLASS_HANDLE delegateTypeHnd);
2850 var_types impImportCall(OPCODE opcode,
2851 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2852 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2854 GenTreePtr newobjThis,
2856 CORINFO_CALL_INFO* callInfo,
2857 IL_OFFSET rawILOffset);
2859 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2861 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2863 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2866 var_types impImportJitTestLabelMark(int numArgs);
2869 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2871 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2873 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2874 CORINFO_ACCESS_FLAGS access,
2875 CORINFO_FIELD_INFO* pFieldInfo,
2878 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2880 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2882 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2884 void impImportLeave(BasicBlock* block);
2885 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2886 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2887 CORINFO_CLASS_HANDLE clsHnd,
2888 CORINFO_METHOD_HANDLE method,
2889 CORINFO_SIG_INFO* sig,
2893 CorInfoIntrinsics* pIntrinsicID);
2894 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2895 CORINFO_SIG_INFO* sig,
2898 CorInfoIntrinsics intrinsicID);
2899 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2901 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2903 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2904 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2905 CORINFO_THIS_TRANSFORM transform);
2907 //----------------- Manipulating the trees and stmts ----------------------
2909 GenTreePtr impTreeList; // Trees for the BB being imported
2910 GenTreePtr impTreeLast; // The last tree for the current BB
2914 CHECK_SPILL_ALL = -1,
2915 CHECK_SPILL_NONE = -2
2919 void impBeginTreeList();
2920 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2921 void impEndTreeList(BasicBlock* block);
2922 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2923 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2924 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2925 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2926 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2927 void impAssignTempGen(unsigned tmp,
2930 GenTreePtr* pAfterStmt = nullptr,
2931 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2932 BasicBlock* block = nullptr);
2933 void impAssignTempGen(unsigned tmpNum,
2935 CORINFO_CLASS_HANDLE structHnd,
2937 GenTreePtr* pAfterStmt = nullptr,
2938 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2939 BasicBlock* block = nullptr);
2940 GenTreePtr impCloneExpr(GenTreePtr tree,
2942 CORINFO_CLASS_HANDLE structHnd,
2944 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2945 GenTreePtr impAssignStruct(GenTreePtr dest,
2947 CORINFO_CLASS_HANDLE structHnd,
2949 GenTreePtr* pAfterStmt = nullptr,
2950 BasicBlock* block = nullptr);
2951 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2953 CORINFO_CLASS_HANDLE structHnd,
2955 GenTreePtr* pAfterStmt = nullptr,
2956 BasicBlock* block = nullptr);
2958 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2959 CORINFO_CLASS_HANDLE structHnd,
2963 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2964 BYTE* gcLayout = nullptr,
2965 unsigned* numGCVars = nullptr,
2966 var_types* simdBaseType = nullptr);
2968 GenTreePtr impNormStructVal(GenTreePtr structVal,
2969 CORINFO_CLASS_HANDLE structHnd,
2971 bool forceNormalization = false);
2973 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2974 BOOL* pRuntimeLookup = nullptr,
2975 BOOL mustRestoreHandle = FALSE,
2976 BOOL importParent = FALSE);
2978 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2979 BOOL* pRuntimeLookup = nullptr,
2980 BOOL mustRestoreHandle = FALSE)
2982 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2985 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2986 CORINFO_LOOKUP* pLookup,
2988 void* compileTimeHandle);
2990 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2992 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2993 CORINFO_LOOKUP* pLookup,
2994 void* compileTimeHandle);
2996 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2998 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2999 CorInfoHelpFunc helper,
3001 GenTreeArgList* arg = nullptr,
3002 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3004 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
3006 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3009 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
3010 CORINFO_CLASS_HANDLE typeClass,
3014 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3015 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3016 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3017 static bool IsMathIntrinsic(GenTreePtr tree);
3020 //----------------- Importing the method ----------------------------------
3022 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3025 unsigned impCurOpcOffs;
3026 const char* impCurOpcName;
3027 bool impNestedStackSpill;
3029 // For displaying instrs with generated native code (-n:B)
3030 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3031 void impNoteLastILoffs();
3034 /* IL offset of the stmt currently being imported. It gets set to
3035 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3036 updated at IL offsets for which we have to report mapping info.
3037 It also includes flag bits, so use jitGetILoffs()
3038 to get the actual IL offset value.
3041 IL_OFFSETX impCurStmtOffs;
3042 void impCurStmtOffsSet(IL_OFFSET offs);
3044 void impNoteBranchOffs();
3046 unsigned impInitBlockLineInfo();
3048 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3049 bool impIsThis(GenTreePtr obj);
3050 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3051 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3052 bool impIsAnySTLOC(OPCODE opcode)
3054 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3055 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3058 GenTreeArgList* impPopList(unsigned count,
3060 CORINFO_SIG_INFO* sig,
3061 GenTreeArgList* prefixTree = nullptr);
3063 GenTreeArgList* impPopRevList(unsigned count,
3065 CORINFO_SIG_INFO* sig,
3066 unsigned skipReverseCount = 0);
3069 * Get current IL offset with stack-empty info incoporated
3071 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3073 //---------------- Spilling the importer stack ----------------------------
3079 SavedStack pdSavedStack;
3080 ThisInitState pdThisPtrInit;
3083 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3084 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3086 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3087 ExpandArray<BYTE> impPendingBlockMembers;
3089 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3090 // Operates on the map in the top-level ancestor.
3091 BYTE impGetPendingBlockMember(BasicBlock* blk)
3093 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3096 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3097 // Operates on the map in the top-level ancestor.
3098 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3100 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3103 bool impCanReimport;
3105 bool impSpillStackEntry(unsigned level,
3109 bool bAssertOnRecursion,
3114 void impSpillStackEnsure(bool spillLeaves = false);
3115 void impEvalSideEffects();
3116 void impSpillSpecialSideEff();
3117 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3118 void impSpillValueClasses();
3119 void impSpillEvalStack();
3120 static fgWalkPreFn impFindValueClasses;
3121 void impSpillLclRefs(ssize_t lclNum);
3123 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3125 void impImportBlockCode(BasicBlock* block);
3127 void impReimportMarkBlock(BasicBlock* block);
3128 void impReimportMarkSuccessors(BasicBlock* block);
3130 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3132 void impImportBlockPending(BasicBlock* block);
3134 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3135 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3136 // for the block, but instead, just re-uses the block's existing EntryState.
3137 void impReimportBlockPending(BasicBlock* block);
3139 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3141 void impImportBlock(BasicBlock* block);
3143 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3144 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3145 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3146 // the variables that will be used -- and for all the predecessors of those successors, and the
3147 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3148 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3149 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3150 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3151 // of local variable numbers, so we represent them with the base local variable number), returns that.
3152 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3153 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3154 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3155 // on which kind of member of the clique the block is).
3156 unsigned impGetSpillTmpBase(BasicBlock* block);
3158 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3159 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3160 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3161 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3162 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3163 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3164 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3165 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3166 // successors receive a native int. Similarly float and double are unified to double.
3167 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3168 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3169 // predecessors, so they insert an upcast if needed).
3170 void impReimportSpillClique(BasicBlock* block);
3172 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3173 // block, and represent the predecessor and successor members of the clique currently being computed.
3174 // *** Access to these will need to be locked in a parallel compiler.
3175 ExpandArray<BYTE> impSpillCliquePredMembers;
3176 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3184 // Abstract class for receiving a callback while walking a spill clique
3185 class SpillCliqueWalker
3188 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3191 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3192 class SetSpillTempsBase : public SpillCliqueWalker
3197 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3200 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3203 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3204 class ReimportSpillClique : public SpillCliqueWalker
3209 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3212 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3215 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3216 // predecessor or successor within the spill clique
3217 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3219 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3220 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3221 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3222 void impRetypeEntryStateTemps(BasicBlock* blk);
3224 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3225 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3227 void impPushVar(GenTree* op, typeInfo tiRetVal);
3228 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3229 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3231 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3233 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3234 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3235 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3238 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3241 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3242 struct BlockListNode
3245 BlockListNode* m_next;
3246 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3249 void* operator new(size_t sz, Compiler* comp);
3251 BlockListNode* impBlockListNodeFreeList;
3253 BlockListNode* AllocBlockListNode();
3254 void FreeBlockListNode(BlockListNode* node);
3256 bool impIsValueType(typeInfo* pTypeInfo);
3257 var_types mangleVarArgsType(var_types type);
3260 regNumber getCallArgIntRegister(regNumber floatReg);
3261 regNumber getCallArgFloatRegister(regNumber intReg);
3262 #endif // FEATURE_VARARG
3265 static unsigned jitTotalMethodCompiled;
3269 static LONG jitNestingLevel;
3272 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3274 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3276 // STATIC inlining decision based on the IL code.
3277 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3278 CORINFO_METHOD_INFO* methInfo,
3280 InlineResult* inlineResult);
3282 void impCheckCanInline(GenTreePtr call,
3283 CORINFO_METHOD_HANDLE fncHandle,
3285 CORINFO_CONTEXT_HANDLE exactContextHnd,
3286 InlineCandidateInfo** ppInlineCandidateInfo,
3287 InlineResult* inlineResult);
3289 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3290 GenTreePtr curArgVal,
3292 InlineResult* inlineResult);
3294 void impInlineInitVars(InlineInfo* pInlineInfo);
3296 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3298 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3300 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3302 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3303 GenTreePtr variableBeingDereferenced,
3304 InlArgInfo* inlArgInfo);
3306 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3308 bool impTailCallRetTypeCompatible(var_types callerRetType,
3309 CORINFO_CLASS_HANDLE callerRetTypeClass,
3310 var_types calleeRetType,
3311 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3313 bool impIsTailCallILPattern(bool tailPrefixed,
3315 const BYTE* codeAddrOfNextOpcode,
3316 const BYTE* codeEnd,
3318 bool* IsCallPopRet = nullptr);
3320 bool impIsImplicitTailCallCandidate(
3321 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3324 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3325 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3329 XX Info about the basic-blocks, their contents and the flow analysis XX
3331 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3332 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3336 BasicBlock* fgFirstBB; // Beginning of the basic block list
3337 BasicBlock* fgLastBB; // End of the basic block list
3338 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3339 #if FEATURE_EH_FUNCLETS
3340 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3342 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3344 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3345 unsigned fgEdgeCount; // # of control flow edges between the BBs
3346 unsigned fgBBcount; // # of BBs in the method
3348 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3350 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3351 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3352 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3353 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3355 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3356 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3357 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3358 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3359 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3360 // index). The arrays are of size fgBBNumMax + 1.
3361 unsigned* fgDomTreePreOrder;
3362 unsigned* fgDomTreePostOrder;
3364 bool fgBBVarSetsInited;
3366 // Allocate array like T* a = new T[fgBBNumMax + 1];
3367 // Using helper so we don't keep forgetting +1.
3368 template <typename T>
3369 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3371 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3374 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3375 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3376 // cannot be meaningfully combined. Note that new blocks can be created with higher
3377 // block numbers without changing the basic block epoch. These blocks *cannot*
3378 // participate in a block set until the blocks are all renumbered, causing the epoch
3379 // to change. This is useful if continuing to use previous block sets is valuable.
3380 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3381 unsigned fgCurBBEpoch;
3383 unsigned GetCurBasicBlockEpoch()
3385 return fgCurBBEpoch;
3388 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3389 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3390 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3391 unsigned fgCurBBEpochSize;
3393 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3394 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3395 unsigned fgBBSetCountInSizeTUnits;
3397 void NewBasicBlockEpoch()
3399 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3401 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3403 fgCurBBEpochSize = fgBBNumMax + 1;
3404 fgBBSetCountInSizeTUnits =
3405 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3408 // All BlockSet objects are now invalid!
3409 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3410 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3414 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3415 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3416 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3417 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3419 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3420 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3421 // array of size_t bitsets), then print that out.
3422 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3429 void EnsureBasicBlockEpoch()
3431 if (fgCurBBEpochSize != fgBBNumMax + 1)
3433 NewBasicBlockEpoch();
3437 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3438 void fgEnsureFirstBBisScratch();
3439 bool fgFirstBBisScratch();
3440 bool fgBBisScratch(BasicBlock* block);
3442 void fgExtendEHRegionBefore(BasicBlock* block);
3443 void fgExtendEHRegionAfter(BasicBlock* block);
3445 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3447 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3449 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3452 BasicBlock* nearBlk,
3453 bool putInFilter = false,
3454 bool runRarely = false,
3455 bool insertAtEnd = false);
3457 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3459 bool runRarely = false,
3460 bool insertAtEnd = false);
3462 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3464 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3465 BasicBlock* afterBlk,
3466 unsigned xcptnIndex,
3467 bool putInTryRegion);
3469 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3470 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3471 void fgUnlinkBlock(BasicBlock* block);
3473 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3474 bool fgMultipleNots;
3477 bool fgModified; // True if the flow graph has been modified recently
3478 bool fgComputePredsDone; // Have we computed the bbPreds list
3479 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3480 bool fgDomsComputed; // Have we computed the dominator sets?
3481 bool fgOptimizedFinally; // Did we optimize any try-finallys?
3483 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3484 bool fgHasPostfix; // any postfix ++/-- found?
3485 unsigned fgIncrCount; // number of increment nodes found
3487 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3491 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3492 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3495 bool fgRemoveRestOfBlock; // true if we know that we will throw
3496 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3498 // There are two modes for ordering of the trees.
3499 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3500 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3501 // by traversing the tree according to the order of the operands.
3502 // - In FGOrderLinear, the dominant ordering is the linear order.
3509 FlowGraphOrder fgOrder;
3511 // The following are boolean flags that keep track of the state of internal data structures
3513 bool fgStmtListThreaded;
3514 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3515 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3516 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3517 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3518 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3519 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3520 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3521 // This is derived from the profile data
3522 // or is BB_UNITY_WEIGHT when we don't have profile data
3524 #if FEATURE_EH_FUNCLETS
3525 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3526 #endif // FEATURE_EH_FUNCLETS
3528 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3529 // since fgMorphTree can be called from several places
3530 bool fgExpandInline; // indicates that we are creating tree for the inliner
3532 bool impBoxTempInUse; // the temp below is valid and available
3533 unsigned impBoxTemp; // a temporary that is used for boxing
3536 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3537 // and we are trying to compile again in a "safer", minopts mode?
3541 unsigned impInlinedCodeSize;
3544 //-------------------------------------------------------------------------
3552 void fgRemoveEmptyTry();
3554 void fgRemoveEmptyFinally();
3556 void fgCloneFinally();
3558 void fgCleanupContinuation(BasicBlock* continuation);
3560 void fgUpdateFinallyTargetFlags();
3562 GenTreePtr fgGetCritSectOfStaticMethod();
3564 #if !defined(_TARGET_X86_)
3566 void fgAddSyncMethodEnterExit();
3568 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3570 void fgConvertSyncReturnToLeave(BasicBlock* block);
3572 #endif // !_TARGET_X86_
3574 void fgAddReversePInvokeEnterExit();
3576 bool fgMoreThanOneReturnBlock();
3578 // The number of separate return points in the method.
3579 unsigned fgReturnCount;
3581 void fgAddInternal();
3583 bool fgFoldConditional(BasicBlock* block);
3585 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3586 void fgMorphBlocks();
3588 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3590 void fgCheckArgCnt();
3591 void fgSetOptions();
3594 static fgWalkPreFn fgAssertNoQmark;
3595 void fgPreExpandQmarkChecks(GenTreePtr expr);
3596 void fgPostExpandQmarkChecks();
3597 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3600 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3602 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3603 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3604 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3605 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3606 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3608 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3609 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3610 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3611 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3613 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3614 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3615 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3616 void fgExpandQmarkNodes();
3620 // Do "simple lowering." This functionality is (conceptually) part of "general"
3621 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3622 void fgSimpleLowering();
3624 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3626 GenTreePtr fgInitThisClass();
3628 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3630 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3632 void fgLocalVarLiveness();
3634 void fgLocalVarLivenessInit();
3636 #ifdef LEGACY_BACKEND
3637 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3639 void fgPerNodeLocalVarLiveness(GenTree* node);
3641 void fgPerBlockLocalVarLiveness();
3643 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3645 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3647 // This is used in the liveness computation, as a temporary. When we use the
3648 // arbitrary-length VarSet representation, it is better not to allocate a new one
3650 VARSET_TP fgMarkIntfUnionVS;
3652 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3654 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3656 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3658 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3660 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3662 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3664 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3665 GenTreePtr startNode,
3667 VARSET_VALARG_TP volatileVars,
3668 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3670 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3672 bool fgRemoveDeadStore(GenTree** pTree,
3676 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3678 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3680 // For updating liveset during traversal AFTER fgComputeLife has completed
3681 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3682 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3684 // Returns the set of live variables after endTree,
3685 // assuming that liveSet is the set of live variables BEFORE tree.
3686 // Requires that fgComputeLife has completed, and that tree is in the same
3687 // statement as endTree, and that it comes before endTree in execution order
3689 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3691 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3692 while (tree != nullptr && tree != endTree->gtNext)
3694 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3695 tree = tree->gtNext;
3697 assert(tree == endTree->gtNext);
3701 void fgInterBlockLocalVarLiveness();
3703 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3704 // "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
3705 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3706 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3707 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3708 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3709 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3711 if (m_opAsgnVarDefSsaNums == nullptr)
3713 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3715 return m_opAsgnVarDefSsaNums;
3718 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3719 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3720 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3722 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3724 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3725 // Except: assumes that lcl is a def, and if it is
3726 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3727 // rather than the "use" SSA number recorded in the tree "lcl".
3728 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3730 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3731 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3732 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3733 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3734 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3736 // (byref addrS1 = &s1,
3737 // *(addrS1 * offsetof(f0)) = s2f0,
3739 // *(addrS1 * offsetof(fn)) = s2fn)
3741 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3742 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3743 // give it SSA names and value numbers?
3745 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3746 // end with an instance of the structure below, whose fields are described in the declaration.
3747 struct IndirectAssignmentAnnotation
3749 unsigned m_lclNum; // The local num that is being indirectly assigned.
3750 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3751 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3752 // be the singleton field sequence "g". The individual assignments would
3753 // further append the fields of "s.g" to that.
3754 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3755 // structure has a single field).
3756 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3757 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3760 IndirectAssignmentAnnotation(unsigned lclNum,
3761 FieldSeqNode* fldSeq,
3763 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3764 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3765 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3769 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3770 NodeToIndirAssignMap;
3771 NodeToIndirAssignMap* m_indirAssignMap;
3772 NodeToIndirAssignMap* GetIndirAssignMap()
3774 if (m_indirAssignMap == nullptr)
3776 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3777 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3778 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3780 return m_indirAssignMap;
3783 // Performs SSA conversion.
3786 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3787 void fgResetForSsa();
3789 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3791 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3792 inline bool fgExcludeFromSsa(unsigned lclNum);
3794 // The value numbers for this compilation.
3795 ValueNumStore* vnStore;
3798 ValueNumStore* GetValueNumStore()
3803 // Do value numbering (assign a value number to each
3805 void fgValueNumber();
3807 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3808 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3809 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3810 // match the element type of the array or fldSeq. When this type doesn't match
3811 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3813 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3816 FieldSeqNode* fldSeq,
3820 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3821 // has been parsed to yield the other input arguments. If evaluation of the address
3822 // can raise exceptions, those should be captured in the exception set "excVN."
3823 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3824 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3825 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3826 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3827 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3829 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3830 CORINFO_CLASS_HANDLE elemTypeEq,
3834 FieldSeqNode* fldSeq);
3836 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3837 // by evaluating the array index expression "tree". Returns the value number resulting from
3838 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
3839 // "GT_IND" that does the dereference, and it is given the returned value number.
3840 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3842 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
3843 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
3845 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3847 // Utility functions for fgValueNumber.
3849 // Perform value-numbering for the trees in "blk".
3850 void fgValueNumberBlock(BasicBlock* blk);
3852 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3853 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3854 // assumed for the memoryKind at the start "entryBlk".
3855 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3857 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3858 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
3859 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3861 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
3863 void fgMutateAddressExposedLocal(GenTreePtr tree DEBUGARG(const char* msg));
3865 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
3866 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
3867 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3869 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
3870 void recordAddressExposedLocalStore(GenTreePtr curTree, ValueNum memoryVN DEBUGARG(const char* msg));
3872 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3873 // value in that SSA #.
3874 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3876 // The input 'tree' is a leaf node that is a constant
3877 // Assign the proper value number to the tree
3878 void fgValueNumberTreeConst(GenTreePtr tree);
3880 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3881 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3883 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3885 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3887 // Does value-numbering for a block assignment.
3888 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3890 // Does value-numbering for a cast tree.
3891 void fgValueNumberCastTree(GenTreePtr tree);
3893 // Does value-numbering for an intrinsic tree.
3894 void fgValueNumberIntrinsic(GenTreePtr tree);
3896 // Does value-numbering for a call. We interpret some helper calls.
3897 void fgValueNumberCall(GenTreeCall* call);
3899 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3900 void fgUpdateArgListVNs(GenTreeArgList* args);
3902 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3903 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3905 // Requires "helpCall" to be a helper call. Assigns it a value number;
3906 // we understand the semantics of some of the calls. Returns "true" if
3907 // the call may modify the heap (we assume arbitrary memory side effects if so).
3908 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3910 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3911 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3913 // These are the current value number for the memory implicit variables while
3914 // doing value numbering. These are the value numbers under the "liberal" interpretation
3915 // of memory values; the "conservative" interpretation needs no VN, since every access of
3916 // memory yields an unknown value.
3917 ValueNum fgCurMemoryVN[MemoryKindCount];
3919 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3920 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3921 // is 1, and the rest is an encoding of "elemTyp".
3922 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3924 if (elemStructType != nullptr)
3926 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3927 varTypeIsIntegral(elemTyp));
3928 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3929 return elemStructType;
3933 elemTyp = varTypeUnsignedToSigned(elemTyp);
3934 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3937 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3938 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3939 // the struct type of the element).
3940 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3942 size_t clsHndVal = size_t(clsHnd);
3943 if (clsHndVal & 0x1)
3945 return var_types(clsHndVal >> 1);
3953 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3954 var_types getJitGCType(BYTE gcType);
3956 enum structPassingKind
3958 SPK_Unknown, // Invalid value, never returned
3959 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3960 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3961 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3962 // parameters registers are used, then the stack will be used)
3963 // for X86 passed on the stack, for ARM32 passed in registers
3964 // or the stack or split between registers and the stack.
3965 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3967 }; // The struct is passed/returned by reference to a copy/buffer.
3969 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3970 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3971 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3972 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3974 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3976 // Get the type that is used to pass values of the given struct type.
3977 // If you have already retrieved the struct size then pass it as the optional third argument
3979 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3980 structPassingKind* wbPassStruct,
3981 unsigned structSize = 0);
3983 // Get the type that is used to return values of the given struct type.
3984 // If you have already retrieved the struct size then pass it as the optional third argument
3986 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3987 structPassingKind* wbPassStruct = nullptr,
3988 unsigned structSize = 0);
3991 // Print a representation of "vnp" or "vn" on standard output.
3992 // If "level" is non-zero, we also print out a partial expansion of the value.
3993 void vnpPrint(ValueNumPair vnp, unsigned level);
3994 void vnPrint(ValueNum vn, unsigned level);
3997 // Dominator computation member functions
3998 // Not exposed outside Compiler
4000 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4002 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4004 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4005 // flow graph. We first assume the fields bbIDom on each
4006 // basic block are invalid. This computation is needed later
4007 // by fgBuildDomTree to build the dominance tree structure.
4008 // Based on: A Simple, Fast Dominance Algorithm
4009 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4011 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4012 // Note: this is relatively slow compared to calling fgDominate(),
4013 // especially if dealing with a single block versus block check.
4015 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4017 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4019 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4021 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4023 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4025 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4026 // processed in topological sort, this function takes care of that.
4028 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4030 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4031 // Returns this as a set.
4033 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4034 // root nodes. Returns this as a set.
4037 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4040 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4041 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4044 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4045 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4046 // && postOrder(A) >= postOrder(B) making the computation O(1).
4047 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4049 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4051 void fgUpdateChangedFlowGraph();
4054 // Compute the predecessors of the blocks in the control flow graph.
4055 void fgComputePreds();
4057 // Remove all predecessor information.
4058 void fgRemovePreds();
4060 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4061 // before the full predecessors lists are computed.
4062 void fgComputeCheapPreds();
4065 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4067 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4077 // Initialize the per-block variable sets (used for liveness analysis).
4078 void fgInitBlockVarSets();
4080 // true if we've gone through and created GC Poll calls.
4081 bool fgGCPollsCreated;
4082 void fgMarkGCPollBlocks();
4083 void fgCreateGCPolls();
4084 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4086 // Requires that "block" is a block that returns from
4087 // a finally. Returns the number of successors (jump targets of
4088 // of blocks in the covered "try" that did a "LEAVE".)
4089 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4091 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4092 // a finally. Returns its "i"th successor (jump targets of
4093 // of blocks in the covered "try" that did a "LEAVE".)
4094 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4095 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4098 // Factor out common portions of the impls of the methods above.
4099 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4102 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4103 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4104 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4105 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4106 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4107 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4108 // we leave the entry associated with the block, but it will no longer be accessed.)
4109 struct SwitchUniqueSuccSet
4111 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4112 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4115 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4116 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4117 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4118 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4121 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4122 BlockToSwitchDescMap;
4125 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4126 // iteration over only the distinct successors.
4127 BlockToSwitchDescMap* m_switchDescMap;
4130 BlockToSwitchDescMap* GetSwitchDescMap()
4132 if (m_switchDescMap == nullptr)
4134 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4136 return m_switchDescMap;
4139 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4140 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4141 // we don't accidentally look up and return the wrong switch data.
4142 void InvalidateUniqueSwitchSuccMap()
4144 m_switchDescMap = nullptr;
4147 // Requires "switchBlock" to be a block that ends in a switch. Returns
4148 // the corresponding SwitchUniqueSuccSet.
4149 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4151 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4152 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4153 // remove it from "this", and ensure that "to" is a member.
4154 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4156 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4157 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4159 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4161 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4163 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4165 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4167 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4169 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4171 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4173 void fgRemoveBlockAsPred(BasicBlock* block);
4175 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4177 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4179 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4181 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4183 flowList* fgAddRefPred(BasicBlock* block,
4184 BasicBlock* blockPred,
4185 flowList* oldEdge = nullptr,
4186 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4189 void fgFindBasicBlocks();
4191 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4193 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4195 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4196 bool putInTryRegion,
4197 BasicBlock* startBlk,
4199 BasicBlock* nearBlk,
4200 BasicBlock* jumpBlk,
4203 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4205 void fgRemoveEmptyBlocks();
4207 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4209 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4211 void fgCreateLoopPreHeader(unsigned lnum);
4213 void fgUnreachableBlock(BasicBlock* block);
4215 void fgRemoveConditionalJump(BasicBlock* block);
4217 BasicBlock* fgLastBBInMainFunction();
4219 BasicBlock* fgEndBBAfterMainFunction();
4221 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4223 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4225 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4227 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4229 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4231 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4233 bool fgRenumberBlocks();
4235 bool fgExpandRarelyRunBlocks();
4237 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4239 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4241 enum FG_RELOCATE_TYPE
4243 FG_RELOCATE_TRY, // relocate the 'try' region
4244 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4246 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4248 #if FEATURE_EH_FUNCLETS
4249 #if defined(_TARGET_ARM_)
4250 void fgClearFinallyTargetBit(BasicBlock* block);
4251 #endif // defined(_TARGET_ARM_)
4252 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4253 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4254 void fgInsertFuncletPrologBlock(BasicBlock* block);
4255 void fgCreateFuncletPrologBlocks();
4256 void fgCreateFunclets();
4257 #else // !FEATURE_EH_FUNCLETS
4258 bool fgRelocateEHRegions();
4259 #endif // !FEATURE_EH_FUNCLETS
4261 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4263 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4265 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4267 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4269 bool fgOptimizeEmptyBlock(BasicBlock* block);
4271 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4273 bool fgOptimizeBranch(BasicBlock* bJump);
4275 bool fgOptimizeSwitchBranches(BasicBlock* block);
4277 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4279 bool fgOptimizeSwitchJumps();
4281 void fgPrintEdgeWeights();
4283 void fgComputeEdgeWeights();
4285 void fgReorderBlocks();
4287 void fgDetermineFirstColdBlock();
4289 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4291 bool fgUpdateFlowGraph(bool doTailDup = false);
4293 void fgFindOperOrder();
4295 // method that returns if you should split here
4296 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4298 void fgSetBlockOrder();
4300 void fgRemoveReturnBlock(BasicBlock* block);
4302 /* Helper code that has been factored out */
4303 inline void fgConvertBBToThrowBB(BasicBlock* block);
4305 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4306 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4307 GenTreePtr fgMakeTmpArgNode(
4308 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4310 // The following check for loops that don't execute calls
4311 bool fgLoopCallMarked;
4313 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4314 void fgLoopCallMark();
4316 void fgMarkLoopHead(BasicBlock* block);
4318 unsigned fgGetCodeEstimate(BasicBlock* block);
4321 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4322 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4323 bool fgDumpFlowGraph(Phases phase);
4325 #endif // DUMP_FLOWGRAPHS
4330 void fgDispBBLiveness(BasicBlock* block);
4331 void fgDispBBLiveness();
4332 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4333 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4334 void fgDispBasicBlocks(bool dumpTrees = false);
4335 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4336 void fgDumpBlock(BasicBlock* block);
4337 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4339 static fgWalkPreFn fgStress64RsltMulCB;
4340 void fgStress64RsltMul();
4341 void fgDebugCheckUpdate();
4342 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4343 void fgDebugCheckBlockLinks();
4344 void fgDebugCheckLinks(bool morphTrees = false);
4345 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4346 void fgDebugCheckFlags(GenTreePtr tree);
4347 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4348 void fgDebugCheckTryFinallyExits();
4351 #ifdef LEGACY_BACKEND
4352 static void fgOrderBlockOps(GenTreePtr tree,
4356 GenTreePtr* opsPtr, // OUT
4357 regMaskTP* regsPtr); // OUT
4358 #endif // LEGACY_BACKEND
4360 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4361 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4363 inline bool fgIsInlining()
4365 return fgExpandInline;
4368 void fgTraverseRPO();
4370 //--------------------- Walking the trees in the IR -----------------------
4375 fgWalkPreFn* wtprVisitorFn;
4376 fgWalkPostFn* wtpoVisitorFn;
4377 void* pCallbackData; // user-provided data
4378 bool wtprLclsOnly; // whether to only visit lclvar nodes
4379 GenTreePtr parent; // parent of current node, provided to callback
4380 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4382 bool printModified; // callback can use this
4386 template <bool computeStack>
4387 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4389 // general purpose tree-walker that is capable of doing pre- and post- order
4390 // callbacks at the same time
4391 template <bool doPreOrder, bool doPostOrder>
4392 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4394 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4395 fgWalkPreFn* visitor,
4396 void* pCallBackData = nullptr,
4397 bool lclVarsOnly = false,
4398 bool computeStack = false);
4400 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4401 fgWalkPreFn* preVisitor,
4402 fgWalkPostFn* postVisitor,
4403 void* pCallBackData = nullptr);
4405 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4409 template <bool computeStack>
4410 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4412 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4413 fgWalkPostFn* visitor,
4414 void* pCallBackData = nullptr,
4415 bool computeStack = false);
4417 // An fgWalkPreFn that looks for expressions that have inline throws in
4418 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4419 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4420 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4421 // properly propagated to parent trees). It returns WALK_CONTINUE
4423 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4424 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4425 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4427 /**************************************************************************
4429 *************************************************************************/
4432 friend class SsaBuilder;
4433 friend struct ValueNumberState;
4435 //--------------------- Detect the basic blocks ---------------------------
4437 BasicBlock** fgBBs; // Table of pointers to the BBs
4439 void fgInitBBLookup();
4440 BasicBlock* fgLookupBB(unsigned addr);
4442 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4444 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4446 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4448 void fgLinkBasicBlocks();
4450 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4452 void fgCheckBasicBlockControlFlow();
4454 void fgControlFlowPermitted(BasicBlock* blkSrc,
4455 BasicBlock* blkDest,
4456 BOOL IsLeave = false /* is the src a leave block */);
4458 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4460 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4462 void fgAdjustForAddressExposedOrWrittenThis();
4464 bool fgProfileData_ILSizeMismatch;
4465 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4466 ULONG fgProfileBufferCount;
4467 ULONG fgNumProfileRuns;
4469 unsigned fgStressBBProf()
4472 unsigned result = JitConfig.JitStressBBProf();
4475 if (compStressCompile(STRESS_BB_PROFILE, 15))
4486 bool fgHaveProfileData();
4487 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4489 bool fgIsUsingProfileWeights()
4491 return (fgHaveProfileData() || fgStressBBProf());
4493 void fgInstrumentMethod();
4495 //-------- Insert a statement at the start or end of a basic block --------
4499 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4503 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4505 public: // Used by linear scan register allocation
4506 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4509 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4510 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4512 public: // Used by linear scan register allocation
4513 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4516 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4518 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4520 // Create a new temporary variable to hold the result of *ppTree,
4521 // and transform the graph accordingly.
4522 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4523 GenTree* fgMakeMultiUse(GenTree** ppTree);
4526 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4527 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4528 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4530 //-------- Determine the order in which the trees will be evaluated -------
4532 unsigned fgTreeSeqNum;
4533 GenTree* fgTreeSeqLst;
4534 GenTree* fgTreeSeqBeg;
4536 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4537 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4538 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4539 void fgSetStmtSeq(GenTree* tree);
4540 void fgSetBlockOrder(BasicBlock* block);
4542 //------------------------- Morphing --------------------------------------
4544 unsigned fgPtrArgCntCur;
4545 unsigned fgPtrArgCntMax;
4546 hashBv* fgOutgoingArgTemps;
4547 hashBv* fgCurrentlyInUseArgTemps;
4549 bool compCanEncodePtrArgCntMax();
4551 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4554 void fgMoveOpsLeft(GenTreePtr tree);
4557 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4559 bool fgIsThrow(GenTreePtr tree);
4561 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4562 bool fgIsBlockCold(BasicBlock* block);
4564 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4566 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4568 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4570 bool fgMorphRelopToQmark(GenTreePtr tree);
4572 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4573 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4574 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4575 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4576 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4577 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4578 // small; hence the other fields of MorphAddrContext.
4579 enum MorphAddrContextKind
4584 struct MorphAddrContext
4586 MorphAddrContextKind m_kind;
4587 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4588 // top-level indirection and here have been constants.
4589 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4590 // In that case, is the sum of those constant offsets.
4592 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4597 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4598 static MorphAddrContext s_CopyBlockMAC;
4601 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4602 var_types* baseTypeOut,
4604 unsigned* simdSizeOut,
4605 bool ignoreUsedInSIMDIntrinsic = false);
4606 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4607 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4608 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4609 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4611 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4612 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4613 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4615 #endif // FEATURE_SIMD
4616 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4617 GenTreePtr fgMorphCast(GenTreePtr tree);
4618 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4619 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4621 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4624 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4625 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4627 void fgFixupStructReturn(GenTreePtr call);
4628 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4629 bool fgAddrCouldBeNull(GenTreePtr addr);
4630 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4631 bool fgCanFastTailCall(GenTreeCall* call);
4632 void fgMorphTailCall(GenTreeCall* call);
4633 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4634 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4635 fgArgTabEntryPtr argTabEntry,
4637 IL_OFFSETX callILOffset,
4638 GenTreePtr tmpAssignmentInsertionPoint,
4639 GenTreePtr paramAssignmentInsertionPoint);
4640 static int fgEstimateCallStackSize(GenTreeCall* call);
4641 GenTreePtr fgMorphCall(GenTreeCall* call);
4642 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4643 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4645 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4646 static fgWalkPreFn fgFindNonInlineCandidate;
4648 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4649 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4650 void fgAssignSetVarDef(GenTreePtr tree);
4651 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4652 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4653 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4654 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4655 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4656 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4657 void fgMorphUnsafeBlk(GenTreeObj* obj);
4658 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4659 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4660 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4661 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4662 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4663 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4664 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4666 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4667 GenTreePtr fgMorphConst(GenTreePtr tree);
4670 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4673 #if LOCAL_ASSERTION_PROP
4674 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4675 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4677 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4679 GenTreeStmt* fgMorphStmt;
4681 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4682 // used when morphing big offset.
4684 //----------------------- Liveness analysis -------------------------------
4686 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4687 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4689 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4690 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4691 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4693 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4695 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4697 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4698 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4700 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4701 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4703 void fgExtendDbgScopes();
4704 void fgExtendDbgLifetimes();
4707 void fgDispDebugScopes();
4710 //-------------------------------------------------------------------------
4712 // The following keeps track of any code we've added for things like array
4713 // range checking or explicit calls to enable GC, and so on.
4718 AddCodeDsc* acdNext;
4719 BasicBlock* acdDstBlk; // block to which we jump
4721 SpecialCodeKind acdKind; // what kind of a special block is this?
4722 unsigned short acdStkLvl;
4726 static unsigned acdHelper(SpecialCodeKind codeKind);
4728 AddCodeDsc* fgAddCodeList;
4730 bool fgRngChkThrowAdded;
4731 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4733 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4735 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4738 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4741 bool fgIsCodeAdded();
4743 bool fgIsThrowHlpBlk(BasicBlock* block);
4744 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4746 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4748 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4749 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4750 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4751 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4752 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4754 #if FEATURE_MULTIREG_RET
4755 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4756 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4757 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4758 #endif // FEATURE_MULTIREG_RET
4760 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4763 static fgWalkPreFn fgDebugCheckInlineCandidates;
4766 void fgPromoteStructs();
4767 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4768 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4769 void fgMarkImplicitByRefArgs();
4770 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4771 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4772 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4773 void fgMarkAddressExposedLocals();
4774 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4776 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4778 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4780 // The given local variable, required to be a struct variable, is being assigned via
4781 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4782 // the variable is not enregistered, and is therefore not promoted independently.
4783 void fgLclFldAssign(unsigned lclNum);
4785 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4786 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4787 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4788 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4791 bool fgPrintInlinedMethods;
4794 bool fgIsBigOffset(size_t offset);
4796 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4797 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4798 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4799 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4800 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4803 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4804 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4808 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4809 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4816 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4819 void optRemoveRangeCheck(
4820 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4821 bool optIsRangeCheckRemovable(GenTreePtr tree);
4824 static fgWalkPreFn optValidRangeCheckIndex;
4825 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4828 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4830 /**************************************************************************
4832 *************************************************************************/
4835 // Do hoisting for all loops.
4836 void optHoistLoopCode();
4838 // To represent sets of VN's that have already been hoisted in outer loops.
4839 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4840 typedef VNToBoolMap VNSet;
4842 struct LoopHoistContext
4845 // The set of variables hoisted in the current loop (or nullptr if there are none).
4846 VNSet* m_pHoistedInCurLoop;
4849 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4850 VNSet m_hoistedInParentLoops;
4851 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4852 // Previous decisions on loop-invariance of value numbers in the current loop.
4853 VNToBoolMap m_curLoopVnInvariantCache;
4855 VNSet* GetHoistedInCurLoop(Compiler* comp)
4857 if (m_pHoistedInCurLoop == nullptr)
4859 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4861 return m_pHoistedInCurLoop;
4864 VNSet* ExtractHoistedInCurLoop()
4866 VNSet* res = m_pHoistedInCurLoop;
4867 m_pHoistedInCurLoop = nullptr;
4871 LoopHoistContext(Compiler* comp)
4872 : m_pHoistedInCurLoop(nullptr)
4873 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4874 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4879 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4880 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4881 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4882 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4884 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4885 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4886 // "m_hoistedInParentLoops".
4888 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4890 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4891 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4892 // expressions to "hoistInLoop".
4893 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4895 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4896 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4898 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4899 // that are invariant in loop "lnum" (an index into the optLoopTable)
4900 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4901 // expressions to "hoistInLoop".
4902 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4903 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4904 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4905 bool optHoistLoopExprsForTree(GenTreePtr tree,
4907 LoopHoistContext* hoistCtxt,
4908 bool* firstBlockAndBeforeSideEffect,
4911 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4912 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4914 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4915 // Constants and init values are always loop invariant.
4916 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4917 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4919 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4920 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4921 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4922 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4923 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4925 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4926 // in the loop table.
4927 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4929 // Records the set of "side effects" of all loops: fields (object instance and static)
4930 // written to, and SZ-array element type equivalence classes updated.
4931 void optComputeLoopSideEffects();
4934 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4935 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4936 // static) written to, and SZ-array element type equivalence classes updated.
4937 void optComputeLoopNestSideEffects(unsigned lnum);
4939 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4940 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4942 // Hoist the expression "expr" out of loop "lnum".
4943 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4946 void optOptimizeBools();
4949 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4951 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4954 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4956 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4957 // the loop into a "do-while" loop
4958 // Also finds all natural loops and records them in the loop table
4960 // Optionally clone loops in the loop table.
4961 void optCloneLoops();
4963 // Clone loop "loopInd" in the loop table.
4964 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4966 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4967 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4968 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4970 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4972 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4975 // This enumeration describes what is killed by a call.
4979 CALLINT_NONE, // no interference (most helpers)
4980 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4981 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4982 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4983 CALLINT_ALL, // kills everything (normal method call)
4987 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4988 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4989 // in bbNext order; we use comparisons on the bbNum to decide order.)
4990 // The blocks that define the body are
4991 // first <= top <= entry <= bottom .
4992 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4993 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4994 // Compiler::optFindNaturalLoops().
4997 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4998 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4999 // loop, but not the outer loop.)
5000 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5002 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5003 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5004 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5006 callInterf lpAsgCall; // "callInterf" for calls in the loop
5007 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5008 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5010 unsigned short lpFlags; // Mask of the LPFLG_* constants
5012 unsigned char lpExitCnt; // number of exits from the loop
5014 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5015 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5016 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5017 // (Actually, an "immediately" nested loop --
5018 // no other child of this loop is a parent of lpChild.)
5019 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5020 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5021 // by following "lpChild" then "lpSibling" links.
5023 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5024 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5026 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5027 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5028 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5030 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5031 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5033 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5034 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5035 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5036 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5038 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5039 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5040 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5042 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5043 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5044 // type are assigned to.
5046 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5047 // memory side effects. If this is set, the fields below
5048 // may not be accurate (since they become irrelevant.)
5049 bool lpContainsCall; // True if executing the loop body *may* execute a call
5051 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5052 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5054 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5056 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5057 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5059 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5061 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5062 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5064 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5065 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5067 JitSimplerHashBehavior>
5069 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5070 // instance fields modified
5073 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5074 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5076 JitSimplerHashBehavior>
5078 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5079 // arrays of that type are modified
5082 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5083 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5085 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5086 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5087 // (shifted left, with a low-order bit set to distinguish.)
5088 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5089 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5091 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5093 GenTreePtr lpIterTree; // The "i <op>= const" tree
5094 unsigned lpIterVar(); // iterator variable #
5095 int lpIterConst(); // the constant with which the iterator is incremented
5096 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5097 void VERIFY_lpIterTree();
5099 var_types lpIterOperType(); // For overflow instructions
5102 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5103 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5107 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5109 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5110 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5111 void VERIFY_lpTestTree();
5113 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5114 GenTreePtr lpIterator(); // the iterator node in the loop test
5115 GenTreePtr lpLimit(); // the limit node in the loop test
5117 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5118 // LPFLG_CONST_LIMIT
5119 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5121 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5122 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5123 // LPFLG_ARRLEN_LIMIT
5125 // Returns "true" iff "*this" contains the blk.
5126 bool lpContains(BasicBlock* blk)
5128 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5130 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5131 // to be equal, but requiring bottoms to be different.)
5132 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5134 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5137 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5138 // bottoms to be different.)
5139 bool lpContains(const LoopDsc& lp2)
5141 return lpContains(lp2.lpFirst, lp2.lpBottom);
5144 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5145 // (allowing firsts to be equal, but requiring bottoms to be different.)
5146 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5148 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5151 // Returns "true" iff "*this" is (properly) contained by "lp2"
5152 // (allowing firsts to be equal, but requiring bottoms to be different.)
5153 bool lpContainedBy(const LoopDsc& lp2)
5155 return lpContains(lp2.lpFirst, lp2.lpBottom);
5158 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5159 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5161 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5163 // Returns "true" iff "*this" is disjoint from "lp2".
5164 bool lpDisjoint(const LoopDsc& lp2)
5166 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5168 // Returns "true" iff the loop is well-formed (see code for defn).
5171 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5172 lpEntry->bbNum <= lpBottom->bbNum &&
5173 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5178 bool fgMightHaveLoop(); // returns true if there are any backedges
5179 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5182 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5183 unsigned char optLoopCount; // number of tracked loops
5186 unsigned optCallCount; // number of calls made in the method
5187 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5188 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5189 unsigned optLoopsCloned; // number of loops cloned in the current method.
5192 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5193 void optPrintLoopInfo(unsigned loopNum,
5195 BasicBlock* lpFirst,
5197 BasicBlock* lpEntry,
5198 BasicBlock* lpBottom,
5199 unsigned char lpExitCnt,
5201 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5202 void optPrintLoopInfo(unsigned lnum);
5203 void optPrintLoopRecording(unsigned lnum);
5205 void optCheckPreds();
5208 void optSetBlockWeights();
5210 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5212 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5214 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5216 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5217 unsigned optIsLoopIncrTree(GenTreePtr incr);
5218 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5219 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5220 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5221 bool optExtractInitTestIncr(BasicBlock* head,
5226 GenTreePtr* ppIncr);
5228 void optRecordLoop(BasicBlock* head,
5234 unsigned char exitCnt);
5236 void optFindNaturalLoops();
5238 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5239 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5240 bool optCanonicalizeLoopNest(unsigned char loopInd);
5242 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5243 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5244 bool optCanonicalizeLoop(unsigned char loopInd);
5246 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5247 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5248 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5249 bool optLoopContains(unsigned l1, unsigned l2);
5251 // Requires "loopInd" to be a valid index into the loop table.
5252 // Updates the loop table by changing loop "loopInd", whose head is required
5253 // to be "from", to be "to". Also performs this transformation for any
5254 // loop nested in "loopInd" that shares the same head as "loopInd".
5255 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5257 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5258 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5259 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5261 // Marks the containsCall information to "lnum" and any parent loops.
5262 void AddContainsCallAllContainingLoops(unsigned lnum);
5263 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5264 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5265 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5266 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5267 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5268 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5270 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5271 // of "from".) Copies the jump destination from "from" to "to".
5272 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5274 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5275 unsigned optLoopDepth(unsigned lnum)
5277 unsigned par = optLoopTable[lnum].lpParent;
5278 if (par == BasicBlock::NOT_IN_LOOP)
5284 return 1 + optLoopDepth(par);
5288 void fgOptWhileLoop(BasicBlock* block);
5290 bool optComputeLoopRep(int constInit,
5293 genTreeOps iterOper,
5295 genTreeOps testOper,
5298 unsigned* iterCount);
5299 #if FEATURE_STACK_FP_X87
5302 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5303 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5304 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5305 #endif // FEATURE_STACK_FP_X87
5308 static fgWalkPreFn optIsVarAssgCB;
5311 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5313 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5315 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5317 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5319 /**************************************************************************
5320 * Optimization conditions
5321 *************************************************************************/
5323 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5324 bool optPentium4(void);
5325 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5326 bool optAvoidIntMult(void);
5331 // The following is the upper limit on how many expressions we'll keep track
5332 // of for the CSE analysis.
5334 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5336 static const int MIN_CSE_COST = 2;
5338 // Keeps tracked cse indices
5339 BitVecTraits* cseTraits;
5343 /* Generic list of nodes - used by the CSE logic */
5351 typedef struct treeLst* treeLstPtr;
5355 treeStmtLst* tslNext;
5356 GenTreePtr tslTree; // tree node
5357 GenTreePtr tslStmt; // statement containing the tree
5358 BasicBlock* tslBlock; // block containing the statement
5361 typedef struct treeStmtLst* treeStmtLstPtr;
5363 // The following logic keeps track of expressions via a simple hash table.
5367 CSEdsc* csdNextInBucket; // used by the hash table
5369 unsigned csdHashValue; // the orginal hashkey
5371 unsigned csdIndex; // 1..optCSECandidateCount
5372 char csdLiveAcrossCall; // 0 or 1
5374 unsigned short csdDefCount; // definition count
5375 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5377 unsigned csdDefWtCnt; // weighted def count
5378 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5380 GenTreePtr csdTree; // treenode containing the 1st occurance
5381 GenTreePtr csdStmt; // stmt containing the 1st occurance
5382 BasicBlock* csdBlock; // block containing the 1st occurance
5384 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5385 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5388 static const size_t s_optCSEhashSize;
5389 CSEdsc** optCSEhash;
5394 CSEdsc* optCSEfindDsc(unsigned index);
5395 void optUnmarkCSE(GenTreePtr tree);
5397 // user defined callback data for the tree walk function optCSE_MaskHelper()
5398 struct optCSE_MaskData
5400 EXPSET_TP CSE_defMask;
5401 EXPSET_TP CSE_useMask;
5404 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5405 static fgWalkPreFn optCSE_MaskHelper;
5407 // This function walks all the node for an given tree
5408 // and return the mask of CSE definitions and uses for the tree
5410 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5412 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5413 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5414 bool optCSE_canSwap(GenTree* tree);
5416 static fgWalkPostFn optPropagateNonCSE;
5417 static fgWalkPreFn optHasNonCSEChild;
5419 static fgWalkPreFn optUnmarkCSEs;
5421 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5422 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5424 void optCleanupCSEs();
5427 void optEnsureClearCSEInfo();
5430 #endif // FEATURE_ANYCSE
5432 #if FEATURE_VALNUM_CSE
5433 /**************************************************************************
5434 * Value Number based CSEs
5435 *************************************************************************/
5438 void optOptimizeValnumCSEs();
5441 void optValnumCSE_Init();
5442 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5443 unsigned optValnumCSE_Locate();
5444 void optValnumCSE_InitDataFlow();
5445 void optValnumCSE_DataFlow();
5446 void optValnumCSE_Availablity();
5447 void optValnumCSE_Heuristic();
5448 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5450 #endif // FEATURE_VALNUM_CSE
5453 bool optDoCSE; // True when we have found a duplicate CSE tree
5454 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5455 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5456 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5457 unsigned optCSEstart; // The first local variable number that is a CSE
5458 unsigned optCSEcount; // The total count of CSE's introduced.
5459 unsigned optCSEweight; // The weight of the current block when we are
5460 // scanning for CSE expressions
5462 bool optIsCSEcandidate(GenTreePtr tree);
5464 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5466 bool lclNumIsTrueCSE(unsigned lclNum) const
5468 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5471 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5473 bool lclNumIsCSE(unsigned lclNum) const
5475 return lvaTable[lclNum].lvIsCSE;
5479 bool optConfigDisableCSE();
5480 bool optConfigDisableCSE2();
5482 void optOptimizeCSEs();
5484 #endif // FEATURE_ANYCSE
5492 unsigned ivaVar; // Variable we are interested in, or -1
5493 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5494 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5495 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5496 callInterf ivaMaskCall; // What kind of calls are there?
5499 static callInterf optCallInterf(GenTreePtr call);
5502 // VN based copy propagation.
5503 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5504 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5505 LclNumToGenTreePtrStack;
5507 // Kill set to track variables with intervening definitions.
5508 VARSET_TP optCopyPropKillSet;
5510 // Copy propagation functions.
5511 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5512 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5513 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5514 bool optIsSsaLocal(GenTreePtr tree);
5515 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5516 void optVnCopyProp();
5518 /**************************************************************************
5519 * Early value propagation
5520 *************************************************************************/
5526 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5530 static unsigned GetHashCode(SSAName ssaNm)
5532 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5535 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5537 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5541 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5542 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5543 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5544 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5545 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5547 unsigned optMethodFlags;
5549 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5550 // No throughput diff was found with backward walk bound between 3-8.
5551 static const int optEarlyPropRecurBound = 5;
5553 enum class optPropKind
5561 bool gtIsVtableRef(GenTreePtr tree);
5562 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5563 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5564 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5565 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5566 bool optEarlyPropRewriteTree(GenTreePtr tree);
5567 bool optDoEarlyPropForBlock(BasicBlock* block);
5568 bool optDoEarlyPropForFunc();
5569 void optEarlyProp();
5570 void optFoldNullCheck(GenTreePtr tree);
5571 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5574 /**************************************************************************
5575 * Value/Assertion propagation
5576 *************************************************************************/
5578 // Data structures for assertion prop
5579 BitVecTraits* apTraits;
5583 enum optAssertionKind
5598 O1K_ARRLEN_OPER_BND,
5599 O1K_ARRLEN_LOOP_BND,
5600 O1K_CONSTANT_LOOP_BND,
5621 optAssertionKind assertionKind;
5624 unsigned lclNum; // assigned to or property of this local var number
5632 struct AssertionDscOp1
5634 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5641 struct AssertionDscOp2
5643 optOp2Kind kind; // a const or copy assignment
5647 ssize_t iconVal; // integer
5648 unsigned iconFlags; // gtFlags
5650 struct Range // integer subrange
5664 bool IsArrLenArithBound()
5666 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5668 bool IsArrLenBound()
5670 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5672 bool IsConstantBound()
5674 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5675 op1.kind == O1K_CONSTANT_LOOP_BND);
5677 bool IsBoundsCheckNoThrow()
5679 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5682 bool IsCopyAssertion()
5684 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5687 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5689 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5690 a1->op2.kind == a2->op2.kind;
5693 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5695 if (kind == OAK_EQUAL)
5697 return kind2 == OAK_NOT_EQUAL;
5699 else if (kind == OAK_NOT_EQUAL)
5701 return kind2 == OAK_EQUAL;
5706 static ssize_t GetLowerBoundForIntegralType(var_types type)
5726 static ssize_t GetUpperBoundForIntegralType(var_types type)
5750 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5752 return (op1.kind == that->op1.kind) &&
5753 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5756 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5758 if (op2.kind != that->op2.kind)
5764 case O2K_IND_CNS_INT:
5766 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5768 case O2K_CONST_LONG:
5769 return (op2.lconVal == that->op2.lconVal);
5771 case O2K_CONST_DOUBLE:
5772 // exact match because of positive and negative zero.
5773 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5775 case O2K_LCLVAR_COPY:
5777 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5778 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5781 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5784 // we will return false
5788 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5794 bool Complementary(AssertionDsc* that, bool vnBased)
5796 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5797 HasSameOp2(that, vnBased);
5800 bool Equals(AssertionDsc* that, bool vnBased)
5802 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5806 typedef unsigned short AssertionIndex;
5809 static fgWalkPreFn optAddCopiesCallback;
5810 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5811 unsigned optAddCopyLclNum;
5812 GenTreePtr optAddCopyAsgnNode;
5814 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5815 bool optAssertionPropagated; // set to true if we modified the trees
5816 bool optAssertionPropagatedCurrentStmt;
5818 GenTreePtr optAssertionPropCurrentTree;
5820 AssertionIndex* optComplementaryAssertionMap;
5821 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5822 // using the value of a local var) for each local var
5823 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5824 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5825 AssertionIndex optMaxAssertionCount;
5828 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5829 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5830 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5831 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5832 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5833 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5835 AssertionIndex GetAssertionCount()
5837 return optAssertionCount;
5839 ASSERT_TP* bbJtrueAssertionOut;
5840 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5841 ValueNumToAssertsMap;
5842 ValueNumToAssertsMap* optValueNumToAsserts;
5844 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5846 // Assertion prop helpers.
5847 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5848 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5849 void optAssertionInit(bool isLocalProp);
5850 void optAssertionTraitsInit(AssertionIndex assertionCount);
5851 #if LOCAL_ASSERTION_PROP
5852 void optAssertionReset(AssertionIndex limit);
5853 void optAssertionRemove(AssertionIndex index);
5856 // Assertion prop data flow functions.
5857 void optAssertionPropMain();
5858 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5859 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5860 ASSERT_TP* optInitAssertionDataflowFlags();
5861 ASSERT_TP* optComputeAssertionGen();
5863 // Assertion Gen functions.
5864 void optAssertionGen(GenTreePtr tree);
5865 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5866 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5867 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5868 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5869 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5870 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5872 // Assertion creation functions.
5873 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5874 AssertionIndex optCreateAssertion(GenTreePtr op1,
5876 optAssertionKind assertionKind,
5877 AssertionDsc* assertion);
5878 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5880 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5881 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5882 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5884 void optPrintVnAssertionMapping();
5886 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5888 // Used for respective assertion propagations.
5889 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5890 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5891 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5892 bool optAssertionIsNonNull(GenTreePtr op,
5893 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5895 // Used for Relop propagation.
5896 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5897 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5898 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5900 // Assertion prop for lcl var functions.
5901 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5902 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5904 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5905 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5906 const GenTreePtr tree,
5907 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5908 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5910 // Assertion propagation functions.
5911 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5912 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5913 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5914 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5915 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5916 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5917 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5918 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5919 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5920 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5921 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5922 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5924 // Implied assertion functions.
5925 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5926 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5927 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5928 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5930 ASSERT_VALRET_TP optNewFullAssertSet();
5931 ASSERT_VALRET_TP optNewEmptyAssertSet();
5934 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5935 void optDebugCheckAssertion(AssertionDsc* assertion);
5936 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5938 void optAddCopies();
5939 #endif // ASSERTION_PROP
5941 /**************************************************************************
5943 *************************************************************************/
5946 struct LoopCloneVisitorInfo
5948 LoopCloneContext* context;
5951 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5952 : context(context), loopNum(loopNum), stmt(nullptr)
5957 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5958 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5959 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5960 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5961 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5962 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5963 void optObtainLoopCloningOpts(LoopCloneContext* context);
5964 bool optIsLoopClonable(unsigned loopInd);
5966 bool optCanCloneLoops();
5969 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5971 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5972 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5973 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5974 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5978 void optInsertLoopCloningStress(BasicBlock* head);
5980 #if COUNT_RANGECHECKS
5981 static unsigned optRangeChkRmv;
5982 static unsigned optRangeChkAll;
5991 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5996 RngChkDsc* rcdNextInBucket; // used by the hash table
5998 unsigned short rcdHashValue; // to make matching faster
5999 unsigned short rcdIndex; // 0..optRngChkCount-1
6001 GenTreePtr rcdTree; // the array index tree
6004 unsigned optRngChkCount;
6005 static const size_t optRngChkHashSize;
6007 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
6008 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
6010 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6013 bool optLoopsMarked;
6016 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6017 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6021 XX Does the register allocation and puts the remaining lclVars on the stack XX
6023 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6024 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6028 #ifndef LEGACY_BACKEND
6033 #else // LEGACY_BACKEND
6038 #endif // LEGACY_BACKEND
6040 #ifdef LEGACY_BACKEND
6042 void raAssignVars(); // register allocation
6043 #endif // LEGACY_BACKEND
6045 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6047 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6049 void raMarkStkVars();
6052 // Some things are used by both LSRA and regpredict allocators.
6054 FrameType rpFrameType;
6055 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6057 #ifdef LEGACY_BACKEND
6058 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6060 #endif // LEGACY_BACKEND
6062 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6064 #if FEATURE_FP_REGALLOC
6065 enum enumConfigRegisterFP
6067 CONFIG_REGISTER_FP_NONE = 0x0,
6068 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6069 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6070 CONFIG_REGISTER_FP_FULL = 0x3,
6072 enumConfigRegisterFP raConfigRegisterFP();
6073 #endif // FEATURE_FP_REGALLOC
6076 regMaskTP raConfigRestrictMaskFP();
6079 #ifndef LEGACY_BACKEND
6080 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6081 #else // LEGACY_BACKEND
6082 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6083 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6084 bool raNewBlocks; // True is we added killing blocks for FPU registers
6085 unsigned rpPasses; // Number of passes made by the register predicter
6086 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6087 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6088 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6089 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6090 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6091 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6092 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6093 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6094 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6095 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6096 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6097 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6099 bool rpRegAllocDone; // Set to true after we have completed register allocation
6101 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6103 void raSetupArgMasks(RegState* r);
6105 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6107 void raDumpVarIntf(); // Dump the variable to variable interference graph
6108 void raDumpRegIntf(); // Dump the variable to register interference graph
6110 void raAdjustVarIntf();
6112 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6114 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6116 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6117 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6119 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6121 static fgWalkPreFn rpMarkRegIntf;
6123 regMaskTP rpPredictAddressMode(
6124 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6126 void rpPredictRefAssign(unsigned lclNum);
6128 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6130 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6132 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6134 void rpPredictRegUse(); // Entry point
6136 unsigned raPredictTreeRegUse(GenTreePtr tree);
6137 unsigned raPredictListRegUse(GenTreePtr list);
6139 void raSetRegVarOrder(var_types regType,
6140 regNumber* customVarOrder,
6141 unsigned* customVarOrderSize,
6143 regMaskTP avoidReg);
6145 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6146 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6147 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6148 void raAddToStkPredict(unsigned val)
6150 unsigned newStkPredict = rpStkPredict + val;
6151 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6152 rpStkPredict = UINT_MAX - 1;
6154 rpStkPredict = newStkPredict;
6158 #if !FEATURE_FP_REGALLOC
6159 void raDispFPlifeInfo();
6163 regMaskTP genReturnRegForTree(GenTreePtr tree);
6164 #endif // LEGACY_BACKEND
6166 /* raIsVarargsStackArg is called by raMaskStkVars and by
6167 lvaSortByRefCount. It identifies the special case
6168 where a varargs function has a parameter passed on the
6169 stack, other than the special varargs handle. Such parameters
6170 require special treatment, because they cannot be tracked
6171 by the GC (their offsets in the stack are not known
6175 bool raIsVarargsStackArg(unsigned lclNum)
6179 LclVarDsc* varDsc = &lvaTable[lclNum];
6181 assert(varDsc->lvIsParam);
6183 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6185 #else // _TARGET_X86_
6189 #endif // _TARGET_X86_
6192 #ifdef LEGACY_BACKEND
6193 // Records the current prediction, if it's better than any previous recorded prediction.
6194 void rpRecordPrediction();
6195 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6196 void rpUseRecordedPredictionIfBetter();
6198 // Data members used in the methods above.
6199 unsigned rpBestRecordedStkPredict;
6200 struct VarRegPrediction
6202 bool m_isEnregistered;
6203 regNumberSmall m_regNum;
6204 regNumberSmall m_otherReg;
6206 VarRegPrediction* rpBestRecordedPrediction;
6207 #endif // LEGACY_BACKEND
6210 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6211 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6215 XX Get to the class and method info from the Execution Engine given XX
6216 XX tokens for the class and method XX
6218 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6219 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6223 /* These are the different addressing modes used to access a local var.
6224 * The JIT has to report the location of the locals back to the EE
6225 * for debugging purposes.
6231 VLT_REG_BYREF, // this type is currently only used for value types on X64
6234 VLT_STK_BYREF, // this type is currently only used for value types on X64
6248 siVarLocType vlType;
6251 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6253 // VLT_REG_BYREF -- the specified register contains the address of the variable
6261 // VLT_STK -- Any 32 bit value which is on the stack
6262 // eg. [ESP+0x20], or [EBP-0x28]
6263 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6264 // eg. mov EAX, [ESP+0x20]; [EAX]
6268 regNumber vlsBaseReg;
6269 NATIVE_OFFSET vlsOffset;
6272 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6281 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6282 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6290 regNumber vlrssBaseReg;
6291 NATIVE_OFFSET vlrssOffset;
6295 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6296 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6302 regNumber vlsrsBaseReg;
6303 NATIVE_OFFSET vlsrsOffset;
6309 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6310 // eg 2 DWords at [ESP+0x10]
6314 regNumber vls2BaseReg;
6315 NATIVE_OFFSET vls2Offset;
6318 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6319 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6326 // VLT_FIXED_VA -- fixed argument of a varargs function.
6327 // The argument location depends on the size of the variable
6328 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6329 // location of the first arg. This argument can then be accessed
6330 // relative to the position of the first arg
6334 unsigned vlfvOffset;
6341 void* rpValue; // pointer to the in-process
6342 // location of the value.
6348 bool vlIsInReg(regNumber reg);
6349 bool vlIsOnStk(regNumber reg, signed offset);
6352 /*************************************************************************/
6357 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6358 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6359 CORINFO_CALLINFO_FLAGS flags,
6360 CORINFO_CALL_INFO* pResult);
6361 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6363 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6364 CORINFO_ACCESS_FLAGS flags,
6365 CORINFO_FIELD_INFO* pResult);
6369 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6371 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6373 bool IsSuperPMIException(unsigned code)
6375 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6377 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6378 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6379 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6380 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6381 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6382 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6383 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6384 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6388 case EXCEPTIONCODE_DebugBreakorAV:
6389 case EXCEPTIONCODE_MC:
6390 case EXCEPTIONCODE_LWM:
6391 case EXCEPTIONCODE_SASM:
6392 case EXCEPTIONCODE_SSYM:
6393 case EXCEPTIONCODE_CALLUTILS:
6394 case EXCEPTIONCODE_TYPEUTILS:
6395 case EXCEPTIONCODE_ASSERT:
6402 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6403 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6405 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6406 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6409 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6410 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6411 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6413 // VOM info, method sigs
6415 void eeGetSig(unsigned sigTok,
6416 CORINFO_MODULE_HANDLE scope,
6417 CORINFO_CONTEXT_HANDLE context,
6418 CORINFO_SIG_INFO* retSig);
6420 void eeGetCallSiteSig(unsigned sigTok,
6421 CORINFO_MODULE_HANDLE scope,
6422 CORINFO_CONTEXT_HANDLE context,
6423 CORINFO_SIG_INFO* retSig);
6425 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6427 // Method entry-points, instrs
6429 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6431 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6433 CORINFO_EE_INFO eeInfo;
6434 bool eeInfoInitialized;
6436 CORINFO_EE_INFO* eeGetEEInfo();
6438 // Gets the offset of a SDArray's first element
6439 unsigned eeGetArrayDataOffset(var_types type);
6440 // Gets the offset of a MDArray's first element
6441 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6443 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6445 // Returns the page size for the target machine as reported by the EE.
6446 inline size_t eeGetPageSize()
6448 #if COR_JIT_EE_VERSION > 460
6449 return eeGetEEInfo()->osPageSize;
6450 #else // COR_JIT_EE_VERSION <= 460
6451 return CORINFO_PAGE_SIZE;
6452 #endif // COR_JIT_EE_VERSION > 460
6455 // Returns the frame size at which we will generate a loop to probe the stack.
6456 inline size_t getVeryLargeFrameSize()
6459 // The looping probe code is 40 bytes, whereas the straight-line probing for
6460 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6461 // or greater, to generate smaller code.
6462 return 2 * eeGetPageSize();
6464 return 3 * eeGetPageSize();
6468 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6470 #if COR_JIT_EE_VERSION > 460
6471 return eeGetEEInfo()->targetAbi == abi;
6473 return CORINFO_DESKTOP_ABI == abi;
6477 inline bool generateCFIUnwindCodes()
6479 #ifdef UNIX_AMD64_ABI
6480 return IsTargetAbi(CORINFO_CORERT_ABI);
6488 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6490 // Debugging support - Line number info
6492 void eeGetStmtOffsets();
6494 unsigned eeBoundariesCount;
6496 struct boundariesDsc
6498 UNATIVE_OFFSET nativeIP;
6500 unsigned sourceReason;
6501 } * eeBoundaries; // Boundaries to report to EE
6502 void eeSetLIcount(unsigned count);
6503 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6507 static void eeDispILOffs(IL_OFFSET offs);
6508 static void eeDispLineInfo(const boundariesDsc* line);
6509 void eeDispLineInfos();
6512 // Debugging support - Local var info
6516 unsigned eeVarsCount;
6518 struct VarResultInfo
6520 UNATIVE_OFFSET startOffset;
6521 UNATIVE_OFFSET endOffset;
6525 void eeSetLVcount(unsigned count);
6526 void eeSetLVinfo(unsigned which,
6527 UNATIVE_OFFSET startOffs,
6528 UNATIVE_OFFSET length,
6533 const siVarLoc& loc);
6537 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6538 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6541 // ICorJitInfo wrappers
6543 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6545 void eeAllocUnwindInfo(BYTE* pHotCode,
6551 CorJitFuncKind funcKind);
6553 void eeSetEHcount(unsigned cEH);
6555 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6557 WORD eeGetRelocTypeHint(void* target);
6559 // ICorStaticInfo wrapper functions
6561 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6563 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6565 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6568 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6569 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6570 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6571 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6573 template <typename ParamType>
6574 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6576 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6579 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6581 // Utility functions
6583 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6586 const wchar_t* eeGetCPString(size_t stringHandle);
6589 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6591 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6592 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6594 static fgWalkPreFn CountSharedStaticHelper;
6595 static bool IsSharedStaticHelper(GenTreePtr tree);
6596 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6598 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6599 // returns true/false if 'field' is a Jit Data offset
6600 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6601 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6602 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6604 /*****************************************************************************/
6609 enum TEMP_USAGE_TYPE
6615 static var_types tmpNormalizeType(var_types type);
6616 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6617 void tmpRlsTemp(TempDsc* temp);
6618 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6621 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6622 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6626 bool tmpAllFree() const;
6629 #ifndef LEGACY_BACKEND
6630 void tmpPreAllocateTemps(var_types type, unsigned count);
6631 #endif // !LEGACY_BACKEND
6634 #ifdef LEGACY_BACKEND
6635 unsigned tmpIntSpillMax; // number of int-sized spill temps
6636 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6637 #endif // LEGACY_BACKEND
6639 unsigned tmpCount; // Number of temps
6640 unsigned tmpSize; // Size of all the temps
6643 // Used by RegSet::rsSpillChk()
6644 unsigned tmpGetCount; // Temps which haven't been released yet
6647 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6649 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6650 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6653 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6654 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6658 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6659 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6663 CodeGenInterface* codeGen;
6665 // The following holds information about instr offsets in terms of generated code.
6669 IPmappingDsc* ipmdNext; // next line# record
6670 IL_OFFSETX ipmdILoffsx; // the instr offset
6671 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6672 bool ipmdIsLabel; // Can this code be a branch label?
6675 // Record the instr offset mapping to the generated code
6677 IPmappingDsc* genIPmappingList;
6678 IPmappingDsc* genIPmappingLast;
6680 // Managed RetVal - A side hash table meant to record the mapping from a
6681 // GT_CALL node to its IL offset. This info is used to emit sequence points
6682 // that can be used by debugger to determine the native offset at which the
6683 // managed RetVal will be available.
6685 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6686 // favor of a side table for two reasons: 1) We need IL offset for only those
6687 // GT_CALL nodes (created during importation) that correspond to an IL call and
6688 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6689 // structure and IL offset is needed only when generating debuggable code. Therefore
6690 // it is desirable to avoid memory size penalty in retail scenarios.
6691 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6692 CallSiteILOffsetTable;
6693 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6695 unsigned genReturnLocal; // Local number for the return value when applicable.
6696 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6698 // The following properties are part of CodeGenContext. Getters are provided here for
6699 // convenience and backward compatibility, but the properties can only be set by invoking
6700 // the setter on CodeGenContext directly.
6702 __declspec(property(get = getEmitter)) emitter* genEmitter;
6703 emitter* getEmitter()
6705 return codeGen->getEmitter();
6708 const bool isFramePointerUsed()
6710 return codeGen->isFramePointerUsed();
6713 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6714 bool getInterruptible()
6716 return codeGen->genInterruptible;
6718 void setInterruptible(bool value)
6720 codeGen->setInterruptible(value);
6724 const bool genDoubleAlign()
6726 return codeGen->doDoubleAlign();
6728 DWORD getCanDoubleAlign();
6729 bool shouldDoubleAlign(unsigned refCntStk,
6731 unsigned refCntWtdReg,
6732 unsigned refCntStkParam,
6733 unsigned refCntWtdStkDbl);
6734 #endif // DOUBLE_ALIGN
6736 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6737 bool getFullPtrRegMap()
6739 return codeGen->genFullPtrRegMap;
6741 void setFullPtrRegMap(bool value)
6743 codeGen->setFullPtrRegMap(value);
6746 // Things that MAY belong either in CodeGen or CodeGenContext
6748 #if FEATURE_EH_FUNCLETS
6749 FuncInfoDsc* compFuncInfos;
6750 unsigned short compCurrFuncIdx;
6751 unsigned short compFuncInfoCount;
6753 unsigned short compFuncCount()
6755 assert(fgFuncletsCreated);
6756 return compFuncInfoCount;
6759 #else // !FEATURE_EH_FUNCLETS
6761 // This is a no-op when there are no funclets!
6762 void genUpdateCurrentFunclet(BasicBlock* block)
6767 FuncInfoDsc compFuncInfoRoot;
6769 static const unsigned compCurrFuncIdx = 0;
6771 unsigned short compFuncCount()
6776 #endif // !FEATURE_EH_FUNCLETS
6778 FuncInfoDsc* funCurrentFunc();
6779 void funSetCurrentFunc(unsigned funcIdx);
6780 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6781 unsigned int funGetFuncIdx(BasicBlock* block);
6785 VARSET_TP compCurLife; // current live variables
6786 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6788 template <bool ForCodeGen>
6789 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6791 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6793 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6796 template <bool ForCodeGen>
6797 void compUpdateLife(GenTreePtr tree);
6799 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6800 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6801 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6802 template <bool ForCodeGen>
6803 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6805 template <bool ForCodeGen>
6806 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6808 // Gets a register mask that represent the kill set for a helper call since
6809 // not all JIT Helper calls follow the standard ABI on the target architecture.
6810 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6812 // Gets a register mask that represent the kill set for a NoGC helper call.
6813 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6816 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6817 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6818 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6819 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6820 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6821 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6822 #endif // _TARGET_ARM_
6824 // 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
6826 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6828 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6829 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6830 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6831 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6832 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6833 // for the tracked var indices of the field vars, as in a live var set).
6834 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6836 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6838 if (m_promotedStructDeathVars == nullptr)
6840 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6842 return m_promotedStructDeathVars;
6846 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6847 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6851 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6852 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6855 #if !defined(__GNUC__)
6856 #pragma region Unwind information
6861 // Infrastructure functions: start/stop/reserve/emit.
6864 void unwindBegProlog();
6865 void unwindEndProlog();
6866 void unwindBegEpilog();
6867 void unwindEndEpilog();
6868 void unwindReserve();
6869 void unwindEmit(void* pHotCode, void* pColdCode);
6872 // Specific unwind information functions: called by code generation to indicate a particular
6873 // prolog or epilog unwindable instruction has been generated.
6876 void unwindPush(regNumber reg);
6877 void unwindAllocStack(unsigned size);
6878 void unwindSetFrameReg(regNumber reg, unsigned offset);
6879 void unwindSaveReg(regNumber reg, unsigned offset);
6881 #if defined(_TARGET_ARM_)
6882 void unwindPushMaskInt(regMaskTP mask);
6883 void unwindPushMaskFloat(regMaskTP mask);
6884 void unwindPopMaskInt(regMaskTP mask);
6885 void unwindPopMaskFloat(regMaskTP mask);
6886 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6887 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6888 // called via unwindPadding().
6889 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6890 // instruction and the current location.
6891 #endif // _TARGET_ARM_
6893 #if defined(_TARGET_ARM64_)
6895 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6896 // instruction and the current location.
6897 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6898 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6899 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6900 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6901 void unwindSaveNext(); // unwind code: save_next
6902 void unwindReturn(regNumber reg); // ret lr
6903 #endif // defined(_TARGET_ARM64_)
6906 // Private "helper" functions for the unwind implementation.
6910 #if FEATURE_EH_FUNCLETS
6911 void unwindGetFuncLocations(FuncInfoDsc* func,
6912 bool getHotSectionData,
6913 /* OUT */ emitLocation** ppStartLoc,
6914 /* OUT */ emitLocation** ppEndLoc);
6915 #endif // FEATURE_EH_FUNCLETS
6917 void unwindReserveFunc(FuncInfoDsc* func);
6918 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6920 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6922 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6923 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6925 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6927 #if defined(_TARGET_AMD64_)
6929 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6931 void unwindBegPrologWindows();
6932 void unwindPushWindows(regNumber reg);
6933 void unwindAllocStackWindows(unsigned size);
6934 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6935 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6937 #ifdef UNIX_AMD64_ABI
6938 void unwindBegPrologCFI();
6939 void unwindPushCFI(regNumber reg);
6940 void unwindAllocStackCFI(unsigned size);
6941 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6942 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6943 int mapRegNumToDwarfReg(regNumber reg);
6944 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6945 #endif // UNIX_AMD64_ABI
6946 #elif defined(_TARGET_ARM_)
6948 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6949 void unwindPushPopMaskFloat(regMaskTP mask);
6950 void unwindSplit(FuncInfoDsc* func);
6952 #endif // _TARGET_ARM_
6954 #if !defined(__GNUC__)
6955 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6959 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6960 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6964 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6965 XX that contains the distinguished, well-known SIMD type definitions). XX
6967 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6968 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6971 // Get highest available instruction set for floating point codegen
6972 InstructionSet getFloatingPointInstructionSet()
6974 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6977 return InstructionSet_AVX;
6982 return InstructionSet_SSE3_4;
6986 assert(canUseSSE2());
6987 return InstructionSet_SSE2;
6989 assert(!"getFPInstructionSet() is not implemented for target arch");
6991 return InstructionSet_NONE;
6995 // Get highest available instruction set for SIMD codegen
6996 InstructionSet getSIMDInstructionSet()
6998 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6999 return getFloatingPointInstructionSet();
7001 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7003 return InstructionSet_NONE;
7009 // Should we support SIMD intrinsics?
7012 // Have we identified any SIMD types?
7013 // This is currently used by struct promotion to avoid getting type information for a struct
7014 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7016 bool _usesSIMDTypes;
7017 bool usesSIMDTypes()
7019 return _usesSIMDTypes;
7021 void setUsesSIMDTypes(bool value)
7023 _usesSIMDTypes = value;
7026 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7027 // that require indexed access to the individual fields of the vector, which is not well supported
7028 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7029 unsigned lvaSIMDInitTempVarNum;
7032 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7033 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7034 CORINFO_CLASS_HANDLE SIMDIntHandle;
7035 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7036 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7037 CORINFO_CLASS_HANDLE SIMDShortHandle;
7038 CORINFO_CLASS_HANDLE SIMDByteHandle;
7039 CORINFO_CLASS_HANDLE SIMDLongHandle;
7040 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7041 CORINFO_CLASS_HANDLE SIMDULongHandle;
7042 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7043 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7044 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7045 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7047 // Get the handle for a SIMD type.
7048 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7050 if (simdBaseType == TYP_FLOAT)
7055 return SIMDVector2Handle;
7057 return SIMDVector3Handle;
7059 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7061 return SIMDVector4Handle;
7070 assert(simdType == getSIMDVectorType());
7071 switch (simdBaseType)
7074 return SIMDFloatHandle;
7076 return SIMDDoubleHandle;
7078 return SIMDIntHandle;
7080 return SIMDUShortHandle;
7082 return SIMDUShortHandle;
7084 return SIMDUByteHandle;
7086 return SIMDShortHandle;
7088 return SIMDByteHandle;
7090 return SIMDLongHandle;
7092 return SIMDUIntHandle;
7094 return SIMDULongHandle;
7096 assert(!"Didn't find a class handle for simdType");
7098 return NO_CLASS_HANDLE;
7102 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7103 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7104 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7106 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7107 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7108 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7109 bool isSIMDTypeLocal(GenTree* tree)
7111 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7114 // Returns true if the type of the tree is a byref of TYP_SIMD
7115 bool isAddrOfSIMDType(GenTree* tree)
7117 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7119 switch (tree->OperGet())
7122 return varTypeIsSIMD(tree->gtGetOp1());
7124 case GT_LCL_VAR_ADDR:
7125 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7128 return isSIMDTypeLocal(tree);
7135 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7137 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7138 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7139 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7142 // Returns base type of a TYP_SIMD local.
7143 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7144 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7146 if (isSIMDTypeLocal(tree))
7148 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7154 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7156 return info.compCompHnd->isInSIMDModule(clsHnd);
7159 bool isSIMDClass(typeInfo* pTypeInfo)
7161 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7164 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7165 // if it is not a SIMD type or is an unsupported base type.
7166 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7168 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7170 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7173 // Get SIMD Intrinsic info given the method handle.
7174 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7175 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7176 CORINFO_METHOD_HANDLE methodHnd,
7177 CORINFO_SIG_INFO* sig,
7180 var_types* baseType,
7181 unsigned* sizeBytes);
7183 // Pops and returns GenTree node from importers type stack.
7184 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7185 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7187 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7188 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7190 // Creates a GT_SIMD tree for Select operation
7191 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7193 unsigned simdVectorSize,
7198 // Creates a GT_SIMD tree for Min/Max operation
7199 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7200 CORINFO_CLASS_HANDLE typeHnd,
7202 unsigned simdVectorSize,
7206 // Transforms operands and returns the SIMD intrinsic to be applied on
7207 // transformed operands to obtain given relop result.
7208 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7209 CORINFO_CLASS_HANDLE typeHnd,
7210 unsigned simdVectorSize,
7211 var_types* baseType,
7215 // Creates a GT_SIMD tree for Abs intrinsic.
7216 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7218 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7219 // Transforms operands and returns the SIMD intrinsic to be applied on
7220 // transformed operands to obtain == comparison result.
7221 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7222 unsigned simdVectorSize,
7226 // Transforms operands and returns the SIMD intrinsic to be applied on
7227 // transformed operands to obtain > comparison result.
7228 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7229 unsigned simdVectorSize,
7233 // Transforms operands and returns the SIMD intrinsic to be applied on
7234 // transformed operands to obtain >= comparison result.
7235 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7236 unsigned simdVectorSize,
7240 // Transforms operands and returns the SIMD intrinsic to be applied on
7241 // transformed operands to obtain >= comparison result in case of int32
7242 // and small int base type vectors.
7243 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7244 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7245 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7247 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7248 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7249 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7250 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7251 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7253 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7254 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7255 GenTreePtr newobjThis,
7256 CORINFO_CLASS_HANDLE clsHnd,
7257 CORINFO_METHOD_HANDLE method,
7258 CORINFO_SIG_INFO* sig,
7261 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7263 // Whether SIMD vector occupies part of SIMD register.
7264 // SSE2: vector2f/3f are considered sub register SIMD types.
7265 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7266 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7268 unsigned sizeBytes = 0;
7269 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7270 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7273 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7275 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7278 // Get the type for the hardware SIMD vector.
7279 // This is the maximum SIMD type supported for this target.
7280 var_types getSIMDVectorType()
7282 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7289 assert(canUseSSE2());
7293 assert(!"getSIMDVectorType() unimplemented on target arch");
7298 // Get the size of the SIMD type in bytes
7299 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7301 unsigned sizeBytes = 0;
7302 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7306 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7307 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7309 // Get the the number of elements of basetype of SIMD vector given by its type handle
7310 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7312 // Get preferred alignment of SIMD type.
7313 int getSIMDTypeAlignment(var_types simdType);
7315 // Get the number of bytes in a SIMD Vector for the current compilation.
7316 unsigned getSIMDVectorRegisterByteLength()
7318 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7321 return YMM_REGSIZE_BYTES;
7325 assert(canUseSSE2());
7326 return XMM_REGSIZE_BYTES;
7329 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7334 // The minimum and maximum possible number of bytes in a SIMD vector.
7335 unsigned int maxSIMDStructBytes()
7337 return getSIMDVectorRegisterByteLength();
7339 unsigned int minSIMDStructBytes()
7341 return emitTypeSize(TYP_SIMD8);
7344 #ifdef FEATURE_AVX_SUPPORT
7345 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7346 static const unsigned maxPossibleSIMDStructBytes = 32;
7347 #else // !FEATURE_AVX_SUPPORT
7348 static const unsigned maxPossibleSIMDStructBytes = 16;
7349 #endif // !FEATURE_AVX_SUPPORT
7351 // Returns the codegen type for a given SIMD size.
7352 var_types getSIMDTypeForSize(unsigned size)
7354 var_types simdType = TYP_UNDEF;
7357 simdType = TYP_SIMD8;
7359 else if (size == 12)
7361 simdType = TYP_SIMD12;
7363 else if (size == 16)
7365 simdType = TYP_SIMD16;
7367 #ifdef FEATURE_AVX_SUPPORT
7368 else if (size == 32)
7370 simdType = TYP_SIMD32;
7372 #endif // FEATURE_AVX_SUPPORT
7375 noway_assert(!"Unexpected size for SIMD type");
7380 unsigned getSIMDInitTempVarNum()
7382 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7384 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7385 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7387 return lvaSIMDInitTempVarNum;
7390 #endif // FEATURE_SIMD
7393 //------------------------------------------------------------------------
7394 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7396 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7397 // candidate for enregistration.
7399 unsigned largestEnregisterableStructSize()
7402 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7403 if (vectorRegSize > TARGET_POINTER_SIZE)
7405 return vectorRegSize;
7408 #endif // FEATURE_SIMD
7410 return TARGET_POINTER_SIZE;
7415 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7416 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7417 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7419 // Is this var is of type simd struct?
7420 bool lclVarIsSIMDType(unsigned varNum)
7422 LclVarDsc* varDsc = lvaTable + varNum;
7423 return varDsc->lvIsSIMDType();
7426 // Is this Local node a SIMD local?
7427 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7429 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7432 // Returns true if the TYP_SIMD locals on stack are aligned at their
7433 // preferred byte boundary specified by getSIMDTypeAlignment().
7435 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7436 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7437 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7438 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7439 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7440 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7441 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7444 bool isSIMDTypeLocalAligned(unsigned varNum)
7446 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7447 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7450 int off = lvaFrameAddress(varNum, &ebpBased);
7451 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7452 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7453 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7456 #endif // FEATURE_SIMD
7461 // Whether SSE2 is available
7462 bool canUseSSE2() const
7464 #ifdef _TARGET_XARCH_
7465 return opts.compCanUseSSE2;
7471 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7472 bool CanUseSSE3_4() const
7474 #ifdef _TARGET_XARCH_
7475 return opts.compCanUseSSE3_4;
7481 bool canUseAVX() const
7483 #ifdef FEATURE_AVX_SUPPORT
7484 return opts.compCanUseAVX;
7491 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7492 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7496 XX Generic info about the compilation and the method being compiled. XX
7497 XX It is responsible for driving the other phases. XX
7498 XX It is also responsible for all the memory management. XX
7500 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7501 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7505 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7507 InlineResult* compInlineResult; // The result of importing the inlinee method.
7509 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7510 bool compJmpOpUsed; // Does the method do a JMP
7511 bool compLongUsed; // Does the method use TYP_LONG
7512 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7513 bool compTailCallUsed; // Does the method do a tailcall
7514 bool compLocallocUsed; // Does the method use localloc.
7515 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7516 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7517 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7519 // NOTE: These values are only reliable after
7520 // the importing is completely finished.
7522 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7523 // we can iterate over these efficiently.
7525 #if CPU_USES_BLOCK_MOVE
7526 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7530 // State information - which phases have completed?
7531 // These are kept together for easy discoverability
7533 bool bRangeAllowStress;
7534 bool compCodeGenDone;
7535 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7536 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7537 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7538 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7541 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7542 bool fgLocalVarLivenessChanged;
7544 bool compStackProbePrologDone;
7546 #ifndef LEGACY_BACKEND
7548 #endif // !LEGACY_BACKEND
7549 bool compRationalIRForm;
7551 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7553 bool compGeneratingProlog;
7554 bool compGeneratingEpilog;
7555 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7556 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7557 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7558 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7559 bool getNeedsGSSecurityCookie() const
7561 return compNeedsGSSecurityCookie;
7563 void setNeedsGSSecurityCookie()
7565 compNeedsGSSecurityCookie = true;
7568 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7569 // frame layout calculations, this is the level we are currently
7572 //---------------------------- JITing options -----------------------------
7585 JitFlags* jitFlags; // all flags passed from the EE
7586 unsigned compFlags; // method attributes
7588 codeOptimize compCodeOpt; // what type of code optimizations
7592 #ifdef _TARGET_XARCH_
7593 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7594 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7596 #ifdef FEATURE_AVX_SUPPORT
7597 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7598 #endif // FEATURE_AVX_SUPPORT
7599 #endif // _TARGET_XARCH_
7601 // optimize maximally and/or favor speed over size?
7603 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7604 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7605 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7606 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7607 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7609 // Maximun number of locals before turning off the inlining
7610 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7613 unsigned instrCount;
7614 unsigned lvRefCount;
7615 bool compMinOptsIsSet;
7617 bool compMinOptsIsUsed;
7619 inline bool MinOpts()
7621 assert(compMinOptsIsSet);
7622 compMinOptsIsUsed = true;
7625 inline bool IsMinOptsSet()
7627 return compMinOptsIsSet;
7630 inline bool MinOpts()
7634 inline bool IsMinOptsSet()
7636 return compMinOptsIsSet;
7639 inline void SetMinOpts(bool val)
7641 assert(!compMinOptsIsUsed);
7642 assert(!compMinOptsIsSet || (compMinOpts == val));
7644 compMinOptsIsSet = true;
7647 // true if the CLFLG_* for an optimization is set.
7648 inline bool OptEnabled(unsigned optFlag)
7650 return !!(compFlags & optFlag);
7653 #ifdef FEATURE_READYTORUN_COMPILER
7654 inline bool IsReadyToRun()
7656 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7659 inline bool IsReadyToRun()
7665 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7666 // PInvoke transitions inline (e.g. when targeting CoreRT).
7667 inline bool ShouldUsePInvokeHelpers()
7669 #if COR_JIT_EE_VERSION > 460
7670 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7676 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7678 inline bool IsReversePInvoke()
7680 #if COR_JIT_EE_VERSION > 460
7681 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7687 // true if we must generate code compatible with JIT32 quirks
7688 inline bool IsJit32Compat()
7690 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7691 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7697 // true if we must generate code compatible with Jit64 quirks
7698 inline bool IsJit64Compat()
7700 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7701 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7702 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7709 bool compScopeInfo; // Generate the LocalVar info ?
7710 bool compDbgCode; // Generate debugger-friendly code?
7711 bool compDbgInfo; // Gather debugging info?
7714 #ifdef PROFILING_SUPPORTED
7715 bool compNoPInvokeInlineCB;
7717 static const bool compNoPInvokeInlineCB;
7721 bool compGcChecks; // Check arguments and return values to ensure they are sane
7722 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7723 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7727 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7728 // to be allocated on the stack.
7729 // It will be set to true in the following cases:
7730 // 1. When the method being compiled has a declarative security
7731 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7732 // This is also the case when we inject a prolog and epilog in the method.
7734 // 2. When the method being compiled has imperative security (i.e. the method
7735 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7737 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7739 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7740 // which gets reported as a GC root to stackwalker.
7741 // (See also ICodeManager::GetAddrOfSecurityObject.)
7748 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7749 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7753 #ifdef UNIX_AMD64_ABI
7754 // This flag is indicating if there is a need to align the frame.
7755 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7756 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7757 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7758 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7759 // there are calls and making sure the frame alignment logic is executed.
7760 bool compNeedToAlignFrame;
7761 #endif // UNIX_AMD64_ABI
7763 bool compProcedureSplitting; // Separate cold code from hot code
7765 bool genFPorder; // Preserve FP order (operations are non-commutative)
7766 bool genFPopt; // Can we do frame-pointer-omission optimization?
7767 bool altJit; // True if we are an altjit and are compiling this method
7770 bool optRepeat; // Repeat optimizer phases k times
7771 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7772 bool dspCode; // Display native code generated
7773 bool dspEHTable; // Display the EH table reported to the VM
7774 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7775 bool dspEmit; // Display emitter output
7776 bool dspLines; // Display source-code lines intermixed with native code output
7777 bool dmpHex; // Display raw bytes in hex of native code output
7778 bool varNames; // Display variables names in native code output
7779 bool disAsm; // Display native code as it is generated
7780 bool disAsmSpilled; // Display native code when any register spilling occurs
7781 bool disDiffable; // Makes the Disassembly code 'diff-able'
7782 bool disAsm2; // Display native code after it is generated using external disassembler
7783 bool dspOrder; // Display names of each of the methods that we ngen/jit
7784 bool dspUnwind; // Display the unwind info output
7785 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7786 bool compLongAddress; // Force using large pseudo instructions for long address
7787 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7788 bool dspGCtbls; // Display the GC tables
7792 bool doLateDisasm; // Run the late disassembler
7793 #endif // LATE_DISASM
7795 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7796 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7797 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7798 static const bool dspGCtbls = true;
7801 // We need stack probes to guarantee that we won't trigger a stack overflow
7802 // when calling unmanaged code until they get a chance to set up a frame, because
7803 // the EE will have no idea where it is.
7805 // We will only be doing this currently for hosted environments. Unfortunately
7806 // we need to take care of stubs, so potentially, we will have to do the probes
7807 // for any call. We have a plan for not needing for stubs though
7808 bool compNeedStackProbes;
7810 #ifdef PROFILING_SUPPORTED
7811 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7812 // This option helps make the JIT behave as if it is running under a profiler.
7813 bool compJitELTHookEnabled;
7814 #endif // PROFILING_SUPPORTED
7816 #if FEATURE_TAILCALL_OPT
7817 // Whether opportunistic or implicit tail call optimization is enabled.
7818 bool compTailCallOpt;
7819 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7820 bool compTailCallLoopOpt;
7824 static const bool compUseSoftFP = true;
7825 #else // !ARM_SOFTFP
7826 static const bool compUseSoftFP = false;
7829 GCPollType compGCPollType;
7833 static bool s_pAltJitExcludeAssembliesListInitialized;
7834 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7839 template <typename T>
7842 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7845 template <typename T>
7848 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7851 static int dspTreeID(GenTree* tree)
7853 return tree->gtTreeID;
7855 static void printTreeID(GenTree* tree)
7857 if (tree == nullptr)
7863 printf("[%06d]", dspTreeID(tree));
7870 #define STRESS_MODES \
7874 /* "Variations" stress areas which we try to mix up with each other. */ \
7875 /* These should not be exhaustively used as they might */ \
7876 /* hide/trivialize other areas */ \
7878 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7879 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7880 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7881 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7882 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7883 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7884 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7885 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7886 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7887 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7888 STRESS_MODE(NULL_OBJECT_CHECK) \
7889 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7890 STRESS_MODE(RANDOM_INLINE) \
7892 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7894 /* "Check" stress areas that can be exhaustively used if we */ \
7895 /* dont care about performance at all */ \
7897 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7898 STRESS_MODE(CHK_FLOW_UPDATE) \
7899 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7901 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7905 #define STRESS_MODE(mode) STRESS_##mode,
7912 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7913 BYTE compActiveStressModes[STRESS_COUNT];
7916 #define MAX_STRESS_WEIGHT 100
7918 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7922 bool compInlineStress()
7924 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7927 bool compRandomInlineStress()
7929 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7934 bool compTailCallStress()
7937 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7943 codeOptimize compCodeOpt()
7946 // Switching between size & speed has measurable throughput impact
7947 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7948 // DEBUG, but should generate identical code between CHK & RET builds,
7949 // so that's not acceptable.
7950 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7951 // Investigate the cause of the throughput regression.
7953 return opts.compCodeOpt;
7955 return BLENDED_CODE;
7959 //--------------------- Info about the procedure --------------------------
7963 COMP_HANDLE compCompHnd;
7964 CORINFO_MODULE_HANDLE compScopeHnd;
7965 CORINFO_CLASS_HANDLE compClassHnd;
7966 CORINFO_METHOD_HANDLE compMethodHnd;
7967 CORINFO_METHOD_INFO* compMethodInfo;
7969 BOOL hasCircularClassConstraints;
7970 BOOL hasCircularMethodConstraints;
7972 #if defined(DEBUG) || defined(LATE_DISASM)
7973 const char* compMethodName;
7974 const char* compClassName;
7975 const char* compFullName;
7976 #endif // defined(DEBUG) || defined(LATE_DISASM)
7978 #if defined(DEBUG) || defined(INLINE_DATA)
7979 // Method hash is logcally const, but computed
7981 mutable unsigned compMethodHashPrivate;
7982 unsigned compMethodHash() const;
7983 #endif // defined(DEBUG) || defined(INLINE_DATA)
7985 #ifdef PSEUDORANDOM_NOP_INSERTION
7986 // things for pseudorandom nop insertion
7987 unsigned compChecksum;
7991 // The following holds the FLG_xxxx flags for the method we're compiling.
7994 // The following holds the class attributes for the method we're compiling.
7995 unsigned compClassAttr;
7997 const BYTE* compCode;
7998 IL_OFFSET compILCodeSize; // The IL code size
7999 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8000 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8001 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8002 // (2) the code is hot/cold split, and we issued less code than we expected
8003 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8005 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8006 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8007 bool compIsContextful : 1; // contextful method
8008 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8009 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8010 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8011 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8012 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8014 var_types compRetType; // Return type of the method as declared in IL
8015 var_types compRetNativeType; // Normalized return type as per target arch ABI
8016 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8017 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8018 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8019 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8020 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8021 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8022 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8023 unsigned compMaxStack;
8024 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8025 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8027 unsigned compCallUnmanaged; // count of unmanaged calls
8028 unsigned compLvFrameListRoot; // lclNum for the Frame root
8029 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8030 // You should generally use compHndBBtabCount instead: it is the
8031 // current number of EH clauses (after additions like synchronized
8032 // methods and funclets, and removals like unreachable code deletion).
8034 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8035 // and the VM expects that, or the JIT is a "self-host" compiler
8036 // (e.g., x86 hosted targeting x86) and the VM expects that.
8038 /* The following holds IL scope information about local variables.
8041 unsigned compVarScopesCount;
8042 VarScopeDsc* compVarScopes;
8044 /* The following holds information about instr offsets for
8045 * which we need to report IP-mappings
8048 IL_OFFSET* compStmtOffsets; // sorted
8049 unsigned compStmtOffsetsCount;
8050 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8052 #define CPU_X86 0x0100 // The generic X86 CPU
8053 #define CPU_X86_PENTIUM_4 0x0110
8055 #define CPU_X64 0x0200 // The generic x64 CPU
8056 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8057 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8059 #define CPU_ARM 0x0300 // The generic ARM CPU
8061 unsigned genCPU; // What CPU are we running on
8064 // Returns true if the method being compiled returns a non-void and non-struct value.
8065 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8066 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8067 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8068 // Methods returning such structs are considered to return non-struct return value and
8069 // this method returns true in that case.
8070 bool compMethodReturnsNativeScalarType()
8072 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8075 // Returns true if the method being compiled returns RetBuf addr as its return value
8076 bool compMethodReturnsRetBufAddr()
8078 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8079 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8081 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8082 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8083 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8084 // methods with hidden RetBufArg.
8086 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8087 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8088 // returning the address of RetBuf.
8090 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8091 // to be returned in RAX.
8092 CLANG_FORMAT_COMMENT_ANCHOR;
8094 #ifdef _TARGET_AMD64_
8095 return (info.compRetBuffArg != BAD_VAR_NUM);
8096 #else // !_TARGET_AMD64_
8097 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8098 #endif // !_TARGET_AMD64_
8101 // Returns true if the method returns a value in more than one return register
8102 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8103 // TODO-ARM64: Does this apply for ARM64 too?
8104 bool compMethodReturnsMultiRegRetType()
8106 #if FEATURE_MULTIREG_RET
8107 #if defined(_TARGET_X86_)
8108 // On x86 only 64-bit longs are returned in multiple registers
8109 return varTypeIsLong(info.compRetNativeType);
8110 #else // targets: X64-UNIX, ARM64 or ARM32
8111 // On all other targets that support multireg return values:
8112 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8113 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8114 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8115 #endif // TARGET_XXX
8117 #else // not FEATURE_MULTIREG_RET
8119 // For this architecture there are no multireg returns
8122 #endif // FEATURE_MULTIREG_RET
8125 #if FEATURE_MULTIREG_ARGS
8126 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8127 // return the gcPtr layout for the pointers sized fields
8128 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8129 #endif // FEATURE_MULTIREG_ARGS
8131 // Returns true if the method being compiled returns a value
8132 bool compMethodHasRetVal()
8134 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8135 compMethodReturnsMultiRegRetType();
8140 void compDispLocalVars();
8144 //-------------------------- Global Compiler Data ------------------------------------
8147 static unsigned s_compMethodsCount; // to produce unique label names
8148 unsigned compGenTreeID;
8151 BasicBlock* compCurBB; // the current basic block in process
8152 GenTreePtr compCurStmt; // the current statement in process
8154 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8157 // The following is used to create the 'method JIT info' block.
8158 size_t compInfoBlkSize;
8159 BYTE* compInfoBlkAddr;
8161 EHblkDsc* compHndBBtab; // array of EH data
8162 unsigned compHndBBtabCount; // element count of used elements in EH data array
8163 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8165 #if defined(_TARGET_X86_)
8167 //-------------------------------------------------------------------------
8168 // Tracking of region covered by the monitor in synchronized methods
8169 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8170 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8172 #endif // !_TARGET_X86_
8174 Phases previousCompletedPhase; // the most recently completed phase
8176 //-------------------------------------------------------------------------
8177 // The following keeps track of how many bytes of local frame space we've
8178 // grabbed so far in the current function, and how many argument bytes we
8179 // need to pop when we return.
8182 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8184 // Count of callee-saved regs we pushed in the prolog.
8185 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8186 // In case of Amd64 this doesn't include float regs saved on stack.
8187 unsigned compCalleeRegsPushed;
8189 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8190 // Mask of callee saved float regs on stack.
8191 regMaskTP compCalleeFPRegsSavedMask;
8193 #ifdef _TARGET_AMD64_
8194 // Quirk for VS debug-launch scenario to work:
8195 // Bytes of padding between save-reg area and locals.
8196 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8197 unsigned compVSQuirkStackPaddingNeeded;
8198 bool compQuirkForPPPflag;
8201 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8203 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8204 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8205 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8207 //-------------------------------------------------------------------------
8209 static void compStartup(); // One-time initialization
8210 static void compShutdown(); // One-time finalization
8212 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8215 static void compDisplayStaticSizes(FILE* fout);
8217 //------------ Some utility functions --------------
8219 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8220 void** ppIndirection); /* OUT */
8222 // Several JIT/EE interface functions return a CorInfoType, and also return a
8223 // class handle as an out parameter if the type is a value class. Returns the
8224 // size of the type these describe.
8225 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8228 // Components used by the compiler may write unit test suites, and
8229 // have them run within this method. They will be run only once per process, and only
8230 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8231 // These should fail by asserting.
8232 void compDoComponentUnitTestsOnce();
8235 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8236 CORINFO_MODULE_HANDLE classPtr,
8237 COMP_HANDLE compHnd,
8238 CORINFO_METHOD_INFO* methodInfo,
8239 void** methodCodePtr,
8240 ULONG* methodCodeSize,
8241 JitFlags* compileFlags);
8242 void compCompileFinish();
8243 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8244 COMP_HANDLE compHnd,
8245 CORINFO_METHOD_INFO* methodInfo,
8246 void** methodCodePtr,
8247 ULONG* methodCodeSize,
8248 JitFlags* compileFlags,
8249 CorInfoInstantiationVerification instVerInfo);
8251 ArenaAllocator* compGetAllocator();
8253 #if MEASURE_MEM_ALLOC
8255 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8259 unsigned allocCnt; // # of allocs
8260 UINT64 allocSz; // total size of those alloc.
8261 UINT64 allocSzMax; // Maximum single allocation.
8262 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8263 UINT64 nraTotalSizeAlloc;
8264 UINT64 nraTotalSizeUsed;
8266 static const char* s_CompMemKindNames[]; // Names of the kinds.
8268 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8270 for (int i = 0; i < CMK_Count; i++)
8272 allocSzByKind[i] = 0;
8275 MemStats(const MemStats& ms)
8276 : allocCnt(ms.allocCnt)
8277 , allocSz(ms.allocSz)
8278 , allocSzMax(ms.allocSzMax)
8279 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8280 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8282 for (int i = 0; i < CMK_Count; i++)
8284 allocSzByKind[i] = ms.allocSzByKind[i];
8288 // Until we have ubiquitous constructors.
8291 this->MemStats::MemStats();
8294 void AddAlloc(size_t sz, CompMemKind cmk)
8298 if (sz > allocSzMax)
8302 allocSzByKind[cmk] += sz;
8305 void Print(FILE* f); // Print these stats to f.
8306 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8308 MemStats genMemStats;
8310 struct AggregateMemStats : public MemStats
8314 AggregateMemStats() : MemStats(), nMethods(0)
8318 void Add(const MemStats& ms)
8321 allocCnt += ms.allocCnt;
8322 allocSz += ms.allocSz;
8323 allocSzMax = max(allocSzMax, ms.allocSzMax);
8324 for (int i = 0; i < CMK_Count; i++)
8326 allocSzByKind[i] += ms.allocSzByKind[i];
8328 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8329 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8332 void Print(FILE* f); // Print these stats to jitstdout.
8335 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8336 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8337 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8339 #endif // MEASURE_MEM_ALLOC
8341 #if LOOP_HOIST_STATS
8342 unsigned m_loopsConsidered;
8343 bool m_curLoopHasHoistedExpression;
8344 unsigned m_loopsWithHoistedExpressions;
8345 unsigned m_totalHoistedExpressions;
8347 void AddLoopHoistStats();
8348 void PrintPerMethodLoopHoistStats();
8350 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8351 static unsigned s_loopsConsidered;
8352 static unsigned s_loopsWithHoistedExpressions;
8353 static unsigned s_totalHoistedExpressions;
8355 static void PrintAggregateLoopHoistStats(FILE* f);
8356 #endif // LOOP_HOIST_STATS
8358 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8359 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8360 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8361 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8362 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8363 void compFreeMem(void*);
8365 bool compIsForImportOnly();
8366 bool compIsForInlining();
8367 bool compDonotInline();
8370 const char* compLocalVarName(unsigned varNum, unsigned offs);
8371 VarName compVarName(regNumber reg, bool isFloatReg = false);
8372 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8373 const char* compRegPairName(regPairNo regPair);
8374 const char* compRegNameForSize(regNumber reg, size_t size);
8375 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8376 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8377 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8380 //-------------------------------------------------------------------------
8382 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8384 struct VarScopeMapInfo
8386 VarScopeListNode* head;
8387 VarScopeListNode* tail;
8388 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8390 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8397 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8398 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8400 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8401 VarNumToScopeDscMap;
8403 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8404 VarNumToScopeDscMap* compVarScopeMap;
8406 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8408 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8410 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8412 void compInitVarScopeMap();
8414 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8415 // enter scope, sorted by instr offset
8416 unsigned compNextEnterScope;
8418 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8419 // go out of scope, sorted by instr offset
8420 unsigned compNextExitScope;
8422 void compInitScopeLists();
8424 void compResetScopeLists();
8426 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8428 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8430 void compProcessScopesUntil(unsigned offset,
8432 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8433 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8436 void compDispScopeLists();
8439 bool compIsProfilerHookNeeded();
8441 //-------------------------------------------------------------------------
8442 /* Statistical Data Gathering */
8444 void compJitStats(); // call this function and enable
8445 // various ifdef's below for statistical data
8448 void compCallArgStats();
8449 static void compDispCallArgStats(FILE* fout);
8452 //-------------------------------------------------------------------------
8459 ArenaAllocator* compAllocator;
8462 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8463 // suitable for use by utilcode collection types.
8464 IAllocator* compAsIAllocator;
8466 #if MEASURE_MEM_ALLOC
8467 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8468 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8469 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8471 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8473 #endif // MEASURE_MEM_ALLOC
8475 void compFunctionTraceStart();
8476 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8479 size_t compMaxUncheckedOffsetForNullObject;
8481 void compInitOptions(JitFlags* compileFlags);
8483 void compSetProcessor();
8484 void compInitDebuggingInfo();
8485 void compSetOptimizationLevel();
8486 #ifdef _TARGET_ARMARCH_
8487 bool compRsvdRegCheck(FrameLayoutState curState);
8489 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8491 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8492 void ResetOptAnnotations();
8494 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8495 void RecomputeLoopInfo();
8497 #ifdef PROFILING_SUPPORTED
8498 // Data required for generating profiler Enter/Leave/TailCall hooks
8500 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8501 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8502 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8505 #ifdef _TARGET_AMD64_
8506 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8509 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8510 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8512 IAllocator* getAllocator()
8514 return compAsIAllocator;
8517 #if MEASURE_MEM_ALLOC
8518 IAllocator* getAllocatorBitset()
8520 return compAsIAllocatorBitset;
8522 IAllocator* getAllocatorGC()
8524 return compAsIAllocatorGC;
8526 IAllocator* getAllocatorLoopHoist()
8528 return compAsIAllocatorLoopHoist;
8530 #else // !MEASURE_MEM_ALLOC
8531 IAllocator* getAllocatorBitset()
8533 return compAsIAllocator;
8535 IAllocator* getAllocatorGC()
8537 return compAsIAllocator;
8539 IAllocator* getAllocatorLoopHoist()
8541 return compAsIAllocator;
8543 #endif // !MEASURE_MEM_ALLOC
8546 IAllocator* getAllocatorDebugOnly()
8548 #if MEASURE_MEM_ALLOC
8549 return compAsIAllocatorDebugOnly;
8550 #else // !MEASURE_MEM_ALLOC
8551 return compAsIAllocator;
8552 #endif // !MEASURE_MEM_ALLOC
8557 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8558 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8562 XX Checks for type compatibility and merges types XX
8564 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8565 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8569 // Set to TRUE if verification cannot be skipped for this method
8570 // If we detect unverifiable code, we will lazily check
8571 // canSkipMethodVerification() to see if verification is REALLY needed.
8572 BOOL tiVerificationNeeded;
8574 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8575 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8576 BOOL tiIsVerifiableCode;
8578 // Set to TRUE if runtime callout is needed for this method
8579 BOOL tiRuntimeCalloutNeeded;
8581 // Set to TRUE if security prolog/epilog callout is needed for this method
8582 // Note: This flag is different than compNeedSecurityCheck.
8583 // compNeedSecurityCheck means whether or not a security object needs
8584 // to be allocated on the stack, which is currently true for EnC as well.
8585 // tiSecurityCalloutNeeded means whether or not security callouts need
8586 // to be inserted in the jitted code.
8587 BOOL tiSecurityCalloutNeeded;
8589 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8590 // This support is necessary to suport attributes that are not described in
8591 // for example, signatures. For example, the permanent home byref (byref that
8592 // points to the gc heap), isn't a property of method signatures, therefore,
8593 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8594 // but when deciding if we need to reimport a block, we need to take these
8596 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8598 // Returns TRUE if child is equal to or a subtype of parent.
8599 // normalisedForStack indicates that both types are normalised for the stack
8600 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8602 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8603 // *pDest is modified to represent the merged type. Sets "*changed" to true
8604 // if this changes "*pDest".
8605 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8607 // Set pDest from the primitive value type.
8608 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8610 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8613 // <BUGNUM> VSW 471305
8614 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8615 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8616 // We use a "short" as we need to push/pop this scope.
8618 short compRegSetCheckLevel;
8622 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8623 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8625 XX IL verification stuff XX
8628 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8629 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8633 // The following is used to track liveness of local variables, initialization
8634 // of valueclass constructors, and type safe use of IL instructions.
8636 // dynamic state info needed for verification
8637 EntryState verCurrentState;
8639 // this ptr of object type .ctors are considered intited only after
8640 // the base class ctor is called, or an alternate ctor is called.
8641 // An uninited this ptr can be used to access fields, but cannot
8642 // be used to call a member function.
8643 BOOL verTrackObjCtorInitState;
8645 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8647 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8648 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8649 void verInitCurrentState();
8650 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8652 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8653 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8654 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8656 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8657 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8658 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8659 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8660 typeInfo verMakeTypeInfo(CorInfoType ciType,
8661 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8662 BOOL verIsSDArray(typeInfo ti);
8663 typeInfo verGetArrayElemType(typeInfo ti);
8665 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8666 BOOL verNeedsVerification();
8667 BOOL verIsByRefLike(const typeInfo& ti);
8668 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8670 // generic type variables range over types that satisfy IsBoxable
8671 BOOL verIsBoxable(const typeInfo& ti);
8673 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8674 DEBUGARG(unsigned line));
8675 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8676 DEBUGARG(unsigned line));
8677 bool verCheckTailCallConstraint(OPCODE opcode,
8678 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8679 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8680 // on a type parameter?
8681 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8682 // return false to the caller.
8683 // If false, it will throw.
8685 bool verIsBoxedValueType(typeInfo ti);
8687 void verVerifyCall(OPCODE opcode,
8688 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8689 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8691 bool readonlyCall, // is this a "readonly." call?
8692 const BYTE* delegateCreateStart,
8693 const BYTE* codeAddr,
8694 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8696 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8698 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8699 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8700 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8701 const CORINFO_FIELD_INFO& fieldInfo,
8702 const typeInfo* tiThis,
8704 BOOL allowPlainStructAsThis = FALSE);
8705 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8706 void verVerifyThisPtrInitialised();
8707 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8709 // Register allocator
8710 void raInitStackFP();
8711 void raEnregisterVarsPrePassStackFP();
8712 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8713 void raEnregisterVarsPostPassStackFP();
8714 void raGenerateFPRefCounts();
8715 void raEnregisterVarsStackFP();
8716 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8718 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8719 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8721 // returns true if enregistering v1 would save more mem accesses than v2
8722 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8725 void raDumpHeightsStackFP();
8726 void raDumpVariableRegIntfFloat();
8729 #if FEATURE_STACK_FP_X87
8731 // Currently, we use FP transition blocks in only 2 situations:
8733 // -conditional jump on longs where FP stack differs with target: it's not strictly
8734 // necessary, but its low frequency and the code would get complicated if we try to
8735 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8736 // minimize the way we generate the jump code.
8737 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8738 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8740 // However, transition blocks have 2 problems
8742 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8743 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8744 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8745 // in the right place without preordering them), this causes us to have to generate the transition
8746 // blocks in the cold area if we want procedure splitting.
8749 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8750 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8751 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8752 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8753 // a big change in the exception.
8755 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8756 // optimizations. For these 2 cases:
8758 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8759 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8760 // a switch statement.
8762 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8763 // current procedure splitting and exception code have.
8764 bool compMayHaveTransitionBlocks;
8766 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8768 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8770 unsigned raCntStkStackFP;
8771 unsigned raCntWtdStkDblStackFP;
8772 unsigned raCntStkParamDblStackFP;
8774 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8775 // TODO: Do we want to put this in LclVarDsc?
8776 unsigned raPayloadStackFP[lclMAX_TRACKED];
8777 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8779 // Useful for debugging
8780 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8782 #endif // FEATURE_STACK_FP_X87
8785 // One line log function. Default level is 0. Increasing it gives you
8786 // more log information
8788 // levels are currently unused: #define JITDUMP(level,...) ();
8789 void JitLogEE(unsigned level, const char* fmt, ...);
8791 bool compDebugBreak;
8793 bool compJitHaltMethod();
8798 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8799 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8801 XX GS Security checks for unsafe buffers XX
8803 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8804 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8807 struct ShadowParamVarInfo
8809 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8810 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8812 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8814 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8815 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8816 // slots and update all trees to refer to shadow slots is done immediately after
8817 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8818 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8819 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8820 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8821 // creating a shadow slot even though this routine returns true.
8823 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8824 // required. There are two cases under which a reg arg could potentially be used from its
8826 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8827 // b) LSRA spills it
8829 // Possible solution to address case (a)
8830 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8831 // in this routine. Note that live out of exception handler is something we may not be
8832 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8833 // Therefore, for methods with exception handling and need GS cookie check we might have
8834 // to take conservative approach.
8836 // Possible solution to address case (b)
8837 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8838 // create a new spill temp if the method needs GS cookie check.
8839 return varDsc->lvIsParam;
8840 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8841 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8848 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8853 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8854 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8855 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8857 void gsGSChecksInitCookie(); // Grabs cookie variable
8858 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8859 bool gsFindVulnerableParams(); // Shadow param analysis code
8860 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8862 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8863 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8865 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8866 // This can be overwritten by setting complus_JITInlineSize env variable.
8868 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8871 #ifdef FEATURE_JIT_METHOD_PERF
8872 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8873 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8875 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8876 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8878 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8880 #if MEASURE_CLRAPI_CALLS
8881 // Thin wrappers that call into JitTimer (if present).
8882 inline void CLRApiCallEnter(unsigned apix);
8883 inline void CLRApiCallLeave(unsigned apix);
8886 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8887 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8892 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8893 // These variables are associated with maintaining SQM data about compile time.
8894 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8895 // in the current compilation.
8896 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8897 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8898 // the inlining phase in the current compilation.
8899 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8901 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8902 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8903 // type-loading and class initialization).
8904 void RecordStateAtEndOfInlining();
8905 // Assumes being called at the end of compilation. Update the SQM state.
8906 void RecordStateAtEndOfCompilation();
8908 #ifdef FEATURE_CLRSQM
8909 // Does anything SQM related necessary at process shutdown time.
8910 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8911 #endif // FEATURE_CLRSQM
8914 #if FUNC_INFO_LOGGING
8915 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8916 // filename to write it to.
8917 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8918 #endif // FUNC_INFO_LOGGING
8920 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8922 // Is the compilation in a full trust context?
8923 bool compIsFullTrust();
8925 #ifndef FEATURE_TRACELOGGING
8926 // Should we actually fire the noway assert body and the exception handler?
8927 bool compShouldThrowOnNoway();
8928 #else // FEATURE_TRACELOGGING
8929 // Should we actually fire the noway assert body and the exception handler?
8930 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8932 // Telemetry instance to use per method compilation.
8933 JitTelemetry compJitTelemetry;
8935 // Get common parameters that have to be logged with most telemetry data.
8936 void compGetTelemetryDefaults(const char** assemblyName,
8937 const char** scopeName,
8938 const char** methodName,
8939 unsigned* methodHash);
8940 #endif // !FEATURE_TRACELOGGING
8944 NodeToTestDataMap* m_nodeTestData;
8946 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8947 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8948 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8949 // Current kept in this.
8951 NodeToTestDataMap* GetNodeTestData()
8953 Compiler* compRoot = impInlineRoot();
8954 if (compRoot->m_nodeTestData == nullptr)
8956 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8958 return compRoot->m_nodeTestData;
8961 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8963 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8964 // currently occur in the AST graph.
8965 NodeToIntMap* FindReachableNodesInNodeTestData();
8967 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8968 // test data, associate that data with "to".
8969 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8971 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8972 // have annotations, attach similar annotations to the corresponding nodes in "to".
8973 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8975 // These are the methods that test that the various conditions implied by the
8976 // test attributes are satisfied.
8977 void JitTestCheckSSA(); // SSA builder tests.
8978 void JitTestCheckVN(); // Value numbering tests.
8981 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8983 FieldSeqStore* m_fieldSeqStore;
8985 FieldSeqStore* GetFieldSeqStore()
8987 Compiler* compRoot = impInlineRoot();
8988 if (compRoot->m_fieldSeqStore == nullptr)
8990 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8991 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8992 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8994 return compRoot->m_fieldSeqStore;
8997 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8999 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9000 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9001 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9002 // attach the field sequence directly to the address node.
9003 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9005 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9007 // Don't need to worry about inlining here
9008 if (m_zeroOffsetFieldMap == nullptr)
9010 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9012 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9013 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9015 return m_zeroOffsetFieldMap;
9018 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9019 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9020 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9021 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9022 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9023 // record the the field sequence using the ZeroOffsetFieldMap described above.
9025 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9026 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9027 // CoreRT. Such case is handled same as the default case.
9028 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9030 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9032 NodeToArrayInfoMap* m_arrayInfoMap;
9034 NodeToArrayInfoMap* GetArrayInfoMap()
9036 Compiler* compRoot = impInlineRoot();
9037 if (compRoot->m_arrayInfoMap == nullptr)
9039 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9040 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9041 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9043 return compRoot->m_arrayInfoMap;
9046 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9048 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9049 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9050 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9051 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9053 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9055 // Use the same map for GCHeap and ByrefExposed when their states match.
9056 memoryKind = ByrefExposed;
9059 assert(memoryKind < MemoryKindCount);
9060 Compiler* compRoot = impInlineRoot();
9061 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9063 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9064 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9065 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9067 return compRoot->m_memorySsaMap[memoryKind];
9070 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9071 CORINFO_CLASS_HANDLE m_refAnyClass;
9072 CORINFO_FIELD_HANDLE GetRefanyDataField()
9074 if (m_refAnyClass == nullptr)
9076 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9078 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9080 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9082 if (m_refAnyClass == nullptr)
9084 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9086 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9090 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9092 #if ALLVARSET_COUNTOPS
9093 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9096 static HelperCallProperties s_helperCallProperties;
9098 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9099 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9100 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9102 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9105 unsigned __int8* offset0,
9106 unsigned __int8* offset1);
9107 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9108 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9110 void fgMorphMultiregStructArgs(GenTreeCall* call);
9111 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9113 }; // end of class Compiler
9115 // Inline methods of CompAllocator.
9116 void* CompAllocator::Alloc(size_t sz)
9118 #if MEASURE_MEM_ALLOC
9119 return m_comp->compGetMem(sz, m_cmk);
9121 return m_comp->compGetMem(sz);
9125 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9127 #if MEASURE_MEM_ALLOC
9128 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9130 return m_comp->compGetMemArray(elems, elemSize);
9134 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9135 inline LclVarDsc::LclVarDsc(Compiler* comp)
9136 : // Initialize the ArgRegs to REG_STK.
9137 // The morph will do the right thing to change
9138 // to the right register if passed in register.
9141 #if FEATURE_MULTIREG_ARGS
9142 _lvOtherArgReg(REG_STK)
9144 #endif // FEATURE_MULTIREG_ARGS
9146 lvRefBlks(BlockSetOps::UninitVal())
9148 #endif // ASSERTION_PROP
9149 lvPerSsaData(comp->getAllocator())
9154 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9155 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9157 XX Miscellaneous Compiler stuff XX
9159 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9160 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9163 // Values used to mark the types a stack slot is used for
9165 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9166 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9167 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9168 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9169 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9170 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9171 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9172 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9174 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9176 /*****************************************************************************
9178 * Variables to keep track of total code amounts.
9183 extern size_t grossVMsize;
9184 extern size_t grossNCsize;
9185 extern size_t totalNCsize;
9187 extern unsigned genMethodICnt;
9188 extern unsigned genMethodNCnt;
9189 extern size_t gcHeaderISize;
9190 extern size_t gcPtrMapISize;
9191 extern size_t gcHeaderNSize;
9192 extern size_t gcPtrMapNSize;
9194 #endif // DISPLAY_SIZES
9196 /*****************************************************************************
9198 * Variables to keep track of basic block counts (more data on 1 BB methods)
9201 #if COUNT_BASIC_BLOCKS
9202 extern Histogram bbCntTable;
9203 extern Histogram bbOneBBSizeTable;
9206 /*****************************************************************************
9208 * Used by optFindNaturalLoops to gather statistical information such as
9209 * - total number of natural loops
9210 * - number of loops with 1, 2, ... exit conditions
9211 * - number of loops that have an iterator (for like)
9212 * - number of loops that have a constant iterator
9217 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9218 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9219 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9220 extern unsigned totalLoopCount; // counts the total number of natural loops
9221 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9222 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9223 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9224 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9226 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9227 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9228 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9229 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9230 extern Histogram loopCountTable; // Histogram of loop counts
9231 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9233 #endif // COUNT_LOOPS
9235 /*****************************************************************************
9236 * variables to keep track of how many iterations we go in a dataflow pass
9241 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9242 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9244 #endif // DATAFLOW_ITER
9246 #if MEASURE_BLOCK_SIZE
9247 extern size_t genFlowNodeSize;
9248 extern size_t genFlowNodeCnt;
9249 #endif // MEASURE_BLOCK_SIZE
9251 #if MEASURE_NODE_SIZE
9252 struct NodeSizeStats
9257 genTreeNodeSize = 0;
9258 genTreeNodeActualSize = 0;
9261 size_t genTreeNodeCnt;
9262 size_t genTreeNodeSize; // The size we allocate
9263 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9264 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9265 // a smaller node to a larger one. TODO-Cleanup: add stats on
9266 // SetOper()/ChangeOper() usage to quanitfy this.
9268 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9269 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9270 extern Histogram genTreeNcntHist;
9271 extern Histogram genTreeNsizHist;
9272 #endif // MEASURE_NODE_SIZE
9274 /*****************************************************************************
9275 * Count fatal errors (including noway_asserts).
9279 extern unsigned fatal_badCode;
9280 extern unsigned fatal_noWay;
9281 extern unsigned fatal_NOMEM;
9282 extern unsigned fatal_noWayAssertBody;
9284 extern unsigned fatal_noWayAssertBodyArgs;
9286 extern unsigned fatal_NYI;
9287 #endif // MEASURE_FATAL
9289 /*****************************************************************************
9293 #ifdef _TARGET_XARCH_
9295 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9296 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9297 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9299 const instruction INS_AND = INS_and;
9300 const instruction INS_OR = INS_or;
9301 const instruction INS_XOR = INS_xor;
9302 const instruction INS_NEG = INS_neg;
9303 const instruction INS_TEST = INS_test;
9304 const instruction INS_MUL = INS_imul;
9305 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9306 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9307 const instruction INS_BREAKPOINT = INS_int3;
9308 const instruction INS_ADDC = INS_adc;
9309 const instruction INS_SUBC = INS_sbb;
9310 const instruction INS_NOT = INS_not;
9316 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9317 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9318 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9320 const instruction INS_AND = INS_and;
9321 const instruction INS_OR = INS_orr;
9322 const instruction INS_XOR = INS_eor;
9323 const instruction INS_NEG = INS_rsb;
9324 const instruction INS_TEST = INS_tst;
9325 const instruction INS_MUL = INS_mul;
9326 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9327 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9328 const instruction INS_BREAKPOINT = INS_bkpt;
9329 const instruction INS_ADDC = INS_adc;
9330 const instruction INS_SUBC = INS_sbc;
9331 const instruction INS_NOT = INS_mvn;
9335 #ifdef _TARGET_ARM64_
9337 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9338 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9339 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9341 const instruction INS_AND = INS_and;
9342 const instruction INS_OR = INS_orr;
9343 const instruction INS_XOR = INS_eor;
9344 const instruction INS_NEG = INS_neg;
9345 const instruction INS_TEST = INS_tst;
9346 const instruction INS_MUL = INS_mul;
9347 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9348 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9349 const instruction INS_BREAKPOINT = INS_bkpt;
9350 const instruction INS_ADDC = INS_adc;
9351 const instruction INS_SUBC = INS_sbc;
9352 const instruction INS_NOT = INS_mvn;
9356 /*****************************************************************************/
9358 extern const BYTE genTypeSizes[];
9359 extern const BYTE genTypeAlignments[];
9360 extern const BYTE genTypeStSzs[];
9361 extern const BYTE genActualTypes[];
9363 /*****************************************************************************/
9365 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9366 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9369 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9370 #elif defined(_TARGET_ARM64_)
9371 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9374 /*****************************************************************************/
9376 #define REG_CORRUPT regNumber(REG_NA + 1)
9377 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9378 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9380 /*****************************************************************************/
9382 extern BasicBlock dummyBB;
9384 /*****************************************************************************/
9385 /*****************************************************************************/
9387 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9388 // nodes of a statement in execution order.
9389 // __stmt: a GT_STMT type GenTree*
9390 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9392 #define foreach_treenode_execution_order(__node, __stmt) \
9393 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9395 // foreach_block: An iterator over all blocks in the function.
9396 // __compiler: the Compiler* object
9397 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9399 #define foreach_block(__compiler, __block) \
9400 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9402 /*****************************************************************************/
9403 /*****************************************************************************/
9407 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9409 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9410 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9412 XX Debugging helpers XX
9414 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9418 /*****************************************************************************/
9419 /* The following functions are intended to be called from the debugger, to dump
9420 * various data structures. The can be used in the debugger Watch or Quick Watch
9421 * windows. They are designed to be short to type and take as few arguments as
9422 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9423 * See the function definition comment for more details.
9426 void cBlock(Compiler* comp, BasicBlock* block);
9427 void cBlocks(Compiler* comp);
9428 void cBlocksV(Compiler* comp);
9429 void cTree(Compiler* comp, GenTree* tree);
9430 void cTrees(Compiler* comp);
9431 void cEH(Compiler* comp);
9432 void cVar(Compiler* comp, unsigned lclNum);
9433 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9434 void cVars(Compiler* comp);
9435 void cVarsFinal(Compiler* comp);
9436 void cBlockPreds(Compiler* comp, BasicBlock* block);
9437 void cReach(Compiler* comp);
9438 void cDoms(Compiler* comp);
9439 void cLiveness(Compiler* comp);
9440 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9442 void cFuncIR(Compiler* comp);
9443 void cBlockIR(Compiler* comp, BasicBlock* block);
9444 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9445 void cTreeIR(Compiler* comp, GenTree* tree);
9446 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9447 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9448 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9449 int cOperandIR(Compiler* comp, GenTree* operand);
9450 int cLeafIR(Compiler* comp, GenTree* tree);
9451 int cIndirIR(Compiler* comp, GenTree* tree);
9452 int cListIR(Compiler* comp, GenTree* list);
9453 int cSsaNumIR(Compiler* comp, GenTree* tree);
9454 int cValNumIR(Compiler* comp, GenTree* tree);
9455 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9457 void dBlock(BasicBlock* block);
9460 void dTree(GenTree* tree);
9463 void dVar(unsigned lclNum);
9464 void dVarDsc(LclVarDsc* varDsc);
9467 void dBlockPreds(BasicBlock* block);
9471 void dCVarSet(VARSET_VALARG_TP vars);
9473 void dVarSet(VARSET_VALARG_TP vars);
9474 void dRegMask(regMaskTP mask);
9477 void dBlockIR(BasicBlock* block);
9478 void dTreeIR(GenTree* tree);
9479 void dLoopIR(Compiler::LoopDsc* loop);
9480 void dLoopNumIR(unsigned loopNum);
9481 int dTabStopIR(int curr, int tabstop);
9482 int dTreeTypeIR(GenTree* tree);
9483 int dTreeKindsIR(GenTree* tree);
9484 int dTreeFlagsIR(GenTree* tree);
9485 int dOperandIR(GenTree* operand);
9486 int dLeafIR(GenTree* tree);
9487 int dIndirIR(GenTree* tree);
9488 int dListIR(GenTree* list);
9489 int dSsaNumIR(GenTree* tree);
9490 int dValNumIR(GenTree* tree);
9491 int dDependsIR(GenTree* comma);
9494 GenTree* dFindTree(GenTree* tree, unsigned id);
9495 GenTree* dFindTree(unsigned id);
9496 GenTreeStmt* dFindStmt(unsigned id);
9497 BasicBlock* dFindBlock(unsigned bbNum);
9501 #include "compiler.hpp" // All the shared inline functions
9503 /*****************************************************************************/
9504 #endif //_COMPILER_H_
9505 /*****************************************************************************/